On the Boundaries of Phonology and Phonetics - Faculteit der ...
On the Boundaries of Phonology and Phonetics - Faculteit der ...
On the Boundaries of Phonology and Phonetics - Faculteit der ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
2Q WKH %RXQGDULHV RI<br />
3KRQRORJ\ DQG 3KRQHWLFV
Sponsored by<br />
Ne<strong>der</strong>l<strong>and</strong>se Vereniging voor Fonetische Wetenschappen<br />
Center for Language <strong>and</strong> Cognition Groningen<br />
Stichting Groninger Universiteitsfonds<br />
Department <strong>of</strong> Linguistics, University <strong>of</strong> Groningen<br />
1 st edition, January 2004<br />
2 nd edition, February 2004<br />
ISBN 90 367 1930 5
UNIVERSITY OF GRONINGEN<br />
2Q WKH %RXQGDULHV RI<br />
3KRQRORJ\ DQG 3KRQHWLFV<br />
Edited by<br />
Dicky Gilbers<br />
Maartje Schreu<strong>der</strong><br />
Nienke Knevel<br />
To honour Tjeerd de Graaf
Contents<br />
<strong>On</strong> <strong>the</strong> <strong>Boundaries</strong> <strong>of</strong> <strong>Phonology</strong> <strong>and</strong> <strong>Phonetics</strong> 7<br />
The Editors: Dicky Gilbers, Maartje Schreu<strong>der</strong> <strong>and</strong> Nienke Knevel<br />
Tjeerd de Graaf 15<br />
Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
Tseard de Graaf 31<br />
Oerset troch Jurjen van <strong>der</strong> Kooi<br />
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 37<br />
Vincent J. van Heuven<br />
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 61<br />
Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
Learning Phonotactics with Simple Processors 89<br />
John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
Weak Interactions 123<br />
Tamás Bíró<br />
Prosodic Acquisition: a Comparison <strong>of</strong> Two Theories 147<br />
Angela Grimm<br />
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 159<br />
Hidetoshi Shiraishi<br />
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 183<br />
Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
List <strong>of</strong> Addresses 203
<strong>On</strong> <strong>the</strong> <strong>Boundaries</strong> <strong>of</strong> <strong>Phonology</strong> <strong>and</strong> <strong>Phonetics</strong><br />
The Editors: Dicky Gilbers, Maartje Schreu<strong>der</strong> <strong>and</strong><br />
Nienke Knevel<br />
In this volume a collection <strong>of</strong> papers is presented in which <strong>the</strong> boundaries<br />
<strong>of</strong> phonology <strong>and</strong> phonetics are explored. In current phonological research,<br />
<strong>the</strong> distinction between phonology, as <strong>the</strong> study <strong>of</strong> sound systems <strong>of</strong><br />
languages, <strong>and</strong> phonetics, as <strong>the</strong> study <strong>of</strong> <strong>the</strong> characteristics <strong>of</strong> human<br />
(speech) sound making, seems to be blurred.<br />
Consi<strong>der</strong> an example <strong>of</strong> <strong>the</strong> phonological process <strong>of</strong> /l/-substitution as<br />
exemplified in <strong>the</strong> data in Table 1.<br />
Table 1. /l/ substitutions<br />
Historical Dutch data:<br />
/l/ → [w]<br />
alt/olt oud 'old'<br />
kalt/kolt koud 'cold'<br />
schoo[l] schoo[w] 'school'<br />
First Language Acquisition data (Dutch):<br />
hallo ha[w]o 'hello'<br />
lief [w]ief 'sweet'<br />
blauw b[w]auw 'blue'<br />
In phonology, <strong>the</strong> substitution segment is expected to be a minimal<br />
deviation from <strong>the</strong> target segment. For example, boot ‘boat’ could be<br />
realized as [pot], but not as [lot], since <strong>the</strong> target /b/ <strong>and</strong> <strong>the</strong> output [l] differ<br />
in too many dimensions. In o<strong>the</strong>r words, sound substitutions should be<br />
characterized more commonly by single feature changes than by several<br />
feature changes. The widely attested substitution <strong>of</strong> /l/ by [w], however,<br />
cannot be accounted for adequately as a minimal deviation from <strong>the</strong> target<br />
based on articulatorily defined features, as shown in Figure 1.
8 The Editors: Dicky Gilbers, Maartje Schreu<strong>der</strong> <strong>and</strong> Nienke Knevel<br />
Figure 1. /l/-substitutions<br />
/l/ → [w]<br />
+ son + son<br />
+ cons - cons<br />
+ cont + cont<br />
+ lat - lat<br />
- lab + lab<br />
+ ant - ant<br />
+ cor - cor<br />
- high + high<br />
- back + back<br />
- round + round<br />
From an acoustic point <strong>of</strong> view, liquid-glide alternations can be described<br />
as minimal changes. The differences between <strong>the</strong> individual glides <strong>and</strong><br />
liquids can be related to <strong>the</strong>ir relative second <strong>and</strong> third formant locus<br />
frequencies. Ainsworth <strong>and</strong> Paliwal (1984) found that in a perceptualidentification<br />
experiment liquids such as [l] having a mid F2 locus<br />
frequency were classified as [w] if <strong>the</strong>y had a low F2 locus frequency <strong>and</strong><br />
as [j] if <strong>the</strong>y had a high F2 locus frequency.<br />
3160 Hz w w w l l l l j j j<br />
↑ w w w l l l l j j j<br />
F3 locus freq. w w w r r r l j j j<br />
↓ w w w r r r j j j j<br />
1540 Hz w w r r r r r j j j<br />
760 Hz ← F2 locus freq. → 2380 Hz<br />
Figure 2. Typical set <strong>of</strong> responses obtained from listening to glide/liquid-vowel<br />
syn<strong>the</strong>tic stimuli (after Ainsworth & Paliwal, 1984 (simplified))<br />
Based on <strong>the</strong>se acoustic characteristics, liquid-glide substitutions can be<br />
described as a minimal change from <strong>the</strong> target, which cannot be done in <strong>the</strong><br />
phonological representation <strong>of</strong> <strong>the</strong>se sounds. Obviously, phonology needs
<strong>On</strong> <strong>the</strong> <strong>Boundaries</strong> <strong>of</strong> <strong>Phonology</strong> <strong>and</strong> <strong>Phonetics</strong> 9<br />
phonetic information to explain a phonological process <strong>of</strong> this kind (cf.<br />
Gilbers, 2002).<br />
Now consi<strong>der</strong> <strong>the</strong> Dutch process <strong>of</strong> schwa insertion as exhibited in<br />
Table 2.<br />
Table 2. schwa insertion in Dutch<br />
helm [��l�m] 'helmet' darm [��r�m] 'intestine'<br />
half [��l�f] 'half' durf [��r�f] 'courage'<br />
melk [��l�k] 'milk' hark [��r�k] 'rake'<br />
not in: vals 'out <strong>of</strong> tune', hals 'neck', hart 'heart', start 'start'<br />
Schwa may be inserted between a liquid /l,r/ <strong>and</strong> a non-homorganic<br />
consonant (i.e. a consonant that differs in place <strong>of</strong> articulation with /l,r/) at<br />
<strong>the</strong> end <strong>of</strong> a syllable. Therefore, schwa may be inserted between coronal /l/<br />
or /r/ <strong>and</strong> non-coronal /m/, /f/, /k/, etc. Schwa is not allowed, however,<br />
between /l/ or /r/ <strong>and</strong> a coronal obstruent /s/ or /t/. Now, Dutch has at least<br />
two different varieties <strong>of</strong> /r/: an alveolar [r] <strong>and</strong> a uvular [�]. Since <strong>the</strong>re is<br />
no functional difference between realizations such as [���] <strong>and</strong> [���] for rat<br />
'rat', however, <strong>the</strong>re is only one phoneme /r/ in <strong>the</strong> Dutch system with its<br />
allophones [r] <strong>and</strong> [�]. Interestingly, even Dutch speakers with a uvular [�]<br />
do not show schwa insertion between <strong>the</strong>ir [�] <strong>and</strong> non-homorganic coronal<br />
obstruent /s/ or /t/. The process <strong>of</strong> schwa insertion, apparently, takes place<br />
before <strong>the</strong> phonetic level <strong>of</strong> actual realization <strong>of</strong> segments, i.e. on <strong>the</strong><br />
abstract phonological level, where /r/, /s/ <strong>and</strong> /t/ share <strong>the</strong>ir place feature<br />
[coronal]. Synchronically, <strong>the</strong> process can only be described in a<br />
phonological way, even though it may have had a phonetic - articulatory -<br />
base originally. We assume that uvular [�] is a later variant <strong>of</strong> Dutch /r/<br />
than coronal [r], just as <strong>the</strong> even younger, recently observed allophonic<br />
variant [�] in Western Dutch dialects: raar 'strange' realized as [ra:�]. These<br />
allophones date from times when <strong>the</strong> process <strong>of</strong> schwa insertion between<br />
non-homorganic, syllable-final liquid-consonant clusters was already<br />
'fossilized' in <strong>the</strong> Dutch system.<br />
The above-mentioned two accounts <strong>of</strong> phonological processes indicate<br />
<strong>the</strong> way many phonologists approach <strong>the</strong>ir research objects nowadays.<br />
More <strong>and</strong> more <strong>the</strong> distinction between phonology <strong>and</strong> phonetics is<br />
challenged in attempts to provide adequate accounts <strong>of</strong> <strong>the</strong> phonological<br />
phenomena. In this way, <strong>the</strong> phonologists <strong>of</strong> <strong>the</strong> so-called CLCG Klankleer
10 The Editors: Dicky Gilbers, Maartje Schreu<strong>der</strong> <strong>and</strong> Nienke Knevel<br />
group in Groningen study <strong>the</strong> phonology-phonetics interface, whereas o<strong>the</strong>r<br />
members <strong>of</strong> <strong>the</strong> group cross <strong>the</strong> boundaries <strong>of</strong> phonology <strong>and</strong> phonetics by<br />
combining <strong>the</strong> study <strong>of</strong> sound patterns with dialectology, computational<br />
linguistics, musicology, first language acquisition or ethnolinguistics.<br />
The Center for Language <strong>and</strong> Cognition Groningen (CLCG) is a<br />
research institute within <strong>the</strong> Faculty <strong>of</strong> Arts <strong>of</strong> <strong>the</strong> University <strong>of</strong> Groningen.<br />
It comprises most <strong>of</strong> <strong>the</strong> linguistic research that is being carried out within<br />
<strong>the</strong> Faculty <strong>of</strong> Arts. <strong>On</strong>e <strong>of</strong> <strong>the</strong> research groups <strong>of</strong> CLCG is this 'Klankleer'<br />
group (<strong>Phonology</strong> <strong>and</strong> <strong>Phonetics</strong>), which focuses on <strong>the</strong> structure <strong>and</strong><br />
contents <strong>of</strong> <strong>the</strong> sounds <strong>of</strong> language.<br />
This volume <strong>of</strong> papers by members <strong>of</strong> <strong>the</strong> Klankleer group is dedicated<br />
to Tjeerd de Graaf, who was <strong>the</strong> coordinator <strong>of</strong> this group from 1999 until<br />
2003. It does not mean that Tjeerd no longer participates in <strong>the</strong> group,<br />
because he still supervises two PhD projects. These projects by Hidetoshi<br />
Shiraishi <strong>and</strong> Markus Bergmann combine phonetics <strong>and</strong> phonology with<br />
ethnolinguistics. As mentioned above, <strong>the</strong> research <strong>of</strong> most members <strong>of</strong> <strong>the</strong><br />
group involves combinations <strong>of</strong> different (linguistic) areas. Wilbert<br />
Heeringa, Charlotte Gooskens <strong>and</strong> Roberto Bolognesi apply phonetics to<br />
<strong>the</strong> study <strong>of</strong> dialectology. Nanne Streekstra is one <strong>of</strong> <strong>the</strong> first linguists in<br />
our group who was interested in <strong>the</strong> phonology-phonetics interface. Wouter<br />
Jansen's work is exemplary for this so-called 'laboratory phonology'. He<br />
provides acoustic studies <strong>of</strong> voicing assimilation in obstruent clusters in<br />
Germanic languages. Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers combine<br />
phonetics <strong>and</strong> phonology with areas beyond linguistics, such as music<br />
<strong>the</strong>ory. Former member Klarien van <strong>der</strong> Linde <strong>and</strong> Angela Grimm study<br />
first language acquisition, whereas W<strong>and</strong>er Lowie studies second language<br />
acquisition. Finally, Tjeerd de Graaf started his academic life as a<br />
researcher in <strong>the</strong>oretical physics, switched to phonetics, whereas his main<br />
interest is now in ethnolinguistics. This homo universalis also plays piano<br />
<strong>and</strong> oboe <strong>and</strong> speaks nine different languages. This Festschrift, however, is<br />
dedicated to <strong>the</strong> phonetician Tjeerd de Graaf. The papers cover a wide<br />
range <strong>of</strong> topics varying from ethnolinguistics to computational linguistics<br />
<strong>and</strong> from first language acquisition to dialectology. The common<br />
denominator is that all researchers work on <strong>the</strong> boundaries <strong>of</strong> phonology<br />
<strong>and</strong> phonetics.<br />
Vincent van Heuven, as a guest author from University <strong>of</strong> Leiden,<br />
won<strong>der</strong>s whe<strong>the</strong>r certain distinctions in <strong>the</strong> speech signal are phonological<br />
or phonetic. He investigates whe<strong>the</strong>r different prosodic boundary tones<br />
form a continuum or whe<strong>the</strong>r <strong>the</strong>y are categorical. He finds a categorical
<strong>On</strong> <strong>the</strong> <strong>Boundaries</strong> <strong>of</strong> <strong>Phonology</strong> <strong>and</strong> <strong>Phonetics</strong> 11<br />
division between low (declarative) <strong>and</strong> non-low tones, but within <strong>the</strong> nonlow<br />
category <strong>the</strong> cross-over from continuation to question is ra<strong>the</strong>r gradual.<br />
Charlotte Gooskens <strong>and</strong> Wilbert Heeringa measured linguistic distances<br />
between Frisian dialects <strong>and</strong> <strong>the</strong> o<strong>the</strong>r Germanic languages in or<strong>der</strong> to get<br />
an impression <strong>of</strong> <strong>the</strong> effect <strong>of</strong> genetic relationship <strong>and</strong> language contact on<br />
<strong>the</strong> position <strong>of</strong> <strong>the</strong> mo<strong>der</strong>n Frisian language on <strong>the</strong> Germanic language<br />
map. Wilbert is a member <strong>of</strong> <strong>the</strong> CLCG group 'Computational Linguistics'.<br />
John Nerbonne participates as head <strong>of</strong> CLCG. His paper with Ivilin<br />
Stoianov explores <strong>the</strong> learning <strong>of</strong> phonotactics in neural networks, in<br />
particular <strong>the</strong> so-called Simple Recurrent Networks (SRNs). SRNs provide<br />
a valuable means <strong>of</strong> exploring what information in <strong>the</strong> linguistic signal<br />
could in principle be acquired by a very primitive learning mechanism.<br />
Tamás Bíró, who is also a member <strong>of</strong> 'Computational Linguistics' <strong>and</strong><br />
interested in phonology, claims that <strong>the</strong> types <strong>of</strong> interactions between<br />
languages can be extremely diverse, depending on a number <strong>of</strong> factors. The<br />
paper analyses three case studies, namely <strong>the</strong> influence <strong>of</strong> Yiddish on<br />
Hungarian, Mo<strong>der</strong>n Hebrew <strong>and</strong> Esperanto.<br />
Angela Grimm discusses a number <strong>of</strong> empirical <strong>and</strong> <strong>the</strong>oretical<br />
problems with respect to two models <strong>of</strong> prosodic acquisition: a template<br />
mapping model <strong>and</strong> a prosodic hierarchy model. Both models assume that<br />
<strong>the</strong> acquisition <strong>of</strong> word prosody is guided by universal prosodic principles.<br />
Toshi Shiraishi discusses phonological asymmetries between nominal<br />
<strong>and</strong> verbal stems <strong>of</strong> Nivkh, a minority language spoken on <strong>the</strong> isl<strong>and</strong> <strong>of</strong><br />
Sakhalin. These asymmetries are observed in two phonological phenomena:<br />
consonant alternation <strong>and</strong> final fricative devoicing. Though <strong>the</strong><br />
asymmetries <strong>the</strong>mselves look very different on <strong>the</strong> surface, Toshi's paper<br />
makes explicit that <strong>the</strong>y are subject to a common generalization, Base-<br />
Identity.<br />
Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers won<strong>der</strong>ed whe<strong>the</strong>r <strong>the</strong> influence<br />
<strong>of</strong> a higher speech rate leads to adjustment <strong>of</strong> <strong>the</strong> rhythmic pattern, as it<br />
does in music, or just to 'phonetic compression' with preservation <strong>of</strong> <strong>the</strong><br />
phonological structure. An example <strong>of</strong> an item <strong>the</strong>y examined is <strong>the</strong> Dutch<br />
word perfèctioníst, which can get <strong>the</strong> rhythmic structure pèrfectioníst in fast<br />
tempo. The results indeed showed a preference for restructured rhythms in<br />
fast speech.<br />
With this very diverse collection <strong>of</strong> papers, we hope to present <strong>the</strong><br />
phonetician Tjeerd de Graaf a representative selection <strong>of</strong> <strong>the</strong> current<br />
activities <strong>of</strong> his CLCG-Klankleer group.
12 The Editors: Dicky Gilbers, Maartje Schreu<strong>der</strong> <strong>and</strong> Nienke Knevel<br />
In <strong>the</strong> 1970's <strong>and</strong> 1980's Tjeerd's phonetic research stood miles away<br />
from <strong>the</strong> feature geometries <strong>and</strong> grid representations that were customary in<br />
phonology. He used to make sonagrams, i.e. visual displays <strong>of</strong> sound<br />
spectrograms, <strong>of</strong> e.g. [p�], [si] <strong>and</strong> [r�]. But when <strong>the</strong> violin string <strong>of</strong> his<br />
sonagraph broke, he wasn't able to do phonetic research anymore <strong>and</strong> that is<br />
when ethnolinguistics stole his heart. Nowadays, it is much easier to do<br />
phonetic analyses on <strong>the</strong> computer using programs, such as PRAAT<br />
(Boersma <strong>and</strong> Weenink, 1992-2003). Whereas phonetics <strong>and</strong> phonology<br />
grew apart from each o<strong>the</strong>r since <strong>the</strong>y were installed as two distinct<br />
disciplines <strong>of</strong> linguistics at <strong>the</strong> First International Congress <strong>of</strong> Linguists<br />
(The Hague 1928), current laboratory phonological research may even<br />
suggest that phonetics <strong>and</strong> phonology coincide. However, as shown in <strong>the</strong><br />
two examples in this introductory paper, /l/-substitution <strong>and</strong> schwainsertion,<br />
<strong>the</strong> role <strong>of</strong> both disciplines is still distinguishable. That does not<br />
alter <strong>the</strong> fact that co-operation between phoneticians <strong>and</strong> phonologists must<br />
be an integral part <strong>of</strong> <strong>the</strong> study <strong>of</strong> sound patterns. Some sound phenomena,<br />
such as ethnolinguistic <strong>and</strong> dialect differences or acquisition data, can only<br />
be explained adequately if both phonological <strong>and</strong> phonetic characteristics<br />
<strong>of</strong> sounds are consi<strong>der</strong>ed.<br />
University <strong>of</strong> Groningen, January 2004<br />
This volume was presented to Tjeerd de Graaf on January 30, 2004 at <strong>the</strong><br />
workshop '<strong>On</strong> <strong>the</strong> <strong>Boundaries</strong> <strong>of</strong> <strong>Phonology</strong> <strong>and</strong> <strong>Phonetics</strong>'. The CLCG <strong>and</strong><br />
<strong>the</strong> Department <strong>of</strong> Linguistics <strong>of</strong> <strong>the</strong> University <strong>of</strong> Groningen, 'de<br />
Ne<strong>der</strong>l<strong>and</strong>se Vereniging voor Fonetische Wetenschappen' <strong>and</strong> GUF<br />
(Stichting Groninger Universiteitsfonds) sponsored this workshop. Keynote<br />
speakers were Vincent van Heuven <strong>and</strong> Carlos Gussenhoven.<br />
References<br />
Ainsworth, W.A. & K.K. Paliwal (1984). Correlation between <strong>the</strong> production<br />
<strong>and</strong> perception <strong>of</strong> <strong>the</strong> English glides /w,r,l,j/. Journal <strong>of</strong><br />
<strong>Phonetics</strong>, 12: 237-243.<br />
Boersma, Paul, <strong>and</strong> David Weenink (1992-2003). PRAAT, phonetics by<br />
computer. Available at http://www.praat.org. University <strong>of</strong><br />
Amsterdam.
<strong>On</strong> <strong>the</strong> <strong>Boundaries</strong> <strong>of</strong> <strong>Phonology</strong> <strong>and</strong> <strong>Phonetics</strong> 13<br />
Gilbers, D.G. (2002). Conflicting phonologically based <strong>and</strong> phonetically based<br />
constraints in <strong>the</strong> analysis <strong>of</strong> /l/-substitutions. In: M. Beers, P.<br />
Jongmans & A. Wijn<strong>and</strong>s (eds). Netwerk Eerste Taalverwerving,<br />
Net-bulletin 2001. Leiden, 22-40.
Tjeerd de Graaf<br />
Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi<br />
Shiraishi<br />
Tjeerd de Graaf was born on January 27th 1938 in Leeuwarden, <strong>the</strong> capital<br />
<strong>of</strong> <strong>the</strong> province Fryslân in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s. Fryslân is <strong>the</strong> largest <strong>of</strong> several<br />
regions on <strong>the</strong> North Sea where Frisian is spoken, a West Germanic<br />
language whose genetically closest relative is English.<br />
Tjeerd’s parents were both Frisians, <strong>and</strong> at home <strong>the</strong>y spoke exclusively<br />
Frisian. As most o<strong>the</strong>r children in Fryslân at that time, Tjeerd grew up<br />
bilingually. His first native language was Frisian, <strong>and</strong> at school he learned<br />
Dutch, <strong>the</strong> <strong>of</strong>ficial language <strong>of</strong> <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s.<br />
The coexistence <strong>of</strong> Frisian at home <strong>and</strong> Dutch at school was Tjeerd’s<br />
first experience in a fascinating world <strong>of</strong> different languages. For Tjeerd,<br />
<strong>the</strong> difference between <strong>the</strong> two languages had a very illustrative spatial<br />
implication: when he <strong>and</strong> <strong>the</strong> o<strong>the</strong>r children in his neighborhood went to<br />
school in <strong>the</strong> mornings, <strong>the</strong>re was a railway crossing along <strong>the</strong> way. <strong>On</strong>ce<br />
<strong>the</strong>y had crossed it <strong>the</strong>y stopped speaking Frisian <strong>and</strong> switched to Dutch,<br />
<strong>the</strong>ir <strong>of</strong>ficial school language.<br />
At <strong>the</strong> age <strong>of</strong> 18, in 1956, Tjeerd graduated from <strong>the</strong> Leeuwarden High<br />
School <strong>and</strong> became interested in languages. His o<strong>the</strong>r big passion was <strong>the</strong><br />
science <strong>of</strong> physics <strong>and</strong> astronomy. The oldest planetarium in <strong>the</strong> world is<br />
located in Franeker, an old academic place in Fryslân. Intrigued by <strong>the</strong> laws<br />
governing space <strong>and</strong> time, Tjeerd studied physics at <strong>the</strong> University <strong>of</strong><br />
Groningen from 1956 to 1963. In 1963 he received his master’s degree in<br />
science (Doctoraal examen) in <strong>the</strong>oretical physics, a combination <strong>of</strong><br />
physics, ma<strong>the</strong>matics <strong>and</strong> astronomy. From 1963 until 1969 he continued<br />
as a research associate at <strong>the</strong> Institute <strong>of</strong> Theoretical Physics at <strong>the</strong><br />
University <strong>of</strong> Groningen.<br />
Tjeerd was already a “polyglot” at that time, speaking not only Frisian<br />
<strong>and</strong> Dutch, but also German, English <strong>and</strong> French. O<strong>the</strong>r languages would<br />
follow. In <strong>the</strong> former Soviet Union <strong>the</strong> study <strong>of</strong> astronomical sciences was<br />
enjoying an era <strong>of</strong> superiority. Tjeerd un<strong>der</strong>stood that learning Russian <strong>and</strong><br />
o<strong>the</strong>r East European languages would be <strong>the</strong> key to enter <strong>the</strong> field <strong>of</strong><br />
scientific knowledge. Along with his <strong>the</strong>oretical physics’ studies, he also
16 Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
enrolled for <strong>the</strong> study <strong>of</strong> Slavic languages. The new technologies <strong>and</strong> <strong>the</strong>ir<br />
application for future research fascinated him. In 1967 he received his<br />
Master <strong>of</strong> Arts degree (K<strong>and</strong>idaatsexamen) in Slavic languages <strong>and</strong><br />
computer linguistics. In <strong>the</strong> meantime, after having obtained his MS, he<br />
continued his research in <strong>the</strong>oretical physics, combined with a study abroad<br />
in Pol<strong>and</strong>, where he lived for half a year <strong>and</strong> mastered <strong>the</strong> language.<br />
By 1969, he finished his dissertation entitled “Aspects <strong>of</strong> Neutrino<br />
Astrophysics”.<br />
The cover page <strong>of</strong> Tjeerd’s dissertation in Theoretical Physics in 1969<br />
Tjeerd’s quenchless thirst for knowledge led him to Engl<strong>and</strong> toge<strong>the</strong>r with<br />
his wife Nynke <strong>and</strong> <strong>the</strong>ir children where <strong>the</strong>y spent a year from 1970 to<br />
1971 <strong>and</strong> where he worked as a research associate at <strong>the</strong> Institute <strong>of</strong><br />
Theoretical Physics at <strong>the</strong> University <strong>of</strong> Cambridge.<br />
Upon <strong>the</strong>ir return to Groningen, Tjeerd became assistant pr<strong>of</strong>essor in<br />
physics at <strong>the</strong> Institute <strong>of</strong> Astronomy, a post he held until 1975. This was to<br />
be a turning point in his pr<strong>of</strong>essional career when he decided to switch to<br />
his second passion, namely <strong>the</strong> study <strong>of</strong> languages. <strong>On</strong>e <strong>of</strong> his dissertational<br />
<strong>the</strong>ses dealt with <strong>the</strong> question as to how exact a person’s identity could be<br />
defined by his or her speech. This <strong>the</strong>sis symbolically defined one <strong>of</strong><br />
Tjeerd’s later linguistic interests: <strong>the</strong> aspects <strong>of</strong> spoken language, <strong>the</strong> study<br />
<strong>of</strong> phonetics.<br />
In 1975, Tjeerd became associate pr<strong>of</strong>essor at <strong>the</strong> Institute <strong>of</strong> Phonetic<br />
Sciences, Department <strong>of</strong> Linguistics, University <strong>of</strong> Groningen.<br />
Being a native bilingual in Frisian <strong>and</strong> Dutch, Tjeerd was aware <strong>of</strong> <strong>the</strong><br />
numerous phonetic differences between <strong>the</strong> languages. Having studied
Tjeerd de Graaf 17<br />
many o<strong>the</strong>r languages as well, Tjeerd un<strong>der</strong>stood how important phonetic<br />
descriptions are not only for <strong>the</strong>oretical linguistics, but also in learning <strong>and</strong><br />
teaching foreign languages.<br />
Language coexistence <strong>and</strong> language change would become ano<strong>the</strong>r focal<br />
point <strong>of</strong> his research. In most regions <strong>of</strong> <strong>the</strong> world, people are bilingual or<br />
even multilingual. Language variety appears both in space <strong>and</strong> time.<br />
Listening to radio programs or TV broadcasts dating back ten or twenty<br />
years, reveals a distinct difference in speech as compared with today’s<br />
custom <strong>of</strong> speaking. It is still <strong>the</strong> same language, <strong>the</strong> same place, <strong>and</strong> yet<br />
<strong>the</strong> speech is not <strong>the</strong> same as before. Not only <strong>the</strong> lexicon <strong>of</strong> a language<br />
changes but also <strong>the</strong> manner in which people speak, <strong>the</strong>ir pronunciation <strong>and</strong><br />
intonation. This is an extremely intriguing topic for a person interested in<br />
languages <strong>and</strong> <strong>the</strong>ir varieties.<br />
Tjeerd started to trace <strong>the</strong> oldest recordings <strong>of</strong> spoken examples <strong>of</strong><br />
languages. He analyzed Frisian recordings from <strong>the</strong> province <strong>of</strong> Fryslân as<br />
well as recordings from North <strong>and</strong> East Frisian regions. Recordings <strong>of</strong> <strong>the</strong><br />
spoken language <strong>of</strong> former times are not only a historically important<br />
heritage, but <strong>the</strong>y also <strong>of</strong>fer valuable information pertaining to language<br />
shift processes. A practical problem with <strong>the</strong> oldest sound recordings is that<br />
<strong>the</strong>y were made on wax cylin<strong>der</strong>s <strong>and</strong> <strong>the</strong>ir quality decreases tremendously<br />
every time <strong>the</strong>y are listened to. Tjeerd was aware <strong>of</strong> <strong>the</strong> fact that one <strong>of</strong> <strong>the</strong><br />
main tasks was to transfer <strong>the</strong>se recordings to mo<strong>der</strong>n media in or<strong>der</strong> to<br />
preserve <strong>the</strong>m. In <strong>the</strong> beginning <strong>of</strong> <strong>the</strong> 1990s, toge<strong>the</strong>r with Japanese<br />
colleagues, Tjeerd started to investigate <strong>the</strong> possibilities <strong>of</strong> preserving old<br />
language recordings via mo<strong>der</strong>n audio technology. At that time, Tjeerd<br />
acquired yet ano<strong>the</strong>r language, namely Japanese.<br />
Tjeerd working on wax cylin<strong>der</strong>s with old recordings <strong>of</strong> Dutch
18 Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
Tjeerd started to contact <strong>the</strong> most important sound archives <strong>of</strong> <strong>the</strong> world,<br />
which are in Vienna, Berlin, <strong>and</strong> St. Petersburg. Through his collaboration<br />
with <strong>the</strong> sound archive <strong>of</strong> <strong>the</strong> Academy <strong>of</strong> Sciences in St. Petersburg in <strong>the</strong><br />
1990s, he renewed his contact with Russia, which had begun with his<br />
studies <strong>of</strong> Slavic languages in <strong>the</strong> 1960s.<br />
After 1990, <strong>the</strong> world had experienced dramatic changes. The Iron<br />
Curtain had disappeared <strong>and</strong> Russia had once again opened her “Window to<br />
<strong>the</strong> West”. When Tjeerd came back to St. Petersburg in <strong>the</strong> 1990s, he was<br />
immediately fascinated by this city he had visited for <strong>the</strong> first time some<br />
twenty years ago when it was still known as Leningrad. As a Frisian <strong>and</strong> a<br />
Dutchman, he felt at home <strong>the</strong>re. The picturesque canals <strong>and</strong> paths along<br />
<strong>the</strong> wide boulevards reminded him <strong>of</strong> his home region. This was no<br />
coincidence: Czar Peter <strong>the</strong> Great, some 300 years ago, had chosen Holl<strong>and</strong><br />
as <strong>the</strong> model for his new capital.<br />
In <strong>the</strong> following years, Tjeerd organized joint projects with <strong>the</strong> Russian<br />
Academy <strong>of</strong> Sciences <strong>and</strong> St. Petersburg State University to preserve <strong>and</strong><br />
transfer old Russian sound recordings onto mo<strong>der</strong>n digital audio media.<br />
Research on a vast collection <strong>of</strong> <strong>the</strong> most various sound recordings<br />
resulting from many linguistic field work expeditions from <strong>the</strong> end <strong>of</strong> <strong>the</strong><br />
XIX <strong>and</strong> XX centuries served as an incentive for several projects related to<br />
different languages spoken in Russia.<br />
Tjeerd started to initiate research projects on <strong>the</strong> language spoken by <strong>the</strong><br />
Mennonites, a group <strong>of</strong> people in Siberia, who had originally come from<br />
regions in <strong>the</strong> Nor<strong>the</strong>rn Ne<strong>the</strong>rl<strong>and</strong>s <strong>and</strong> Germany <strong>and</strong> still speak <strong>the</strong><br />
language <strong>of</strong> <strong>the</strong>ir ancestors – in fact a language with great similarities to <strong>the</strong><br />
mo<strong>der</strong>n dialects spoken in North-Germany <strong>and</strong> nor<strong>the</strong>rn parts <strong>of</strong> <strong>the</strong><br />
Ne<strong>the</strong>rl<strong>and</strong>s. The Dutch press even reported about “Siberians speak<br />
Gronings”.<br />
Languages do not only divide people <strong>of</strong> different nations, but also build<br />
a bridge between <strong>the</strong>m. Tjeerd showed this with his research work. Even in<br />
far-away Siberia <strong>the</strong>re are people speaking almost <strong>the</strong> same language as in<br />
Groningen. When planning his expeditions, Tjeerd was concerned with<br />
both scientific aims <strong>and</strong> <strong>the</strong> organization <strong>of</strong> humanitarian aid from<br />
Groningen to <strong>the</strong> Siberian villages he visited.<br />
Language as a cultural heritage became <strong>the</strong> core <strong>of</strong> Tjeerd’s linguistic<br />
activities. With his bilingual origin, he set <strong>the</strong> perfect example. Throughout<br />
his life, he showed that each individual can contribute to <strong>the</strong> survival <strong>of</strong> a<br />
language. With his Frisian wife Nynke, whom he met in his student years,<br />
Tjeerd used to converse in Dutch. After <strong>the</strong>ir parents had passed away, <strong>the</strong>y
Tjeerd de Graaf 19<br />
decided to switch to Frisian. They personally experienced how a language<br />
slowly starts to become extinct if <strong>the</strong> children do not carry on <strong>the</strong> language.<br />
This attitude defined Tjeerd’s successive research activities in Russia.<br />
Subsequent projects, which he coordinated now, had two goals:<br />
documentation <strong>of</strong> endangered languages, <strong>and</strong> revitalizing <strong>and</strong> preserving<br />
<strong>the</strong>m for future generations. In <strong>the</strong> following projects, both aspects –<br />
preservation <strong>and</strong> fur<strong>the</strong>r development – were present. Tjeerd made several<br />
expeditions, among o<strong>the</strong>rs to Yakutia <strong>and</strong> <strong>the</strong> Isl<strong>and</strong> <strong>of</strong> Sakhalin, where he<br />
<strong>and</strong> o<strong>the</strong>r linguists recorded <strong>the</strong> speech <strong>of</strong> <strong>the</strong> local indigenous peoples.<br />
Tjeerd de Graaf with a group <strong>of</strong> speakers <strong>of</strong> indigenous languages <strong>of</strong> <strong>the</strong> Isl<strong>and</strong> <strong>of</strong><br />
Sakhalin in <strong>the</strong> Far East <strong>of</strong> Russia: Uiltas <strong>and</strong> Nivkhs, in <strong>the</strong> 1990s.<br />
In <strong>the</strong> second half <strong>of</strong> <strong>the</strong> 1990s, Tjeerd coordinated several projects with<br />
Institutions throughout <strong>the</strong> Russian Fe<strong>der</strong>ation funded by <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s<br />
Organization for Scientific Research <strong>and</strong> <strong>the</strong> EU INTAS organization in<br />
Brussels.<br />
His main goal was to make young people aware <strong>of</strong> <strong>the</strong>ir unique<br />
linguistic heritage <strong>and</strong> stimulate <strong>the</strong>m in supporting minority <strong>and</strong> regional<br />
languages. In 1998, Tjeerd was appointed Knight in <strong>the</strong> Or<strong>der</strong> <strong>of</strong> <strong>the</strong> Dutch<br />
Lion for his research <strong>and</strong> contribution in support <strong>of</strong> <strong>the</strong> preservation <strong>and</strong><br />
construction <strong>of</strong> databases for <strong>the</strong> minority languages in Russia. Later that<br />
same year Tjeerd was awarded an honorary doctorate at <strong>the</strong> University <strong>of</strong><br />
St. Petersburg for his contribution in <strong>the</strong> joint language preservation<br />
projects.
20 Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
Tjeerd de Graaf is appointed Doctor Honoris Causa at <strong>the</strong> University <strong>of</strong><br />
St. Petersburg, November 1998.<br />
Tjeerd has retired from <strong>the</strong> University <strong>of</strong> Groningen in 2003 <strong>and</strong> vacated<br />
<strong>the</strong> chair <strong>of</strong> <strong>the</strong> coordinator <strong>of</strong> <strong>the</strong> 'Klankleer' (<strong>Phonology</strong> <strong>and</strong> <strong>Phonetics</strong>)<br />
group <strong>of</strong> CLCG (Center <strong>of</strong> Language <strong>and</strong> Cognition Groningen). Therefore,<br />
his colleagues compiled this Festschrift exhibiting a diversity <strong>of</strong> research<br />
subjects on <strong>the</strong> boundaries <strong>of</strong> phonology <strong>and</strong> phonetics.<br />
It is not a goodbye to our former coordinator. Tjeerd's passionate<br />
engagement for languages <strong>and</strong> linguistic projects continues. Since his<br />
retirement he became an active honoree member at <strong>the</strong> Frisian Academy in<br />
Leeuwarden <strong>and</strong> he is still in contact with <strong>the</strong> University <strong>of</strong> St. Petersburg<br />
for future research projects. That means more than enough commitments<br />
for Tjeerd combined with his role as a gr<strong>and</strong>fa<strong>the</strong>r for his five<br />
gr<strong>and</strong>children. Tjeerd’s enthusiasm is a stimulation for o<strong>the</strong>r researchers<br />
<strong>and</strong> <strong>the</strong> young generation to continue his research.<br />
Publications by Tjeerd de Graaf<br />
1966<br />
The Annihilation <strong>of</strong> a Neutrino-antineutrino Pair into Photons <strong>and</strong> <strong>the</strong><br />
Neutrino Density in <strong>the</strong> Universe. (With H.A.Tolhoek). Nuclear physics,<br />
81: 596 <strong>and</strong> 99: 695.<br />
Neutrinoprocessen en Neutrino-astronomie [Neutrino Processes <strong>and</strong><br />
Neutrino Astronomy]. Internal Report IR 68, Natuurkundig Laboratorium<br />
Groningen. 58 pp.
Tjeerd de Graaf 21<br />
1968<br />
De Rol van het Neutrino in de Astr<strong>of</strong>ysica [The Role <strong>of</strong> <strong>the</strong> Neutrino in<br />
Astrophysics]. Ne<strong>der</strong>l<strong>and</strong>s tijdschrift voor natuurkunde, 34: 329.<br />
Phase Factors in Discrete Symmetry Operations. (With H.A.Tolhoek).<br />
Intern Rapport IR 85, Natuurkundig Laboratorium Groningen, 96 pp.<br />
Detectie van Neutrino's uit de Zon [Detection <strong>of</strong> Solar Neutrinos].<br />
Ne<strong>der</strong>l<strong>and</strong>s tijdschrift voor natuurkunde, 34: 357.<br />
1969<br />
Phase Factors in Quantum Field Theory. Physica, 43: 142.<br />
Muonen uit Kosmische Straling: het Utah Experiment [Muons from<br />
Cosmic Radiation: <strong>the</strong> Utah Experiment]. (With J. van Klinken).<br />
Ne<strong>der</strong>l<strong>and</strong>s tijdschrift voor natuurkunde, 36: 301.<br />
Aspects <strong>of</strong> Neutrino Astrophysics. Dissertation University <strong>of</strong> Groningen.<br />
Groningen. 119 pp.<br />
Syllabus Beknopte Theoretische Natuurkunde [Syllabus Summary <strong>of</strong><br />
Theoretical Physics]. Natuurkundig Laboratorium Groningen, 190 pp.<br />
1970<br />
<strong>On</strong> a Cosmic Background <strong>of</strong> Low-energy Neutrinos. Astronomy <strong>and</strong><br />
Astrophysics, 5: 335.<br />
Neutrino Processes in <strong>the</strong> Lepton Era <strong>of</strong> <strong>the</strong> Universe. Lettere al Nuovo<br />
Cimento, 4: 638.<br />
Cosmological Neutrinos. Proceedings <strong>of</strong> <strong>the</strong> Cortona Meeting on<br />
Astrophysical Aspects <strong>of</strong> <strong>the</strong> Weak Interactions, 81.<br />
1971<br />
Nucleaire Astr<strong>of</strong>ysica in het Laboratorium [Laboratory Nuclear<br />
Astrophysics]. Ne<strong>der</strong>l<strong>and</strong>s tijdschrift voor natuurkunde, 38: 107.<br />
The Astrophysical Importance <strong>of</strong> Heavy Leptons. Lettere al Nuovo<br />
Cimento, 2: 979.<br />
1972<br />
Lecture Notes on Nuclear Astrophysics. Scuola Normale Superiore, Pisa,<br />
45 pp.<br />
The Lepton Era <strong>of</strong> <strong>the</strong> Big Bang. Proceedings <strong>of</strong> <strong>the</strong> Europhysics<br />
Conference Neutrino'72. Budapest, 167.
22 Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
1973<br />
Neutrinos in <strong>the</strong> Universe. Vistas in Astronomy, 15: 161.<br />
1974<br />
Nuclear Processes in <strong>the</strong> Early Universe. VI th International Seminar on<br />
Nuclear Reactions in <strong>the</strong> Cosmos. Leningrad, 329.<br />
Kernenergie in de Kosmos [Nuclear Energy in <strong>the</strong> Cosmos]. Atoomenergie<br />
en haar Toepassingen, 81.<br />
De Heliumabundantie in het Heelal [The Helium Abundance in <strong>the</strong><br />
Universe]. (With W.J. Weeber). Ne<strong>der</strong>l<strong>and</strong>s tijdschrift voor natuurkunde,<br />
40: 183.<br />
1977<br />
De Computer en de <strong>Faculteit</strong> <strong>der</strong> Letteren [The Computer <strong>and</strong> <strong>the</strong> Faculty<br />
<strong>of</strong> Arts]. Informatiebulletin Computercommissie FdL. Groningen, 38 pp.<br />
1978<br />
Vowel Analysis with <strong>the</strong> Fast Fourier Transform. Acustica, 41: 41<br />
Ienlûd, twa lûden, twalûden [Monophthongs, Two Sounds, Diphthongs].<br />
(with G.L. Meinsma). Us Wurk, 27: 81.<br />
Analyse de voyelles avec des méthodes digitales [Vowel Analysis with<br />
Digital Methods]. Actes des 9èmes Journées d'Etude sur la Parole.<br />
Lannion, 233.<br />
Linear Prediction in Speech Research. Prace XXV Seminarium Otwartego z<br />
Akustyki. Poznań, 19.<br />
1979<br />
Het kenmerk bij hoge gespannen vokalen [The Feature <br />
in High Tense Vowels]. (With N.Streekstra). TABU, 8: 40.<br />
De Computer en Fonetisch <strong>On</strong><strong>der</strong>zoek [The Computer <strong>and</strong> Phonetic<br />
Research]. Informatiebulletin Computercommissie FdL. Groningen, 5 pp.<br />
Vowel Analysis with Linear Prediction. Proceedings <strong>of</strong> <strong>the</strong> 9th<br />
International Congress <strong>of</strong> Phonetic Sciences. Copenhagen, 265.<br />
Digital Methods for <strong>the</strong> Analysis <strong>of</strong> Speech. Proceedings <strong>of</strong> <strong>the</strong> 7th<br />
Colloquium on Acoustics. Budapest, 289.<br />
1980<br />
Phonetic Aspects <strong>of</strong> Breaking in West Frisian. (With P.Tiersma).<br />
Phonetica, 37: 109.
Tjeerd de Graaf 23<br />
De brekking fan sintralisearjende twalûden yn it Frysk [Breaking <strong>of</strong><br />
Centralizing Diphthongs in Frisian]. (With G.L. Meinsma). Us Wurk, 29:<br />
131.<br />
Vannak-e Diftongusok a Magyar Köznyelvben? [Are <strong>the</strong>re Diphthongs in<br />
St<strong>and</strong>ard Hungarian?]. (With A.D. Kylstra). Nyelvtudományi Közlemények,<br />
82: 313.<br />
Applications <strong>of</strong> Linear Predictive Coding in Speech Analysis. Proceedings<br />
<strong>of</strong> <strong>the</strong> Symposium on Speech Acoustics, 57.<br />
1981<br />
Wiskundige Modellen in het Spraakon<strong>der</strong>zoek [Ma<strong>the</strong>matical Models in<br />
Speech Research]. Wiskundige Modellen: Cursusboek Stichting TELEAC,<br />
165.<br />
Syllabegrenzen en Fonetische Experimentatie [Syllable <strong>Boundaries</strong> <strong>and</strong><br />
Phonetic Experiments]. GLOT, Tijdschrift voor Taalwetenschap, 4: 229.<br />
Book Review <strong>of</strong>: Metrical Myths – An Experimental-Phonetic<br />
Investigation into <strong>the</strong> Production <strong>and</strong> Perception <strong>of</strong> Metrical Speech.<br />
Spectator, 10: 385.<br />
1982<br />
Vowel Contrast Reduction in Japanese Compared to Dutch. (With F.J.<br />
Koopmans-van Beinum). Proceedings <strong>of</strong> <strong>the</strong> Institute <strong>of</strong> Phonetic Sciences .<br />
Amsterdam, 7: 27.<br />
A Sociophonetic Study <strong>of</strong> Language Change. Proceedings <strong>of</strong> <strong>the</strong> 13 th<br />
International Conference <strong>of</strong> Linguistics. Tokyo, 602.<br />
1983<br />
Phonetic Sciences in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s, Past <strong>and</strong> Present. (With o<strong>the</strong>r<br />
authors). Publication <strong>of</strong> <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s Association for Phonetic Sciences.<br />
Dordrecht, 32 pp.<br />
<strong>On</strong> <strong>the</strong> Reliability <strong>of</strong> <strong>the</strong> Intraoral Measuring <strong>of</strong> Subglottal Pressure. (With<br />
G.L.J. Nieboer <strong>and</strong> H.K. Schutte). Proceedings <strong>of</strong> <strong>the</strong> 10 th International<br />
Congress <strong>of</strong> Phonetic Sciences. Utrecht, 367.<br />
Phonetic Aspects <strong>of</strong> Vowels <strong>and</strong> Breaking <strong>of</strong> Diphthongs. Fifth<br />
International <strong>Phonology</strong> Meeting. Eisenstadt, 98.<br />
Vowel Contrast Reduction in Finnish, Hungarian <strong>and</strong> O<strong>the</strong>r Languages.<br />
Dritte Tagung für Uralische Phonologie. Eisenstadt, 11.
24 Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
1984<br />
Vowel Contrast Reduction in Terms <strong>of</strong> Acoustic System Contrast. (With<br />
F.J. Koopmans-van Beinum). Proceedings <strong>of</strong> <strong>the</strong> Institute <strong>of</strong> Phonetic<br />
Sciences. Amsterdam, 8: 41.<br />
Vokaalduur en Breking van Diftongen in het Fries [Vowel Duration <strong>and</strong><br />
Breaking <strong>of</strong> Diphthongs in Frisian]. Verslagen van de Ne<strong>der</strong>l<strong>and</strong>se<br />
Vereniging voor Fonetische Wetenschappen, 54.<br />
The Acoustic System Contrast <strong>and</strong> Vowel Contrast Reduction in Various<br />
Languages. Proceedings <strong>of</strong> <strong>the</strong> 23 rd Acoustic Conference on Physiological<br />
<strong>and</strong> Psychological Acoustics. Madrid, 76.<br />
Vowel Data Bases. (With A. Bladon en M. O'Kane). Speech<br />
Communication, 3: 169.<br />
Ne<strong>der</strong>l<strong>and</strong>se Leerboeken voor de Fonetiek van het Engels [Dutch Teaching<br />
Methods on <strong>the</strong> <strong>Phonetics</strong> <strong>of</strong> English]. (With A. van Essen en J.<br />
Posthumus). Toegepaste Taalwetenschap in Artikelen, 20: 123-154.<br />
1985<br />
Phonetic Aspects <strong>of</strong> <strong>the</strong> Frisian Vowel System. NOVELE, 5: 23-42.<br />
Review <strong>of</strong>: Spreken en Verstaan, een nieuwe Inleiding tot de Experimentele<br />
Fonetiek [Speaking <strong>and</strong> Un<strong>der</strong>st<strong>and</strong>ing, A New Introduction to<br />
Experimental <strong>Phonetics</strong>]. (By S. Nooteboom en A. Cohen). Logopedie en<br />
Foniatrie, 57: 106.<br />
De Groninger Button [The Groningen Button]. (With G.L.J. Nieboer <strong>and</strong><br />
H.K. Schutte). Verslagen van de Ne<strong>der</strong>l<strong>and</strong>se Vereniging voor Fonetische<br />
Wetenschappen, 57-62.<br />
1986<br />
S<strong>and</strong>hi Phenomena in West Frisian. (With G. van <strong>der</strong> Meer). S<strong>and</strong>hi<br />
Phenomena in <strong>the</strong> Languages <strong>of</strong> Europe. Berlin, 301-328.<br />
Review <strong>of</strong>: The Production <strong>of</strong> Speech. (By P.F. MacNeilage). Studies in<br />
Language, 10: 273-277.<br />
Production <strong>of</strong> Different Types <strong>of</strong> Esophageal Voice Related to <strong>the</strong> Quality<br />
<strong>and</strong> <strong>the</strong> Intensity <strong>of</strong> <strong>the</strong> Sound Produced. Folia Phoniatrica, 38: 292.<br />
De Uitspraak van het Ne<strong>der</strong>l<strong>and</strong>s door Buitenl<strong>and</strong>ers [The Pronunciation <strong>of</strong><br />
Dutch by Foreigners]. Logopedie en Foniatrie, 58: 343-349.<br />
Sociophonetic Aspects <strong>of</strong> Frisian. Friser Studier IV/V. Odense, 3-21.<br />
Een contrastief fonetisch on<strong>der</strong>zoek Japans-Ne<strong>der</strong>l<strong>and</strong>s [A Contrastive<br />
Phonetic Research Japanese-Dutch]. Verslagen van de Ne<strong>der</strong>l<strong>and</strong>se<br />
Vereniging voor Fonetische Wetenschappen, 15-24.
Tjeerd de Graaf 25<br />
1987<br />
The Retrieval <strong>of</strong> Dialect Material from Old Phonographic Wax Cylin<strong>der</strong>s.<br />
Proceedings <strong>of</strong> <strong>the</strong> Workshop on “New Methods in Dialectology”.<br />
Amsterdam, 117-125.<br />
Acoustic <strong>and</strong> Physiological Properties <strong>of</strong> <strong>the</strong> Laryngeal <strong>and</strong> Alaryngeal<br />
(Esophageal) Voice. Proceedings <strong>of</strong> <strong>the</strong> XXXIV th Open Seminar on<br />
Acoustics. Wrocław, 10-16.<br />
A Contrastive Study <strong>of</strong> Japanese <strong>and</strong> Dutch. Proceedings <strong>of</strong> <strong>the</strong> XI th<br />
International Congress <strong>of</strong> Phonetic Sciences. Tallinn, 124-128.<br />
1988<br />
His Master's Voice: Herkenning van de Spraakmaker [His Master’s Voice:<br />
Recognition <strong>of</strong> <strong>the</strong> Speech Producer]. TER SPRAKE: SPRAAK als<br />
betekenisvol geluid in 36 <strong>the</strong>matische ho<strong>of</strong>dstukken. Dordrecht, 200-208.<br />
Book Review: Fonetiek en Fonologie [<strong>Phonetics</strong> <strong>and</strong> <strong>Phonology</strong>]. (By R.<br />
Collier en F.G. Droste). Logopedie en Foniatrie, 60: 195.<br />
The Frisian Language in America. (With T. Anema <strong>and</strong> H. Schatz).<br />
NOWELE, 6: 91-108.<br />
Esophageal Voice Quality Judgements by Means <strong>of</strong> <strong>the</strong> Semantic<br />
Differential. (With G.L.J. Nieboer <strong>and</strong> H.K. Schutte). Journal <strong>of</strong> <strong>Phonetics</strong>,<br />
16: 417-436.<br />
Book Review: Sprechererkennung [Speaker Recognition]. (By Hermann J.<br />
Künzel). Journal <strong>of</strong> <strong>Phonetics</strong>, 16: 459-463.<br />
1989<br />
Reconstruction, Signal Enhancement <strong>and</strong> Storage <strong>of</strong> Sound Material in<br />
Japan. Proceedings <strong>of</strong> <strong>the</strong> 2 nd International Conference on Japanese<br />
Information in Science, Technology <strong>and</strong> Commerce. Berlin, 367-374.<br />
Aerodynamic <strong>and</strong> Psycho-acoustic Properties <strong>of</strong> Esophageal Voice<br />
Production. (With G.L.J. Nieboer <strong>and</strong> H.K. Schutte). Proceedings <strong>of</strong> <strong>the</strong><br />
Conference on Speech Research '89. Budapest, 53-58.<br />
A Data Base <strong>of</strong> Old Sound Material. Proceedings <strong>of</strong> <strong>the</strong> ESCA Workshop<br />
on Speech Input/Output Assessment <strong>and</strong> Speech Data Bases. Noordwijk,<br />
2.14.1-5.<br />
1990<br />
Een contrastief fonetisch on<strong>der</strong>zoek, in het bijzon<strong>der</strong> Japans-Ne<strong>der</strong>l<strong>and</strong>s<br />
[Contrastive Phonetic Research, in Particular Japanese-Dutch]. Neerl<strong>and</strong>ica<br />
Wratislaviensia IV. Wrocław, 140-148.
26 Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
Book Review: To Siberia <strong>and</strong> Russian America, Three Centuries <strong>of</strong> Russian<br />
Eastward Expansion. Circumpolar Journal, 7: 41-46.<br />
New Technologies in Sound Reconstruction <strong>and</strong> <strong>the</strong>ir Applications to <strong>the</strong><br />
Study <strong>of</strong> <strong>the</strong> Smaller Languages <strong>of</strong> Asia. Proceedings <strong>of</strong> <strong>the</strong> IV th<br />
International Symposium “Uralische Phonologie”. Hamburg, 15-19.<br />
GARASU-GLAS: Fonetische contrasten Japans-Ne<strong>der</strong>l<strong>and</strong>s [GARASU-<br />
GLAS: Phonetic Contrasts Japanese-Dutch]. TABU. Bulletin voor<br />
Taalwetenschap, 20: 49-57.<br />
1991<br />
Aerodynamic <strong>and</strong> Phonetic Properties <strong>of</strong> Voice Production with <strong>the</strong><br />
Groningen Button. TENK jaarboek, 91-97.<br />
Laser-beam Technology in Diachronic Phonetic Research <strong>and</strong><br />
Ethnolinguistic Field Work. Proceedings <strong>of</strong> <strong>the</strong> XII th International<br />
Congress <strong>of</strong> Phonetic Sciences. Amsterdam, 114-118.<br />
Laut aus Wachs: Der Übergang von st<strong>of</strong>fgebundenen zum elektronischen<br />
und optischen Informationstransport [Sound from Wax: The Transition<br />
from Material-Bound to Electronic <strong>and</strong> Optic Information Transport]. TU<br />
International. Berlin, 14/15: 63-66.<br />
1992<br />
The Languages <strong>of</strong> Sakhalin. Small Languages <strong>and</strong> Small Language<br />
Communities: News, Notes, <strong>and</strong> Comments. International Journal <strong>of</strong> <strong>the</strong><br />
Sociology <strong>of</strong> Languages, 94: 185-200.<br />
Dutch Encounters with Sakhalin <strong>and</strong> with <strong>the</strong> Ainu People. Proceedings <strong>of</strong><br />
<strong>the</strong> International Conference 125 th anniversary <strong>of</strong> <strong>the</strong> birth <strong>of</strong> Bronis�aw<br />
Pi�sudski. Sapporo, 108-137.<br />
The Ethnolinguistic Situation on <strong>the</strong> Isl<strong>and</strong> <strong>of</strong> Sakhalin. Circumpolar<br />
Journal, 6: 32-58.<br />
Aerodynamische en fonetische eigenschappen van verschillende soorten<br />
slokdarmstem [Aerodynamic <strong>and</strong> Phonetic Features <strong>of</strong> Different Kinds <strong>of</strong><br />
Esophageal Voice]. (With G.L.J. Nieboer <strong>and</strong> H.K. Schutte). Klinische<br />
Fysica, 8: 64-66.<br />
The Dutch Role in <strong>the</strong> Bor<strong>der</strong> Area between Japan <strong>and</strong> Russian. Round Table<br />
Conference “The Territorial Problem in Russo-Japanese Relations”.<br />
Moscow, 20-26.<br />
De Taal <strong>der</strong> Mennonieten [The Language <strong>of</strong> <strong>the</strong> Mennonites]. Syllabus<br />
NOMES Symposium Groningen, 42 pp.
Tjeerd de Graaf 27<br />
1993<br />
Saharin ni okeru shoosuu minzoku no gengo jookyoo [The Status <strong>of</strong><br />
Minority Languages on Sakhalin]. (With K. Murasaki). Japanese Scientific<br />
Monthly, 46: 18-24.<br />
The Ethnolinguistic Situation on <strong>the</strong> Isl<strong>and</strong> <strong>of</strong> Sakhalin. Ethnic minorities<br />
on Sakhalin. Yokohama, 13-32.<br />
Vstrechi goll<strong>and</strong>tsev c Sakhalinom i Ainami [Meetings <strong>of</strong> <strong>the</strong> Dutch with<br />
Sakhalin <strong>and</strong> <strong>the</strong> Ainu Population]. Proceedings <strong>of</strong> <strong>the</strong> International<br />
Conference “B.O. Pilsudski - issledovatel' narodov Sakhalina”. Yuzhno-<br />
Sakhalinsk, 92-99.<br />
De taal <strong>der</strong> Mennonieten in Siberië en hun relatie met Ne<strong>der</strong>l<strong>and</strong> [The<br />
Language <strong>of</strong> <strong>the</strong> Siberian Mennonites <strong>and</strong> <strong>the</strong>ir Relation with <strong>the</strong><br />
Ne<strong>the</strong>rl<strong>and</strong>s]. (With R. Nieuweboer). Doopsgezinde Bijdragen, 19: 175-<br />
189.<br />
Languages <strong>and</strong> Cultures <strong>of</strong> <strong>the</strong> Arctic Region in <strong>the</strong> Former Soviet Union.<br />
(With R. Nieuweboer). Circumpolar Journal, 1-2: 29-42.<br />
1994<br />
The Dutch Role in <strong>the</strong> Bor<strong>der</strong> Area between Japan <strong>and</strong> Russia.<br />
Circumpolar Journal, 3-4: 1-12.<br />
Ne<strong>der</strong>l<strong>and</strong>s in Siberië [Dutch in Siberia]. (With R. Nieuweboer). TABU<br />
Taalkundig Bulletin, 24: 65-75.<br />
The Language <strong>of</strong> <strong>the</strong> West Siberian Mennonites. (With R. Nieuweboer).<br />
RASK, Internationalt tidsskrift for sprog og kommunikation, 1: 47-63.<br />
1995<br />
Het territoriale geschil tussen Japan en Rusl<strong>and</strong> over de Koerilen [The<br />
Territorial Dispute between Japan <strong>and</strong> Russia about <strong>the</strong> Kuril Isl<strong>and</strong>s].<br />
(With I. van Oosteroom). Internationale Spectator, 49: 41-46.<br />
Dutch Encounters with Sakhalin <strong>and</strong> with <strong>the</strong> Ainu People. Linguistic <strong>and</strong><br />
�������� ������� ���� ������, 35-61.<br />
The Language <strong>of</strong> <strong>the</strong> West Siberian Mennonites. (with R. Nieuweboer).<br />
Proceedings <strong>of</strong> <strong>the</strong> XIII th Congress <strong>of</strong> <strong>Phonetics</strong> Sciences. Stockholm, 4:<br />
180-184.<br />
Pitch Stereotypes in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s <strong>and</strong> Japan. (With R. van Bezooijen<br />
<strong>and</strong> T. Otake). Proceedings <strong>of</strong> <strong>the</strong> XIII th Congress <strong>of</strong> Phonetic Sciences.<br />
Stockholm, 680-684.<br />
The Reconstruction <strong>of</strong> Acoustic Data on <strong>the</strong> Ethnic Minorities <strong>of</strong> Siberia.<br />
Proceedings <strong>of</strong> <strong>the</strong> International Conference on “The Indigenous Peoples
28 Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
<strong>of</strong> Siberia: Studies <strong>of</strong> Endangered Languages <strong>and</strong> Cultures”. Novosibirsk,<br />
1: 381-383.<br />
1996<br />
Book Review <strong>of</strong>: Joshua A. Fishman, Yiddish: Turning to Life. Studies in<br />
Language, 20,1: 191-196.<br />
Language Minorities in <strong>the</strong> Sakha Republic (Yakutia). Report Nagoya City<br />
University. Nagoya, 165-179.<br />
Dutch Encounters with <strong>the</strong> Peoples <strong>of</strong> Eastern Asia. A Frisian <strong>and</strong> Germanic<br />
Miscellany, published in Honour <strong>of</strong> Nils Århammar on his Sixty-Fifth<br />
Birthday. Odense, 377-386.<br />
Dutch Immigrants in Siberia? The Language <strong>of</strong> <strong>the</strong> Mennonites. Charisteria<br />
doctissimo Přemysl Janota oblata, Acta Universitatis Carolinae<br />
viro<br />
Philologica. Prague, 75-86<br />
Archives <strong>of</strong> <strong>the</strong> Languages <strong>of</strong> Russia. (With L.V. Bondarko). Reports on <strong>the</strong><br />
INTAS Project No. 94-4758. St.-Petersburg, 120 pp.<br />
1997<br />
The Reconstruction <strong>of</strong> Acoustic Data <strong>and</strong> Minority Languages in Russia.<br />
Proceedings <strong>of</strong> <strong>the</strong> 2 nd International Congress <strong>of</strong> Dialectologists <strong>and</strong><br />
Geolinguists. Amsterdam.,44-54.<br />
Language <strong>and</strong> Culture <strong>of</strong> <strong>the</strong> Russian Mennonites. Around Peter <strong>the</strong> Great.<br />
Three Centuries <strong>of</strong> Russian-Dutch Relations. Groningen, 132-142.<br />
Resten van het Jiddisch in Groningen en Sint-Petersburg [Remnants <strong>of</strong> <strong>the</strong><br />
Yiddish Language in Groningen <strong>and</strong> Saint-Petersburg]. VDW-berichten,<br />
Vereniging voor Dialectwetenschap, 1: 6-7.<br />
The Reconstruction <strong>of</strong> Acoustic Data <strong>and</strong> <strong>the</strong> Study <strong>of</strong> Language Minorities<br />
in Russia. Language Minorities <strong>and</strong> Minority Language. ������� ��������<br />
1998<br />
Linguistic Databases <strong>and</strong> Language Minorities around <strong>the</strong> North Pacific<br />
Rim. Lecture on <strong>the</strong> Occasion <strong>of</strong> <strong>the</strong> Doctorate Honoris Causa, St.-<br />
Petersburg, 14 pp.<br />
Linguistic Databases: A Link between Archives <strong>and</strong> Users. Journal <strong>of</strong> <strong>the</strong><br />
International Association <strong>of</strong> Sound Archives, 27-34.
Tjeerd de Graaf 29<br />
1999<br />
Russian-Yiddish: Phonetic Aspects <strong>of</strong> Language Interference. (With N.<br />
Svetozarova, Yu. Kleiner <strong>and</strong> R. Nieuweboer). Proceedings <strong>of</strong> <strong>the</strong> 14 th<br />
International Congress <strong>of</strong> Phonetic Sciences. San Francisco., 1397-1401.<br />
Language Contact <strong>and</strong> Sound Archives in Russia. (With L. Bondarko).<br />
Proceedings <strong>of</strong> <strong>the</strong> 14 th International Congress <strong>of</strong> Phonetic Sciences. San<br />
Francisco, 1401-1404.<br />
Lingvisticheskie bazy dannykh i yazykovye men’shinstva po obeim storonam<br />
severnogo tikho-okeanskogo poyasa [Linguistic databases <strong>and</strong> language<br />
minorities at both sides <strong>of</strong> <strong>the</strong> North-Pacific Rim]. Yazyk i rechevaya<br />
deyatel’nost’, 2: 8-18.<br />
2000<br />
Scientific Links between Russia <strong>and</strong> The Ne<strong>the</strong>rl<strong>and</strong>s: A Study <strong>of</strong> <strong>the</strong><br />
Languages <strong>and</strong> Cultures in <strong>the</strong> Russian Fe<strong>der</strong>ation. Proceedings <strong>of</strong> <strong>the</strong><br />
Conference on <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s <strong>and</strong> <strong>the</strong> Russian North. Arkhangelsk. To be<br />
published.<br />
The Language <strong>of</strong> <strong>the</strong> Siberian Mennonites. (With R. Nieuweboer). New<br />
Insights in Germanic Linguistics II. Frankfurt am Main, 21-34.<br />
2001<br />
Nivkh <strong>and</strong> Kashaya: Two endangered Languages in Contact with Russian <strong>and</strong><br />
English. Materialy mezhdunarodnoy konferentsii “100 let eksperimental’noy<br />
fonetike v Rossii”. St.-Petersburg, 78-83.<br />
Data on <strong>the</strong> Languages <strong>of</strong> Russia from Historical Documents, Sound Archives<br />
<strong>and</strong> Fieldwork Expeditions. Recording <strong>and</strong> Restoration <strong>of</strong> Minority<br />
Languages, Sakhalin Ainu <strong>and</strong> Nivkh, ELPR Report A2-009. Kyoto, 13-37.<br />
Kashaya Pomo <strong>and</strong> <strong>the</strong> Russian Influence around <strong>the</strong> North Pacific. Materials<br />
��� ����� ������������� ���������� �� ��������� ��������� ��� ���<br />
��<br />
Scholarly Heritage. Kraków, 385-395.<br />
2002<br />
Yazyk i etnos [Language <strong>and</strong> Ethnos]. (With A.S. Gerd <strong>and</strong> M. Savijärvi).<br />
Texts <strong>and</strong> Comments on Balto-Finnic <strong>and</strong> Northwestern Archaic Russian<br />
Dialects. St.-Petersburg, 206 pp.<br />
Voices from Tundra <strong>and</strong> Taiga: Endangered Languages in Russia on <strong>the</strong><br />
Internet. Conference H<strong>and</strong>book on Endangered Languages. Kyoto, 57-79.
30 Markus Bergmann, Nynke de Graaf <strong>and</strong> Hidetoshi Shiraishi<br />
Phonetic Aspects <strong>of</strong> <strong>the</strong> Frisian Language <strong>and</strong> <strong>the</strong> Use <strong>of</strong> Sound Archives.<br />
Problemy i metody eksperimental’no-foneticheskikh issledovaniy. St.-<br />
Peterburg, 52-57.<br />
Voices from <strong>the</strong> Shtetl: The Past <strong>and</strong> Present <strong>of</strong> <strong>the</strong> Yiddish Language in<br />
Russia. Final Report NWO Russian-Dutch Research Cooperation.<br />
Groningen, 143 pp.<br />
The Use <strong>of</strong> Sound Archives in <strong>the</strong> Study <strong>of</strong> Endangered Languages. Music<br />
Archiving in <strong>the</strong> World, Papers Presented at <strong>the</strong> Conference on <strong>the</strong><br />
Occasion <strong>of</strong> <strong>the</strong> 100 th Anniversary <strong>of</strong> <strong>the</strong> Berlin Phonogramm-Archiv.<br />
Berlin, 101-107.<br />
The Use <strong>of</strong> Acoustic Databases <strong>and</strong> Fieldwork for <strong>the</strong> Study <strong>of</strong> <strong>the</strong><br />
Endangered Languages <strong>of</strong> Russia. Proceedings <strong>of</strong> <strong>the</strong> International LREC<br />
Workshop on Resources <strong>and</strong> Tools in Field Linguistics. Las Palmas, 29.1-4<br />
(CD-ROM).<br />
Yiddish in St.-Petersburg: The Last Sounds <strong>of</strong> a Language. Proceedings <strong>of</strong><br />
<strong>the</strong> Conference “Klezmer, Klassik, jiddisches Lied. Jüdische Musik-Kultur<br />
in Osteuropa.”. Potsdam. To be published.<br />
2003<br />
Yazyki severnoy i vostochnoy Tartarii – o yazykovykh svedeniyakh v<br />
knige N. Vitsena [The Languages <strong>of</strong> North <strong>and</strong> East Tartary – About <strong>the</strong><br />
Linguistic Data in <strong>the</strong> Book <strong>of</strong> N. Witsen]. (With M. Bergmann).<br />
Proceedings <strong>of</strong> <strong>the</strong> Conference on General Linguistics. St.-Petersburg. To<br />
be published.<br />
Description <strong>of</strong> Minority Languages in Russia on <strong>the</strong> Basis <strong>of</strong> Historical<br />
Data <strong>and</strong> Fieldwork. Proceedings <strong>of</strong> <strong>the</strong> XVI th International Congress <strong>of</strong><br />
Linguists. Prague. To be published.<br />
Voices <strong>of</strong> Tundra <strong>and</strong> Taiga: Data on Minority Languages in Russia from<br />
Historical Data <strong>and</strong> Fieldwork. Proceedings <strong>of</strong> <strong>the</strong> Conference “Formation<br />
<strong>of</strong> Educational Programs Aimed at a New Type <strong>of</strong> Humanitarian Education<br />
in Siberian Polyethnic Society, Novosibirsk. To be published.<br />
Endangered Languages in Europe <strong>and</strong> Siberia: State <strong>of</strong> <strong>the</strong> Art, Needs <strong>and</strong><br />
Solutions. International Expert Meeting on UNESCO Programme<br />
“Safeguarding <strong>of</strong> Endangered Languages”. Paris. To be published<br />
Presentation <strong>of</strong> <strong>the</strong> UNESCO Document “Language Vitality <strong>and</strong><br />
Endangerement”. Focus on Linguistic Diversity in <strong>the</strong> New Europe.<br />
European Bureau for Lesser Used Languages, Brussels. To be published.
Tseard de Graaf<br />
Oerset troch Jurjen van <strong>der</strong> Kooi<br />
Tseard de Graaf is berne op 27 jannewaris 1938 yn Ljouwert, de haadstêd<br />
fan de Ne<strong>der</strong>lânske provinsje Fryslân, de grutste regio oan de kusten fan de<br />
Noardsee dêr’t it Frysk, in Westgermaanske taal mei as neiste sibbe it<br />
Ingelsk, sprutsen wurdt.<br />
Tseard syn âlden wiene beide Fries en thús waard allinne Frysk praat.<br />
Syn earste taal wie dan ek it Frysk; op skoalle learde er it Ne<strong>der</strong>lânsk, de<br />
<strong>of</strong>fisjele taal fan Ne<strong>der</strong>lân.<br />
It Frysk waard doe noch net op skoalle jûn en de bern moasten dêr<br />
Ne<strong>der</strong>lânsk leare. Troch it neistinoar fan it Frysk thús en it Ne<strong>der</strong>lânsk op<br />
skoalle kaam Tseard foar it earst yn ’e kunde mei de fassinearjende wrâld<br />
fan ûn<strong>der</strong>skate talen. Foar Tseard hie it ûn<strong>der</strong>skie tusken dy twa talen ek in<br />
besûn<strong>der</strong> romtlik aspekt. As hy en de oare bern út syn buert de moarns nei<br />
skoalle ta gyngen moasten se oer it spoar. Wiene se dêr oer, dan giene se<br />
fan it Frysk oer op it Ne<strong>der</strong>lânsk, de taal fan de skoalle.<br />
18 Jier âld wie er, doe’t er yn 1956 eineksamen middelbere skoalle die<br />
en hy krige niget oan talen. Syn oare grutte leafdes wiene natuer- en<br />
stjerrekunde. It âldste planetarium fan de wrâld is yn Frjentsjer, de âlde<br />
akademystêd fan Fryslân. Yn ’e besnijing fan de wetten dy’t tiid en romte<br />
regeare, studearre Tseard fan 1956 oant 1963 natuerkunde oan de<br />
Universiteit fan Grins. Yn dat lêste jier die er doktoraal eksamen teoretyske<br />
natuerkunde, in kombinaasje fan natuerkunde, wiskunde en stjerrekunde en<br />
dêrnei wie er oant 1969 ûn<strong>der</strong>sykmeiwurker oan it Ynstitút foar Teoretyske<br />
Natuerkunde fan de Grinzer universiteit.<br />
Tseard wie doe al in ‘polyglot’. Hy spruts net allinne Frysk en<br />
Ne<strong>der</strong>lânsk, mar ek Dútsk, Ingelsk en Frânsk. Oare talen soene folgje. Yn<br />
de Sowjet Uny stie de stúdzje fan de astronomy doe op in superieur nivo.<br />
Tseard seach dat en learde it Russysk en oare Eastjeropeeske talen om’t dy<br />
it paad nei nij ynsjoch yn dy fjilden fan wittenskip iepenleinen. Neist syn<br />
stúdzje fan de teoretyske natuerkunde folge er kolleezjes yn de Slavyske<br />
talen. Hy wie fassinearre troch de nije technologyen en har<br />
tapassingsmooglikheden foar takomstich ûn<strong>der</strong>syk en yn 1967 die er<br />
k<strong>and</strong>idaatseksamen Slavyske talen en kompjutertaalkunde. Yntysken wie er
32 Oerset troch Jurjen van <strong>der</strong> Kooi<br />
nei syn doktoraal natuerkunde trochgien mei syn ûn<strong>der</strong>syk yn de teoretyske<br />
natuerkunde, dat er kombinearre mei in heal jier stúdzje yn Poalen, dêr’t er<br />
ek it Poalsk by learde.<br />
Yn 1969 wie er klear mei syn dissertaasje, titele: “Aspects <strong>of</strong> Neutrino<br />
Astrophysics”.<br />
It titelblêd fan Tseard syn dissertaasje teoretyske natuerkunde út 1969<br />
Syn ûndwêstbere toarst nei witten brocht Tseard nei Ingelân, dêr’t er mei<br />
frou en bern fan 1970 oant 1971 in jier tabrocht en dêr’t er wurke as<br />
ûn<strong>der</strong>siker oan it Ynstitút foar Teoretyske Natuerkunde fan de Universiteit<br />
fan Cambridge.<br />
Werom yn Grins waard Tseard universitêr meiwurker natuerkunde oan<br />
it Ynstitút foar Astronomy. Oant 1975. Dat jier waard in kearpunt yn syn<br />
wittenskiplike karriêre om’t er besleat de wei fan syn twadde grutte leafde<br />
te gean, dy fan de bestudearring fan talen. Ien fan de stellingen by syn<br />
dissertaasje gie oer de fraach, hoe krekt <strong>of</strong> immens identiteit definiearre<br />
wurde kin troch syn <strong>of</strong> har taal. Dy stelling kin sjoen wurde as in<br />
symboalyske paadwizer nei syn lettere wei yn de taalkunde, dy’t him liede<br />
soe nei de stúdzje fan aspekten fan de sprutsen taal, nei de fonetyk.<br />
Yn 1975 waard Tseard meiwurker oan it Ynstitút foar Fonetyk fan de<br />
Literêre Fakulteit fan de Grinzer universiteit.<br />
Om’t er fan jongs ôf oan twatalich wie (Frysk-Ne<strong>der</strong>lânsk) hie Tseard in<br />
skerp each foar de ûntelbere fonetyske ferskillen tusken dy talen. En om’t<br />
er oare talen bestudearre hie, wist er hoe wichtich fonetyske beskriuwingen<br />
binne, net allinne foar de teoretyske taalkunde, mar likegoed ek foar it<br />
learen fan en it lesjaan yn frjemde talen.
Tseard de Graaf 33<br />
It neistinoar fan talen en taalferoaring soene oare swiertepunten fan syn<br />
ûn<strong>der</strong>syk wurde. Rûnom yn ’e wrâld binne minsken twa- <strong>of</strong> sels meartalich.<br />
Der is taalfariaasje yn romte èn yn tiid. Harket men nei radioprogramma’s<br />
<strong>of</strong> tillevyzje-útstjoerings fan tsien <strong>of</strong> twintich jier lyn, dan heart men in oare<br />
sprektaal as at no gongber is. It is noch altiten deselde taal en itselde plak,<br />
en dochs is de taal net mear gelyk. Net allinne it leksikon fan in taal<br />
feroaret, mar ek minskene wize fan sprekken, de útspraak en de yntonaasje,<br />
in útsûn<strong>der</strong>lik nijsgjirrich ûn<strong>der</strong>werp foar immen dy’t niget hat oan talen en<br />
har fariabiliteit.<br />
Tseard begûn mei in syktocht nei de âldste registraasjes op lûddragers<br />
fan sprutsen taal. Hy analysearre materiaal net allinne út Westerlauwersk<br />
mar ek út Noard- en Eastfryslân. Sokke registraasjes fan ear<strong>der</strong>e sprutsen<br />
taal binne net allinne wichtich histoarysk erfguod, mar se jouwe ek<br />
weardefolle ynformaasje oange<strong>and</strong>e taalferoaringsprosessen. In praktysk<br />
probleem by dy âldste lûdregistraasjes is dat se makke binne op<br />
waakssilin<strong>der</strong>s en dat de kwaliteit hurd ôfnimt elke kear as se beharke<br />
wurde. Tseard seach yn dat it fan it grutste belang is en bring dizze<br />
registraasjes oer op mo<strong>der</strong>ne lûddragers, sadat se bewarre bliuwe. Yn it<br />
begjin fan de jierren 90 begûn Tseard mei kollega’s út Japan in ûn<strong>der</strong>syk<br />
nei de mooglikheden dêrta. Yn dy tiid makke er him noch in taal eigen, it<br />
Japansk, dat er floeiend sprekken learde.<br />
Tseard oan ’e skrep mei waakssilin<strong>der</strong>s mei âl<strong>der</strong> Ne<strong>der</strong>lânsk<br />
Tseard socht kontakt mei de wichtichste lûdargiven yn de wrâld, dy yn<br />
Wenen, Berlyn en Sint Petersboarch. Troch syn oparbeidzjen yn de 90er<br />
jierren mei it lûdargyf fan de Akademy fan Wittenskippen yn dy lêste stêd
34 Oerset troch Jurjen van <strong>der</strong> Kooi<br />
luts er op ’en nij de relaasjes oan mei Ruslân, dy’t yn de 60er jierren begûn<br />
wiene mei syn bestudearring fan de Slavyske talen.<br />
Sûnt 1990 is wrâld dramatysk feroare. It Izeren Gerdyn is <strong>der</strong> net mear<br />
en Ruslân hat opnij syn ‘Finster op it Westen’ iepenset. Doe’t Tseard nei<br />
1990 weromkaam yn Sint Petersboarch rekke er daliken fassinearre troch<br />
dizze stêd dy’t er foar it earst likernôch 20 jier lyn, doe’t er noch Leningrad<br />
hiet, sjoen hie. As Fries en Ne<strong>der</strong>lanner fielde er him <strong>der</strong> thús. De<br />
skil<strong>der</strong>eftige kanalen en paden lâns de wide bûlevaren diene him tinke oan<br />
thús. Dat wie gjin tafal: tsaar Peter de Grutte hie sa’n 300 jier ear<strong>der</strong> Hollân<br />
keazen as model foar syn nije haadstêd.<br />
Tseard organisearre no mienskiplike projekten mei de Russyske<br />
Akademy fan Wittenskippen en de Steatsuniversiteit fan Sint Petersboarch.<br />
It doel wie âlde Russyske lûdregistraasjes te bewarjen en oer te setten op<br />
mo<strong>der</strong>ne digitale audio media. Un<strong>der</strong>syk nei in grutte samling fan<br />
al<strong>der</strong>h<strong>and</strong>e lûdregistraasjes, resultaat fan withoe folle linguïstyske<br />
ekspedysjes fan ein 19de en út de 20ste ieu, brocht nije projekten<br />
ûn<strong>der</strong>skate yn Ruslân spruten talen oanbelangjende op ’e gleed.<br />
Sels sette er útein mei ûn<strong>der</strong>syk nei de taal fan de Sibearyske<br />
Mennoniten, dêr’t it komôf fan socht wurde moat yn noardlik Ne<strong>der</strong>lân en<br />
Dútslân en dy’t noch altiten de taal fan de foarâlden sprekke – feitliken in<br />
taal dy’t gâns hat fan de dialekten fan it hjoeddeiske Noard-Dútslân en de<br />
noardlike parten fan Ne<strong>der</strong>lân. De Ne<strong>der</strong>lânske parse kaam sels mei de kop<br />
“Sibeariërs sprekke Grinzers”.<br />
Talen skiede net allinne folken en naasjes, se bouwe <strong>der</strong> ek brêgen<br />
tusken. Tseard liet ek dat mei syn ûn<strong>der</strong>syk sjen. Sels yn it fiere Sibearje<br />
wenje minsken dy’t likernôch deselde taal hawwe as de minsken yn<br />
Grinslân. By it plannen fan syn ekspedysjes tocht Tseard net allinne oan de<br />
wittenskip, mar ek om de minsken: hy organisearre ek humanitêre help út<br />
Grins wei foar de doarpen dy’t er yn Sibearje o<strong>and</strong>ie.<br />
Taal as kultureel erfskip waard de kearn fan syn linguïstyske<br />
aktiviteiten. Troch syn twatalich komôf koe er in treflik foarbyld jaan. Syn<br />
hiele libben hat er sjen litten dat elk yndividu bydrage kin oan it oerlibjen<br />
fan in taal. Mei syn frou Nynke, dy’t er met hie yn syn studintetiid en dêr’t<br />
er lang allinne Ne<strong>der</strong>lânsk mei praat hie, praatte er no ôf om oer te stappen<br />
op it Frysk. Nynke is sels in Friesinne en nei de dea fan har âlden murken<br />
hja sels hoe't in taal stadichoan út begjint te stjerren as de bern him net<br />
fier<strong>der</strong> trochjouwe.<br />
Dy taalhâlding waard de rjochtline foar Tseard syn opienfolgjende<br />
ûn<strong>der</strong>sykaktiviteiten yn Ruslân. De projekten dy’t er fan dat stuit ôf oan
Tseard de Graaf 35<br />
koördinearre krigen twa doelstellings: bedrige talen net allinne<br />
dokumintearje, mar ek revitalisearje en yn stân hâlde foar kommende<br />
generaasjes. Tseard die mei oan ferskate ekspedysjes, ûn<strong>der</strong> oaren nei<br />
Yakutia en it eilân Sakhalin, wêr’t er mei oare linguïsten de talen fan de<br />
lokale folken fêstlei.<br />
Tseard de Graaf mei sprekkers fan talen fan it eilân Sakhalin yn it fiere easten fan<br />
Ruslân: Uiltas en Nivkhs (jierren 90)<br />
Yn de twadde helte fan de 90er jierren koördinearre Tseard ûn<strong>der</strong>skate<br />
projekten mei ynstituten rûnom yn de Russyske Fe<strong>der</strong>aasje, foar de<br />
finansearring soargen Ne<strong>der</strong>lânske wittenskiplike organisaasjes en de<br />
INTAS fan de EU.<br />
Al<strong>der</strong>earsten woe er jonge minsken bybringe dat harren taal in unyk<br />
erfskip is en dat hja min<strong>der</strong>heids- en regionale talen stypje moatte. Yn 1998<br />
waard Tseard beneamd ta rid<strong>der</strong> yn de oar<strong>der</strong> fan de Ne<strong>der</strong>lânske liuw<br />
fanwegens syn ûn<strong>der</strong>syk nei en krewearjen foar it behâld fan en it opsetten<br />
fan databanken foar de min<strong>der</strong>heidstalen yn Ruslân. Letter datselde jiers<br />
krige er in earedoktoraat fan de Universiteit fan Sint Petersboarch foar syn<br />
bydragen oan de mienskiplike taalbehâldprojekten.
36 Oerset troch Jurjen van <strong>der</strong> Kooi<br />
Tseard de Graaf earedoktor oan de Universiteit fan Sint Petersboarch, novimber<br />
1998.<br />
Tseard moast yn 2003 mei pinsjoen en syn plak as koördinator fan de<br />
ôfdieling 'Klanklear' (Fonology en Fonetyk) fan it CLGC (Center <strong>of</strong><br />
Language <strong>and</strong> Cognition Groningen) fan de Grinzer Universiteit opjaan. Ta<br />
dy gelegenheid ha syn kollega's dizze earebondel mei in ferskaat oan<br />
bydragen oer ûn<strong>der</strong>syk yn de grinsgebieten fan fonology en fonetyk<br />
gearstald.<br />
Lykwols, it is gjin ôfskie fan ús ear<strong>der</strong>e koördinator. Tseard syn<br />
pasjonearre belutsenens by talen en linguïstyske projekten is bleaun. Sûnt<br />
syn pinsjoen is er aktyf as honorêr meiwurker fan de Fryske Akademy yn<br />
Ljouwert en ek it kontakt mei de Universiteit fan Sint Petersboarch oer<br />
ûn<strong>der</strong>syksprojekten dy't op kommende wei binne is bleaun. Dat betsjut dat<br />
Tseard neist syn rol as pake foar syn fiif pakesizzers noch genôch te dwaan<br />
hat. Syn entûsjasme is in oantrún foar (kommende) ûn<strong>der</strong>sikers om fier<strong>der</strong><br />
te gean mei it ûn<strong>der</strong>syk dat hy op priemmen set hat.
Boundary Tones in Dutch: Phonetic or Phonological<br />
Contrasts?<br />
Vincent J. van Heuven<br />
1. Introduction 1<br />
1.1. Linguistic categorization <strong>of</strong> sound<br />
A basic problem <strong>of</strong> linguistic phonetics is to explain how <strong>the</strong> infinite<br />
variety <strong>of</strong> speech sounds in actual utterances can be described with finite<br />
means, such that <strong>the</strong>y can be dealt with in <strong>the</strong> grammar, i.e. phonology, <strong>of</strong> a<br />
language. The crucial concept that was developed to cope with this<br />
reduction problem is <strong>the</strong> sound category, or – when applied to <strong>the</strong><br />
description <strong>of</strong> segmental phenomena – <strong>the</strong> phoneme. This is best conceived<br />
<strong>of</strong> as an abstract category that contains all possible sounds that are mutually<br />
interchangeable in <strong>the</strong> context <strong>of</strong> a minimal word pair. That is, substitution<br />
<strong>of</strong> one token (allophone) <strong>of</strong> a phoneme for an o<strong>the</strong>r does not yield a<br />
different word (i.e., a string <strong>of</strong> sounds with a different lexical meaning). 2<br />
The phonemes in a language differ from one ano<strong>the</strong>r along a finite<br />
number <strong>of</strong> phonetic dimensions, such as degree <strong>of</strong> voicing, degree <strong>of</strong><br />
noisiness, degree <strong>of</strong> nasality, degree <strong>of</strong> openness, degree <strong>of</strong> backness,<br />
degree <strong>of</strong> rounding, etc. Each phonetic dimension, in turn, is subdivided<br />
into a small number (two to four) <strong>of</strong> phonologically functional categories,<br />
such as voiced/voiceless, (half)closed/(half)open, front/central/back, etc.<br />
Phonetic dimensions generally have multiple acoustical correlates. For<br />
instance, degree <strong>of</strong> voicing correlates with a multitude <strong>of</strong> acoustic cues<br />
such as voice onset time, duration <strong>of</strong> preceding vowel, steepness <strong>of</strong><br />
intensity decay <strong>and</strong> <strong>of</strong> formant bends in preceding vowel, duration <strong>of</strong><br />
intervocalic (near) silence, duration <strong>and</strong> intensity <strong>of</strong> noise burst, steepness<br />
<strong>of</strong> intensity attack <strong>and</strong> formant bends <strong>of</strong> following vowel. These acoustic<br />
properties typically co-vary in preferred patterns, but may be manipulated
38 Vincent J. van Heuven<br />
independently through speech syn<strong>the</strong>sis. When non-typical (‘conflicting’)<br />
combinations <strong>of</strong> parameter values are generated in <strong>the</strong> laboratory, some<br />
cues prove to be more influential than o<strong>the</strong>rs; so-called ‘cue trading<br />
relationships’ have been established for many phonemic contrasts. In<br />
Dutch, for instance, vowel quality (acoustically defined by F1 <strong>and</strong> F2, i.e.,<br />
<strong>the</strong> centre frequencies <strong>of</strong> <strong>the</strong> lowest two resonances in <strong>the</strong> vocal tract) <strong>and</strong><br />
vowel duration were found to be equally influential in cuing <strong>the</strong> tense/laxcontrast<br />
between /�� <strong>and</strong> /�/: a duller vowel quality (lower F1 <strong>and</strong> F2values),<br />
normally cuing /�/ could be compensated for by increasing <strong>the</strong><br />
duration <strong>of</strong> <strong>the</strong> vowel so that native listeners still perceive /a �/ (<strong>and</strong> vice<br />
versa, van Heuven, 1986).<br />
Categorization <strong>of</strong> sounds may proceed along several possible lines.<br />
First, many differences between sounds are simply too small to be heard at<br />
all: <strong>the</strong>se are subliminal. The scientific discipline <strong>of</strong> psycho-acoustics<br />
provides a huge literature on precisely what differences between sounds<br />
can <strong>and</strong> cannot be heard with <strong>the</strong> naked ear. Moreover, research has shown<br />
that <strong>the</strong> human hearing mechanism (<strong>and</strong> that <strong>of</strong> mammals in general) has<br />
developed specific sensitivities to certain differences between sounds <strong>and</strong> is<br />
relatively deaf to o<strong>the</strong>rs. These predilections have been shown to be present<br />
at birth (probably even in utero), <strong>and</strong> need not be acquired through<br />
learning. However, human categorization <strong>of</strong> sound is fur<strong>the</strong>r shaped by<br />
exposure to language. As age progresses from infancy to adulthood, sound<br />
differences that were still above threshold shortly after birth quickly lose<br />
<strong>the</strong>ir distinctivity. An important concept in this context is <strong>the</strong> notion <strong>of</strong><br />
categorical perception. This notion is best explained procedurally in terms<br />
<strong>of</strong> a laboratory experiment.<br />
Imagine a minimal word pair such as English back ~ pack. <strong>On</strong>e<br />
important difference between <strong>the</strong>se two tokens is that <strong>the</strong> onset <strong>of</strong> voicing<br />
in back is more or less coincident with <strong>the</strong> plosive release, whilst <strong>the</strong> voice<br />
onset in pack does not start until some 50 ms after <strong>the</strong> release. It is not too<br />
difficult in <strong>the</strong> laboratory to create a series <strong>of</strong> exemplars by interpolating<br />
<strong>the</strong> voice onset time <strong>of</strong> a prototypical back (0-ms delay) <strong>and</strong> that <strong>of</strong> a<br />
prototypical pack (70-ms delay) in steps <strong>of</strong>, say, 10 ms, so that we now<br />
have an 8-step continuum ranging over 0, 10, 20, 30, 40, 50, 60, <strong>and</strong> 70 ms.<br />
These eight exemplars are shuffled in r<strong>and</strong>om or<strong>der</strong> <strong>and</strong> played to an<br />
audience <strong>of</strong> native English listeners for identification as ei<strong>the</strong>r back or pack<br />
(forced choice). The 0-ms voice delay token will naturally come out with<br />
exclusively back-responses (0% pack); <strong>the</strong> 70-ms token will have 100%<br />
pack-responses. But what results will be obtained for <strong>the</strong> intermediate
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 39<br />
exemplars? If <strong>the</strong> 10-ms changes in voice delay are perceived continuously,<br />
one would predict a constant, gradual increase in %-pack responses for<br />
each 10-ms increment in <strong>the</strong> delay. I.e., when <strong>the</strong> stimulus increment (from<br />
left to right) is plotted against <strong>the</strong> response increment (from bottom to top),<br />
<strong>the</strong> psychometric function (<strong>the</strong> line that captures <strong>the</strong> stimulus-response<br />
relationship) is essentially a straight line (open symbols in Figure 1B). The<br />
typical outcome <strong>of</strong> experiments with voiced/voiceless continua, however, is<br />
non-continuous. For <strong>the</strong> first part <strong>of</strong> <strong>the</strong> continuum all exemplars are<br />
perceived as back-tokens, <strong>the</strong> rightmost two or three exemplars are nearunanimously<br />
perceived as pack. <strong>On</strong>ly for one or two exemplars in <strong>the</strong><br />
middle <strong>of</strong> <strong>the</strong> continuum do we observe uncertainty on <strong>the</strong> part <strong>of</strong> <strong>the</strong><br />
listener: here <strong>the</strong> distribution <strong>of</strong> responses is more or less ambiguous<br />
between back <strong>and</strong> pack. The psychometric function for this so-called<br />
categorical perception is sigmoid, i.e., has <strong>the</strong> shape <strong>of</strong> an S (big solid<br />
symbols in Figure 1B). In <strong>the</strong> idealized case <strong>of</strong> perfect categorical<br />
perception we would, in fact, expect to see a step-function jumping<br />
abruptly from (almost) 0 to (almost) 100% pack-responses somewhere<br />
along <strong>the</strong> continuum (thin black line with small solid symbols in Figure<br />
1B).<br />
The category boundary (at 35-ms VOT in Figure 1B) is defined as <strong>the</strong><br />
(interpolated) point along <strong>the</strong> stimulus axis where <strong>the</strong> distribution <strong>of</strong><br />
responses is completely ambiguous, i.e., 50-50%. For a well-defined crossover<br />
from one category to <strong>the</strong> o<strong>the</strong>r <strong>the</strong>re should be a point along <strong>the</strong><br />
stimulus axis where 75% <strong>of</strong> <strong>the</strong> responses agree on one category, <strong>and</strong> a<br />
second point where <strong>the</strong>re is 75%-agreement on <strong>the</strong> o<strong>the</strong>r category. The<br />
uncertainty margin is defined in absolute terms as <strong>the</strong> distance along <strong>the</strong><br />
stimulus axis between <strong>the</strong> two 75%-points; equivalent relative measures<br />
can be <strong>der</strong>ived from <strong>the</strong> steepness <strong>of</strong> <strong>the</strong> psychometric function (e.g. <strong>the</strong><br />
slope coefficient or <strong>the</strong> st<strong>and</strong>ard deviation <strong>of</strong> <strong>the</strong> cumulative normal<br />
distribution fitted to <strong>the</strong> data points).
40 Vincent J. van Heuven<br />
Figure 1. Panel A. Hypo<strong>the</strong>tical discrimination function for physically same <strong>and</strong><br />
different pairs <strong>of</strong> stimuli (one-step difference) reflecting categorical<br />
perception. Panel B. Illustration <strong>of</strong> continuous (open squares) versus<br />
categorical (big solid squares) perception in <strong>the</strong> identification <strong>and</strong><br />
discrimination paradigm. The thin line with small squares represents <strong>the</strong><br />
ideal step function that should be obtained when categorical perception is<br />
absolute. Category boundary <strong>and</strong> uncertainty margin are indicated<br />
(fur<strong>the</strong>r, see text).<br />
Although a pronounced sigmoid function (such as <strong>the</strong> one drawn in Figure<br />
1B) is a clear sign <strong>of</strong> categorical perception, researchers have always been<br />
reluctant to consi<strong>der</strong> it definitive pro<strong>of</strong>. Listeners, when forced to, tend to<br />
split any continuum down <strong>the</strong> middle. For a continuum to be perceived<br />
categorically, <strong>the</strong>refore, two conditions should be met:
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 41<br />
- results <strong>of</strong> an identification experiment should show a clear sigmoid<br />
function, <strong>and</strong><br />
- <strong>the</strong> discrimination function should show a local peak for stimuli<br />
straddling <strong>the</strong> category boundary.<br />
The discrimination function is determined in a separate experiment in<br />
which ei<strong>the</strong>r (i) identical or (ii) adjacent tokens along <strong>the</strong> stimulus<br />
continuum are presented pair-wise. Listeners <strong>the</strong>n decide for each pair<br />
whe<strong>the</strong>r <strong>the</strong> two tokens are ‘same’ or ‘different’. Two kinds <strong>of</strong> error may<br />
occur in a discrimination task:<br />
- a physically different pair may be heard as ‘same’, <strong>and</strong><br />
- a pair <strong>of</strong> identical tokens may be called ‘different’.<br />
The results <strong>of</strong> a discrimination task are best expressed as <strong>the</strong> percentage <strong>of</strong><br />
correct decisions obtained for a ‘different’ stimulus pair minus <strong>the</strong><br />
percentage <strong>of</strong> errors for ‘same’ pairs constructed from <strong>the</strong>se stimuli (<strong>the</strong><br />
latter percentage is <strong>of</strong>ten called <strong>the</strong> response bias). In <strong>the</strong> case <strong>of</strong> true<br />
categorical perception <strong>the</strong> discrimination scores show a pronounced peak<br />
for <strong>the</strong> stimulus pair straddling <strong>the</strong> category boundary, whilst all o<strong>the</strong>r pairs<br />
are discriminated at or only little above chance level (see panel A in Figure<br />
1). Physically different sounds that fall in <strong>the</strong> same perceptual category are<br />
hard to discriminate. In <strong>the</strong> case <strong>of</strong> continuous perception, <strong>the</strong>re is no local<br />
peak in <strong>the</strong> discrimination function.<br />
1.2. Categorical nature <strong>of</strong> intonational contrasts<br />
By intonation or speech melody we mean <strong>the</strong> pattern <strong>of</strong> rises <strong>and</strong> falls in<br />
<strong>the</strong> time-course <strong>of</strong> <strong>the</strong> pitch <strong>of</strong> spoken sentences. Melodic patterns in<br />
speech vary systematically across languages, <strong>and</strong> even within languages<br />
across dialects. The cross-linguistic differences can be parameterized <strong>and</strong><br />
described in much <strong>the</strong> same way as has been done for <strong>the</strong> segmentals in<br />
language: a set <strong>of</strong> distinctive features defines an inventory <strong>of</strong> abstract units,<br />
which can be organized in higher-or<strong>der</strong> units subject to wellformedness<br />
constraints. Moreover, intonational contrasts are used to perform<br />
grammatical functions that can also be expressed by lexico-syntactic<br />
means, such as turning statements into questions, <strong>and</strong> putting constituents<br />
in focus. For <strong>the</strong>se reasons it has become widely accepted that intonation is
42 Vincent J. van Heuven<br />
part <strong>of</strong> <strong>the</strong> linguistic system (Ladd, 1996: 8). Yet, <strong>the</strong>re have always been<br />
adherents <strong>of</strong> <strong>the</strong> view that speech melody should be consi<strong>der</strong>ed as<br />
something outside <strong>the</strong> realm <strong>of</strong> linguistics proper, i.e., that intonation is a<br />
paralinguistic phenomenon at best, to be treated on a par with <strong>the</strong><br />
expression <strong>of</strong> attitudes or emotions. Typically, <strong>the</strong> communication <strong>of</strong><br />
emotions (such as anger, fear, joy, surprise) or <strong>of</strong> attitudes (such as<br />
sarcasm) is non-categorical: <strong>the</strong> speaker shows himself more or less angry,<br />
fearful, or sarcastic in a continuous, gradient fashion.<br />
A relatively recent insight, <strong>the</strong>refore, is that a division should be made<br />
in melodic phenomena occurring in speech between linguistic versus<br />
paralinguistic contrasts. Obviously, only <strong>the</strong> former but not <strong>the</strong> latter type<br />
<strong>of</strong> phenomena should be described by <strong>the</strong> grammar <strong>and</strong> explained by<br />
linguistic <strong>the</strong>ory. This, however, begs <strong>the</strong> question how <strong>the</strong> difference can<br />
be made between linguistic <strong>and</strong> paralinguistic phenomena within <strong>the</strong> realm<br />
<strong>of</strong> speech melody. 3 Ladd & Morton (1997) were <strong>the</strong> first to suggest that <strong>the</strong><br />
traditional diagnostic for categorical perception should be applicable to<br />
intonational categories in much <strong>the</strong> same ways as it works for segmental<br />
contrasts. <strong>On</strong>ly if a peak in <strong>the</strong> discrimination function is found for adjacent<br />
members on a tone continuum straddling a boundary between tonal<br />
categories, are <strong>the</strong> categories part <strong>of</strong> <strong>the</strong> linguistic system, i.e., phonological<br />
categories. If no categorical perception <strong>of</strong> <strong>the</strong> tone categories can be<br />
established, <strong>the</strong> categories are ‘just’ <strong>the</strong> extremes <strong>of</strong> a paralinguistic or<br />
phonetic tonal continuum. Ladd & Morton tested <strong>the</strong> traditional diagnostic<br />
on a tone continuum between normal <strong>and</strong> emphatic accent in English <strong>and</strong><br />
noted that it failed. This – to me – indicates that <strong>the</strong> contrast is not part <strong>of</strong><br />
<strong>the</strong> phonology <strong>of</strong> English.<br />
Remijsen & van Heuven (1999, 2003) tested <strong>the</strong> traditional diagnostic<br />
on a tone continuum between ‘L%’ <strong>and</strong> ‘H%’ in Dutch, <strong>and</strong> showed that<br />
indeed <strong>the</strong>re was a discrimination peak for adjacent members along <strong>the</strong><br />
continuum straddling <strong>the</strong> boundary – indicating that <strong>the</strong> ‘L%’ <strong>and</strong> ‘H%’<br />
categories are part <strong>of</strong> <strong>the</strong> phonology <strong>of</strong> Dutch. At <strong>the</strong> same time, however,<br />
we had to take recourse to listener-individual normalization <strong>of</strong> <strong>the</strong> category<br />
boundary, a complication that is not generally needed when dealing with<br />
contrasts in <strong>the</strong> segmental phonology. 4<br />
Van Heuven & Kirsner (2002) suggested that <strong>the</strong> relatively weak<br />
categorical effects in Remijsen & van Heuven could have been <strong>the</strong> result <strong>of</strong><br />
an incorrect subdivision <strong>of</strong> <strong>the</strong> ‘L%’ to ‘H%’ tone range. Van Heuven &<br />
Kirsner (2002) showed that Dutch listeners were perfectly able to<br />
categorize a range <strong>of</strong> final pitches between low <strong>and</strong> high in terms <strong>of</strong> three
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 43<br />
categories, functionally denoted as comm<strong>and</strong> intonation, continuation, <strong>and</strong><br />
question. However, we did not run <strong>the</strong> full diagnostic involving both<br />
identification <strong>and</strong> discrimination procedures. Moreover, Van Heuven &<br />
Kirsner forced <strong>the</strong>ir listeners to choose between three response alternatives,<br />
viz. comm<strong>and</strong>, conditional <strong>and</strong> question. Although <strong>the</strong> extremes <strong>of</strong> <strong>the</strong><br />
range, i.e. comm<strong>and</strong> versus question are unchallenged categories, it may<br />
well be <strong>the</strong> case that <strong>the</strong> conditional is not necessarily distinct from <strong>the</strong><br />
question type. After all, in <strong>the</strong> grammar developed by ‘t Hart, Collier &<br />
Cohen (1990) any type <strong>of</strong> non-low terminal pitch falls into <strong>the</strong> same<br />
category, indicating non-finality. It occurred to us that we should take <strong>the</strong><br />
precaution to run <strong>the</strong> experiment several times, using different response<br />
alternatives, such that two separate binary (‘comm<strong>and</strong>’ ~ ‘no comm<strong>and</strong>’<br />
<strong>and</strong> ‘question ~ ‘no question’) response sets as well as <strong>the</strong> ternary response<br />
set (‘comm<strong>and</strong>’ ~ ‘conditional’ ~ ‘question’) were used by <strong>the</strong> same set <strong>of</strong><br />
listeners. If <strong>the</strong> intermediate ‘conditional’ response category does constitute<br />
a clearly defined notion in <strong>the</strong> listeners’ minds, <strong>the</strong> binary <strong>and</strong> ternary<br />
divisions <strong>of</strong> <strong>the</strong> stimulus range should converge on <strong>the</strong> category boundaries.<br />
The present paper seeks to remedy <strong>the</strong> infelicities <strong>of</strong> Van Heuven &<br />
Kirsner (2002). However, before I deal with <strong>the</strong> experiments, it is necessary<br />
to introduce <strong>the</strong> inventory <strong>of</strong> <strong>the</strong> domain-final boundary configurations that<br />
can be found in Dutch.<br />
1.3. Dutch domain-final boundary tones<br />
Over <strong>the</strong> past decades a major research effort has been spent on <strong>the</strong> formal<br />
description <strong>of</strong> <strong>the</strong> sentence melody <strong>of</strong> Dutch. In <strong>the</strong> present paper we<br />
concentrate on one small part <strong>of</strong> <strong>the</strong> intonation system <strong>of</strong> Dutch: <strong>the</strong> options<br />
that are available to <strong>the</strong> speaker to terminate an intonation phrase. It has<br />
become customary to model <strong>the</strong> intonation system <strong>of</strong> a language as a<br />
hierarchically organized structure in which <strong>the</strong> tonal primitives (or ‘atoms’)<br />
are combined into tonal configurations, which in turn combine into<br />
intonation phrases. <strong>On</strong>e or more <strong>of</strong> such intonation phrases are combined<br />
into an utterance, which may combine with o<strong>the</strong>r utterances to form a<br />
prosodic paragraph. The intonation phrase (henceforth IP), <strong>the</strong>n, is situated<br />
roughly in <strong>the</strong> middle <strong>of</strong> <strong>the</strong> prosodic hierarchy. Note that a short utterance<br />
may consist <strong>of</strong> just one IP. An IP is characterized as a stretch <strong>of</strong> speech<br />
between two IP boundaries, i.e., a break in <strong>the</strong> segment string that is<br />
signaled by ei<strong>the</strong>r a pause (physical interruption <strong>of</strong> <strong>the</strong> sound stream), pre-
44 Vincent J. van Heuven<br />
boundary leng<strong>the</strong>ning <strong>and</strong>/or by a boundary-marking tone. If <strong>the</strong> boundary<br />
is sentence medial, <strong>the</strong>n yet ano<strong>the</strong>r IP must follow in or<strong>der</strong> to finish <strong>the</strong><br />
utterance.<br />
The first explicit <strong>and</strong> experimentally verified grammar <strong>of</strong> Dutch<br />
intonation was developed at <strong>the</strong> Institute for Perception Research at<br />
Eindhoven (‘t Hart et al., 1990; Rietveld & van Heuven, 2001: 263-270).<br />
This grammar models <strong>the</strong> sentence melody <strong>of</strong> Dutch as a system <strong>of</strong> two<br />
gently declining reference lines, nominally 6 semitones (half an octave)<br />
apart, between which <strong>the</strong> pitch rises <strong>and</strong> falls in a limited number <strong>of</strong><br />
patterns. The grammar provides for three different ways in which an IP<br />
may be terminated: (i) on <strong>the</strong> low reference line (‘0’), (ii) on <strong>the</strong> high<br />
reference line (‘∅’), or (iii) by executing a steep pitch rise (‘2’). Although<br />
<strong>the</strong> grammar is not completely explicit on this point, it appears that <strong>the</strong><br />
<strong>of</strong>fset <strong>of</strong> rise ‘2’ may exceed <strong>the</strong> level <strong>of</strong> <strong>the</strong> high reference line,<br />
specifically when <strong>the</strong> rise starts at <strong>the</strong> high reference line. The grammar<br />
<strong>the</strong>n allows IPs to end at three different pitches: low, high, <strong>and</strong> extra high.<br />
A more recent account <strong>of</strong> Dutch intonation is given by Gussenhoven <strong>and</strong><br />
co-workers (Gussenhoven, Rietveld & Terken, 1999; Rietveld & van<br />
Heuven, 2001: 270-277). This model is constructed along <strong>the</strong> principles<br />
adopted by autosegmental intonologists, in which a sentence melody is<br />
basically a sequence <strong>of</strong> tonal targets <strong>of</strong> two types: ‘H’ (high) <strong>and</strong> ‘L’ (low).<br />
The ToDI system (Transcription <strong>of</strong> Dutch Intonation), which is an<br />
inventory <strong>of</strong> tonal configurations for surface-level transcriptions <strong>of</strong> Dutch<br />
sentence melodies using <strong>the</strong> autosegmental H/L notation format, provides<br />
three symbols for marking IP boundaries: (i) ‘L%’, i.e., <strong>the</strong> final pitch<br />
target extends below <strong>the</strong> baseline, (ii) ‘%’, i.e., <strong>the</strong> absence <strong>of</strong> a tonal IP<br />
boundary marker, <strong>and</strong> (iii) ‘H%’, i.e., <strong>the</strong> final pitch is higher than <strong>the</strong><br />
preceding pitch. 5 For details <strong>of</strong> <strong>the</strong> ToDI transcription system I refer to <strong>the</strong><br />
ToDI website (www.l<strong>and</strong>s.kun.nl/todi) or to Rietveld & van Heuven (2001:<br />
399-401).<br />
Remijsen & van Heuven (1999, 2003) report an experiment which<br />
sought to establish <strong>the</strong> perceptual boundary between sentence-final<br />
statement <strong>and</strong> question intonation. They did this by varying <strong>the</strong> pitch<br />
configuration on <strong>the</strong> utterance-final syllable <strong>of</strong> <strong>the</strong> verb-less phrase De<br />
Dennenlaan(?) ‘Pine Lane(?)’ between a fall <strong>and</strong> a steep rise in eleven<br />
perceptually equal steps. Listeners were <strong>the</strong>n asked to decide for each <strong>of</strong> <strong>the</strong><br />
eleven pitch patterns whe<strong>the</strong>r <strong>the</strong>y perceived it as a statement or a question.<br />
At <strong>the</strong> time we tacitly assumed that <strong>the</strong> continuum spanned just two<br />
pragmatic categories, i.e. statement versus question, <strong>and</strong> that <strong>the</strong>re was no
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 45<br />
relevant intermediate category that could be interpreted as ‘non-finality’. In<br />
fact, Kirsner & van Heuven (1996) suggested a single abstract meaning for<br />
<strong>the</strong> non-low tonal category: ‘appeal (by <strong>the</strong> speaker to <strong>the</strong> hearer)’, asking<br />
for <strong>the</strong> hearer’s continued attention or for a verbal response to a question or<br />
a non-verbal compliance with a request. However, Caspers (1998)<br />
suggested that <strong>the</strong>re is a functional difference between <strong>the</strong> non-tonal<br />
boundary (‘%’) following an earlier ‘H*’ target <strong>and</strong> <strong>the</strong> high boundary<br />
(‘H%’) following an earlier ‘H*’. She syn<strong>the</strong>sized stimuli in which <strong>the</strong><br />
terminal pitch after <strong>the</strong> accent-marking ‘H*’ was followed by ei<strong>the</strong>r ‘H%’<br />
(where <strong>the</strong> final pitch was raised fur<strong>the</strong>r) or just % (where <strong>the</strong> pitch<br />
remained high but level after <strong>the</strong> accent). Her results indicate that listeners<br />
unequivocally expect <strong>the</strong> speaker to continue after <strong>the</strong> ‘H* ... %’<br />
configuration, in contradistinction to <strong>the</strong> ‘H* ... H%’ pattern, for which <strong>the</strong><br />
responses were equally divided between ‘same speaker will continue’ <strong>and</strong><br />
‘interlocutor will take over (with a response)’.<br />
Note that <strong>the</strong> ‘%’ tone-less boundary as studied by Caspers is found<br />
only after a preceding H* accent. Strictly speaking, <strong>the</strong>n, <strong>the</strong> ‘%’ boundary<br />
cannot be used as an intermediate category in between ‘L%’ <strong>and</strong> ‘H%’<br />
when <strong>the</strong> preceding pitch is low. After ‘L’, any rise in pitch, whe<strong>the</strong>r strong<br />
or intermediate, is a perceptually relevant change in pitch, which must be<br />
coded by an ‘H%’ target. <strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, this formal constraint is in <strong>the</strong><br />
way <strong>of</strong> an attractive generalization which would allow us to view <strong>the</strong> high<br />
level pitch (‘H* ... %’) pattern as a surface realization <strong>of</strong> <strong>the</strong> ‘H*L...%’<br />
pattern from which <strong>the</strong> L target has been deleted – in much <strong>the</strong> same way as<br />
was suggested by Haan (2002) in or<strong>der</strong> to account for <strong>the</strong> functional<br />
similarity between <strong>the</strong> ‘H*...H%’ <strong>and</strong> <strong>the</strong> ‘H*L…H%’ interrogative<br />
patterns, as exemplified in Figure 2.<br />
Figure 2. Un<strong>der</strong>lying tonal shape (dotted) <strong>and</strong> surface realization after ‘L’-deletion<br />
(solid) <strong>of</strong> an ‘H*L … H%’ sequence.<br />
There seems to be a mismatch between <strong>the</strong> functions expressed by Caspers’<br />
‘%’ <strong>and</strong> ‘H%’ after ‘H*’. If we assume an iconic relationship between <strong>the</strong>
46 Vincent J. van Heuven<br />
terminal pitch <strong>of</strong> <strong>the</strong> utterance <strong>and</strong> <strong>the</strong> degree <strong>of</strong> submissiveness <strong>of</strong> <strong>the</strong><br />
speaker towards <strong>the</strong> hearer, <strong>the</strong>n we would reason that ‘H%’ should make<br />
more <strong>of</strong> an appeal to <strong>the</strong> hearer (expressing greater submissiveness) than<br />
just ‘%’. <strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, answering a question seems a bigger favor on<br />
<strong>the</strong> part <strong>of</strong> <strong>the</strong> hearer than merely waiting for <strong>the</strong> speaker to continue <strong>the</strong><br />
utterance. It could be <strong>the</strong> case, <strong>of</strong> course, that even <strong>the</strong> highest terminal<br />
pitches used by Caspers were not high enough to elicit unambiguous ‘o<strong>the</strong>r<br />
speaker will take over’ (i.e. ‘question’) responses. Also, it is unclear if <strong>the</strong><br />
unambiguous ‘same speaker will continue’ response crucially depends on a<br />
flat stretch <strong>of</strong> high declination (as is <strong>the</strong> case after an ‘H*’ accent) or if any<br />
terminal pitch <strong>of</strong> intermediate height would yield <strong>the</strong> same response.<br />
In Caspers’ analysis <strong>the</strong> ‘%’ boundary – <strong>and</strong> arguably an ‘L … H%’<br />
sequence with a mo<strong>der</strong>ately high terminal pitch – unambiguously signals<br />
continuation. This category would <strong>the</strong>n be expected to be firmly<br />
represented in <strong>the</strong> listener’s cognitive system. Varying <strong>the</strong> terminal pitch<br />
from low to extremely high should <strong>the</strong>n elicit two well-defined categories:<br />
(i) unambiguous statement for low pitches, (ii) unambiguous continuations<br />
for intermediate terminal pitches, <strong>and</strong> (iii) a poorly defined or non-unique<br />
interrogative category, which is also compatible with a continuation<br />
reading.<br />
At this time, <strong>the</strong>n, we do not know whe<strong>the</strong>r two or three formal tone<br />
categories should be postulated in IP-final position. It seems that <strong>the</strong> status<br />
<strong>of</strong> ‘L%’ as a linguistic category is unchallenged but <strong>the</strong> non-low part <strong>of</strong> <strong>the</strong><br />
IP-final tone range is very much a matter <strong>of</strong> debate. Does <strong>the</strong> non-low part<br />
<strong>of</strong> <strong>the</strong> range form a continuum expressing lesser or greater appeal by <strong>the</strong><br />
speaker in a paralinguistic manner, or should this part <strong>of</strong> <strong>the</strong> range be split<br />
into two discrete phonological categories, each expressing a distinct<br />
meaning <strong>of</strong> its own (i.e. ‘continuation’ ~ ‘question’, or – even worse – into<br />
two categories <strong>of</strong> which one is specific for ‘continuation’ <strong>and</strong> <strong>the</strong> o<strong>the</strong>r<br />
un<strong>der</strong>specified <strong>and</strong> compatible with both ‘question’ <strong>and</strong> ‘continuation’?<br />
These meanings, <strong>and</strong> a possible way <strong>of</strong> testing <strong>the</strong> categorical nature <strong>of</strong><br />
tonal contrasts expressing <strong>the</strong>m, are <strong>the</strong> topic <strong>of</strong> <strong>the</strong> next section.<br />
1.4. Clause typing<br />
Dutch, like any o<strong>the</strong>r language, has lexico-syntactic means to express a<br />
range <strong>of</strong> clause types, such as statement, comm<strong>and</strong>, exclamation <strong>and</strong><br />
question. Although <strong>the</strong> lexico-syntactic means are generally adequate <strong>and</strong>
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 47<br />
sufficient to express <strong>the</strong> speaker’s pragmatic intention to <strong>the</strong> hearer, several<br />
– if not all – clause types are supported by prosodic means, specifically by<br />
appropriate intonation patterns. In fact, exceptional situations may arise<br />
where <strong>the</strong>re is no lexico-syntactic differentiation between <strong>the</strong> clause types,<br />
<strong>and</strong> where <strong>the</strong> speaker’s intention can only be recovered from melodic<br />
cues. For <strong>the</strong> purposes <strong>of</strong> <strong>the</strong> present experiment we have looked for a<br />
situation in which <strong>the</strong> three prosodic categories may serve as <strong>the</strong> only cue to<br />
a ternary choice among clause types, so that prosody will be exploited to<br />
<strong>the</strong> utmost, <strong>and</strong> <strong>the</strong> listener’s choice will not be co-determined by lexical<br />
<strong>and</strong>/or syntactic cues. Such a situation may be obtained in a V1 sentence,<br />
where <strong>the</strong> finite verb has been moved into <strong>the</strong> sentence-initial position. 6 In<br />
<strong>the</strong> sentence Neemt u de trein naar Wageningen ‘Take you <strong>the</strong> train to<br />
Wageningen’ <strong>the</strong> lexico-syntactic information is compatible with at least<br />
three interpretations: 7<br />
- A polite imperative (Kirsner, van Heuven & Caspers, 1998)<br />
- A conditional clause similar in meaning to ‘If you take <strong>the</strong> train to<br />
Wageningen ...’<br />
- A yes/no question ‘Do you take <strong>the</strong> train to Wageningen?’<br />
Which <strong>of</strong> <strong>the</strong> three readings is intended by <strong>the</strong> speaker, is expressed<br />
through prosody only. In setting up <strong>the</strong> experiment we assumed that <strong>the</strong>re is<br />
no principal difference in <strong>the</strong> speech melody between a statement <strong>and</strong> a<br />
comm<strong>and</strong> in Dutch. 8 Using a range <strong>of</strong> terminal pitch patterns on <strong>the</strong> single<br />
phrase Neemt u de trein naar Wageningen, we can determine <strong>the</strong> category<br />
boundaries between comm<strong>and</strong> (for statement), conditional (for<br />
continuation), <strong>and</strong> question without any interfering differences in lexicosyntactic<br />
structure.<br />
We may conclude this introduction by summarizing <strong>the</strong> research<br />
questions that we will address:<br />
1. Are <strong>the</strong> domain-final boundaries ‘L%’ ~ ‘%’ ~ ‘H%’ contiguous<br />
categories along a single tonal dimension?<br />
2. Is <strong>the</strong>re a one-to-one correspondence between ‘L%’ <strong>and</strong> ‘comm<strong>and</strong>’,<br />
‘%’ <strong>and</strong> ‘conditional’, <strong>and</strong> ‘H%’ <strong>and</strong> ‘question’?<br />
3. Where are <strong>the</strong> category boundaries – if any – along <strong>the</strong> continuum<br />
between (i) ‘L%’ <strong>and</strong> ‘%’ <strong>and</strong> (ii) between ‘%’ <strong>and</strong> ‘H%’?<br />
4. Are <strong>the</strong> category boundaries at <strong>the</strong> same positions along <strong>the</strong> stimulus<br />
range irrespective <strong>of</strong> <strong>the</strong> binary versus ternary response mode?
48 Vincent J. van Heuven<br />
5. Are both boundaries truly categorical in <strong>the</strong> sense that <strong>the</strong>re are<br />
discrimination peaks for adjacent stimulus pairs straddling <strong>the</strong> category<br />
boundaries?<br />
2. Methods<br />
2.1. Stimuli<br />
A male native speaker <strong>of</strong> st<strong>and</strong>ard Dutch read <strong>the</strong> sentence Neemt u de trein<br />
naar WAgeningen? with a single ‘H*L’ accent on <strong>the</strong> first syllable <strong>of</strong><br />
Wageningen. The utterance was recorded onto digital audio tape (DAT)<br />
using a Sennheiser MKH 416 unidirectional condenser microphone,<br />
transferred to computer disk (16 kHz, 16 bits) <strong>and</strong> digitally processed using<br />
<strong>the</strong> Praat speech processing s<strong>of</strong>tware (Boersma & Weenink, 1996; Boersma<br />
& van Heuven, 2001). The intonation pattern <strong>of</strong> <strong>the</strong> utterance was stylized<br />
by h<strong>and</strong> as a sequence <strong>of</strong> straight lines in <strong>the</strong> ERB x linear time<br />
representation. Nine intonationally different versions were <strong>the</strong>n generated<br />
using <strong>the</strong> PSOLA analysis-resyn<strong>the</strong>sis technique (e.g. Moulines & Verhelst,<br />
1995; Rietveld & van Heuven, 2001: 379-380) implemented in <strong>the</strong> Praat<br />
s<strong>of</strong>tware. The nine versions were identical up to <strong>and</strong> including <strong>the</strong> ‘H*L’<br />
configuration on Wageningen. From that point onwards <strong>the</strong> nine versions<br />
diverged into two falls <strong>and</strong> seven rises. The terminal frequencies <strong>of</strong> <strong>the</strong> nine<br />
versions were chosen to be perceptually equidistant, i.e., <strong>the</strong> difference<br />
between any two adjacent terminal frequencies was equal in terms <strong>of</strong> <strong>the</strong><br />
ERB scale. 9 The terminal pitch <strong>of</strong> version 1 equaled 80 Hz, <strong>the</strong> increment<br />
in <strong>the</strong> terminal frequency for each following version was 0,25 ERB. The<br />
nine pitch patterns are shown in Figure 3.
80<br />
60<br />
160<br />
120<br />
80<br />
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 49<br />
neemt u de trein naar WA ge ni ngen<br />
0 0.5<br />
1<br />
1.5<br />
Time (s)<br />
Figure 3. Steps 1 through 9 along resyn<strong>the</strong>sized continuum differing in terminal F0<br />
by 0,25 ERB increments. Intensity contour (dB) <strong>and</strong> segmentation (by<br />
syllables) are indicated.<br />
2.2. Tasks <strong>and</strong> experimental procedures<br />
For <strong>the</strong> discrimination task, which was <strong>the</strong> first task imposed on <strong>the</strong><br />
subjects, we followed Ladd <strong>and</strong> Morton (1997) in using <strong>the</strong> AX<br />
discrimination paradigm. Stimuli were presented in pairs that were ei<strong>the</strong>r<br />
<strong>the</strong> same or one step apart on <strong>the</strong> continuum. In <strong>the</strong> latter case, <strong>the</strong> second<br />
can be higher or lower than <strong>the</strong> first (hereafter AB <strong>and</strong> BA, respectively).<br />
The eight AB stimulus types ran from pair {1,2} to {8,9}; <strong>the</strong> eight<br />
corresponding BA types from {2,1} to {9,8}. This yielded 9 identical pairs<br />
<strong>and</strong> 2 x 8 = 16 different pairs, which occurred in r<strong>and</strong>om or<strong>der</strong>, yielding a<br />
set <strong>of</strong> 25 trials in all, which was presented to each listener four times in<br />
different r<strong>and</strong>om or<strong>der</strong>s, preceded by five practice trials. Stimuli within<br />
pairs were separated by a 500-ms silence, <strong>the</strong> pause between pairs was<br />
3000 ms. A short warning tone was sounded after every tenth trial.<br />
For <strong>the</strong> identification task listeners responded to individual stimuli from<br />
<strong>the</strong> 9-step continuum by classifying each ei<strong>the</strong>r in terms <strong>of</strong> a binary or a<br />
ternary choice:<br />
1. ‘Comm<strong>and</strong>’ ~ ‘no comm<strong>and</strong>’. In one task <strong>the</strong> listeners were instructed<br />
to decide for each stimulus whe<strong>the</strong>r <strong>the</strong>y interpreted it as a comm<strong>and</strong> or<br />
not.<br />
2. ‘Question’ ~ ‘no question’. An alternative task involved <strong>the</strong> decision<br />
whe<strong>the</strong>r <strong>the</strong> stimulus sounded like a question or not.
50 Vincent J. van Heuven<br />
3. ‘Comm<strong>and</strong>’ ~ ‘condition’ ~ ‘question’. The third task was identical to<br />
<strong>the</strong> task imposed in van Heuven & Kirsner (2002).<br />
Half <strong>of</strong> <strong>the</strong> listeners first performed task (1), <strong>the</strong> o<strong>the</strong>r half <strong>of</strong> <strong>the</strong> listeners<br />
began with task (2). Task (3) was always <strong>the</strong> last identification procedure in<br />
<strong>the</strong> array <strong>of</strong> tests. For each task, <strong>the</strong> set <strong>of</strong> nine stimuli were presented five<br />
times to each listener, in different r<strong>and</strong>om or<strong>der</strong>s, <strong>and</strong> preceded by five<br />
practice items, yielding sets <strong>of</strong> 50 identification stimuli per task.<br />
Twenty native Dutch listeners, ten males <strong>and</strong> ten females, took part in<br />
<strong>the</strong> experiment on a voluntary basis. Participants were university students<br />
or members <strong>of</strong> <strong>the</strong>ir families. None <strong>of</strong> <strong>the</strong>m reported any perceptual<br />
deficiencies.<br />
The experiments were run with small groups <strong>of</strong> subjects, who listened to<br />
<strong>the</strong> stimuli at a comfortable loudness level over Quad ESL-63 electrostatic<br />
loudspeakers, while seated in a sound-treated lecture room. Subjects<br />
marked <strong>the</strong>ir responses on printed answer sheets provided to <strong>the</strong>m, always<br />
taking <strong>the</strong> discrimination task first <strong>and</strong> <strong>the</strong> identification tasks last.<br />
3. Results<br />
3.1. Identification<br />
Figures 4 <strong>and</strong> 5 present <strong>the</strong> results obtained in <strong>the</strong> binary identification<br />
tasks, i.e., <strong>the</strong> forced choice between ‘comm<strong>and</strong>’ ~ ‘no comm<strong>and</strong>’ (Figure<br />
4) <strong>and</strong> between ‘question’ ~ ‘no question’ (Figure 5).
Indentifications "comm<strong>and</strong>" (%)<br />
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 51<br />
Figure 4. Percent ‘comm<strong>and</strong>’ responses as a function <strong>of</strong> stimulus step (terminal F0<br />
increments in 0.25 ERB steps) in a binary identification task (‘comm<strong>and</strong>’<br />
~ ’no comm<strong>and</strong>’).<br />
The psychometric function for <strong>the</strong> ‘comm<strong>and</strong>’ responses is very steep. The<br />
category boundary between ‘comm<strong>and</strong>’ <strong>and</strong> ‘no comm<strong>and</strong>’ is located at a<br />
step size <strong>of</strong> 2.7, <strong>and</strong> <strong>the</strong> margin <strong>of</strong> uncertainty runs between 2.2 <strong>and</strong> 3.7,<br />
i.e., a cross-over from 75% to 25% ‘comm<strong>and</strong>’ responses is effected by an<br />
increase in <strong>the</strong> terminal pitch <strong>of</strong> <strong>the</strong> stimulus <strong>of</strong> 1.5 step (i.e., 0.37 ERB).<br />
Indentifications "question" (%)<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
1<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
1<br />
2<br />
2<br />
3<br />
3<br />
Stimulus step<br />
4<br />
4<br />
5<br />
5<br />
6<br />
6<br />
Stimulus step<br />
7<br />
7<br />
8<br />
8<br />
Figure 5. Percent ‘question’ responses as a function <strong>of</strong> stimulus step (terminal F0<br />
increments in 0.25 ERB steps) in a binary identification task (‘question’<br />
~ ’no question’<br />
9<br />
9
52 Vincent J. van Heuven<br />
A complete cross-over is also found for <strong>the</strong> ‘question’ ~ ‘no question’ task.<br />
The category boundary finds itself at a stimulus value <strong>of</strong> 3.6, whilst <strong>the</strong><br />
margin <strong>of</strong> uncertainty runs between 2.3 <strong>and</strong> 4.9, i.e., an interval <strong>of</strong> 2.6<br />
increments <strong>of</strong> 0.25 ERB. We may note that <strong>the</strong> category boundaries in <strong>the</strong><br />
‘comm<strong>and</strong>’ <strong>and</strong> <strong>the</strong> ‘question’ tasks do not coincide, but are separated<br />
along <strong>the</strong> stimulus axis by almost a complete step: 2.7 versus 3.6 or 0.9<br />
step. Note, once more, that none <strong>of</strong> <strong>the</strong> subjects had been alerted to <strong>the</strong><br />
possible existence <strong>of</strong> an intermediate category between ‘comm<strong>and</strong>’ <strong>and</strong><br />
‘question’. Therefore, <strong>the</strong> emergence <strong>of</strong> <strong>the</strong> interval between <strong>the</strong><br />
‘comm<strong>and</strong>’ <strong>and</strong> <strong>the</strong> ‘question’ boundaries might be taken in justification <strong>of</strong><br />
such an intermediate category.<br />
Let us now turn to <strong>the</strong> results <strong>of</strong> <strong>the</strong> ternary identification task in which<br />
all <strong>the</strong> listeners who had already responded to <strong>the</strong> stimuli were now<br />
required to classify <strong>the</strong> nine stimulus types as ei<strong>the</strong>r ‘comm<strong>and</strong>’,<br />
‘conditional subclause’ or ‘question’. These results are shown in Figure 6.<br />
Identification (%)<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
Stimulus step<br />
7<br />
8<br />
9<br />
'comm<strong>and</strong>'<br />
'continuation'<br />
'question'<br />
Figure 6. Ternary identification <strong>of</strong> stimuli as ‘comm<strong>and</strong>’, ‘conditional clause’ or<br />
‘question’. Category boundaries are indicated.<br />
The boundary between ‘comm<strong>and</strong>’ <strong>and</strong> <strong>the</strong> ‘continuation’ categories is at<br />
2.8; this is hardly different than <strong>the</strong> ‘comm<strong>and</strong>’ ~ ’no comm<strong>and</strong>’ boundary<br />
that was found in <strong>the</strong> binary response task. This, <strong>the</strong>n, would seem to be a<br />
very robust boundary, showing that at least ‘comm<strong>and</strong>’ intonation has welldefined<br />
linguistic status. The boundary between ‘continuation’ <strong>and</strong><br />
‘question’ is less clearly defined. Also, <strong>the</strong> maximum scores in <strong>the</strong>se two
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 53<br />
categories are around 80% ra<strong>the</strong>r than 90% or more. Although <strong>the</strong>re is no<br />
ambiguity in <strong>the</strong> listeners’ minds whe<strong>the</strong>r a stimulus is a comm<strong>and</strong> or<br />
something else, <strong>the</strong> choice between ‘continuation’ <strong>and</strong> ‘question’ seems<br />
more ambiguous leaving room for a minority response in <strong>the</strong> or<strong>der</strong> <strong>of</strong> 20%.<br />
This would indicate to us that we are dealing here with a continuum ra<strong>the</strong>r<br />
than with a dichotomy. Finally, we may note that <strong>the</strong> (s<strong>of</strong>t) category<br />
boundary between ‘continuation’ <strong>and</strong> ‘question’ is located at a stimulus<br />
value <strong>of</strong> 7.2. The boundary, <strong>the</strong>n, that sets <strong>of</strong>f ‘question’ from ‘no question’<br />
responses proves very unstable: <strong>the</strong>re is a shift from <strong>the</strong> binary response<br />
task (3.6) to <strong>the</strong> ternary task (7.2) <strong>of</strong> no less than 3.6 points along <strong>the</strong><br />
stimulus continuum.<br />
It would seem, <strong>the</strong>n, that <strong>the</strong> ‘comm<strong>and</strong>’ category is highly stable <strong>and</strong><br />
well-established in <strong>the</strong> minds <strong>of</strong> <strong>the</strong> listeners. The ‘question’ boundary,<br />
however, is ra<strong>the</strong>r poorly defined, as a result <strong>of</strong> several circumstances. The<br />
cross-over points for <strong>the</strong> ‘question’ category <strong>of</strong> individual listeners vary<br />
over a wide range <strong>of</strong> stimulus values, i.e., between 2.2 <strong>and</strong> 8.5 step number,<br />
with a fairly even spread <strong>of</strong> values in between <strong>the</strong>se extremes. Moreover,<br />
for two listeners no cross-over to <strong>the</strong> ‘question’ category could be found at<br />
all; here <strong>the</strong> listeners never gave <strong>the</strong> ‘question’ response in more than 75%.<br />
Also, some listeners have extremely sharp cross-overs to <strong>the</strong> ‘question’<br />
category, but o<strong>the</strong>rs show large margins <strong>of</strong> uncertainty.<br />
3.2. Discrimination<br />
Figure 7 presents <strong>the</strong> mean percentage <strong>of</strong> successfully discriminated stimuli<br />
that were actually different (hereafter ‘hits’), <strong>and</strong> <strong>the</strong> percentage <strong>of</strong> false<br />
alarms, i.e. ‘different’ responses to (identical) AA stimuli. The false-alarm<br />
rate is roughly 20% across <strong>the</strong> entire stimulus continuum. This value can be<br />
seen as a bias for responding ‘different’. Generally, an increment <strong>of</strong> 0.25<br />
ERB is discriminated above <strong>the</strong> 20% bias level, with <strong>the</strong> exception <strong>of</strong> <strong>the</strong><br />
difference between stimulus steps 5 <strong>and</strong> 6. The discrimination function<br />
shows two local peaks. The first one is very large, <strong>and</strong> is located between<br />
stimulus steps 2 <strong>and</strong> 3. This peak obviously coincides with <strong>the</strong> stable<br />
category boundary found between ‘comm<strong>and</strong>’ <strong>and</strong> <strong>the</strong> non-comm<strong>and</strong><br />
responses (whe<strong>the</strong>r binary or ternary). A much smaller second<br />
discrimination peak may be observed between stimulus steps 6 <strong>and</strong> 7,<br />
which location may well reflect <strong>the</strong> ra<strong>the</strong>r poorly defined category<br />
boundary between ‘continuation’ <strong>and</strong> ‘question’.
54 Vincent J. van Heuven<br />
"Different" judgments (%)<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
1 2 3 4 5 6 7 8 9<br />
1-2 2-3 3-4 4-5 5-6 6-7 7-8 8-9<br />
Stimulus step<br />
Figure 7. Percent ‘different’ judgments to nine identical stimulus pairs (false<br />
alarms) <strong>and</strong> eight pairs differing by one step (hits).<br />
4. Conclusions <strong>and</strong> discussion<br />
"same" pair<br />
"diff" pair<br />
Let us now try to formulate answers to <strong>the</strong> research questions that we asked<br />
in section 2. The first two questions, which I will attempt to answer<br />
toge<strong>the</strong>r, asked whe<strong>the</strong>r <strong>the</strong> domain-final boundary tones are contiguous<br />
categories along a single tonal dimension, <strong>and</strong> map onto <strong>the</strong> comm<strong>and</strong>,<br />
continuation <strong>and</strong> question meaning in a one-to-one fashion. The results <strong>of</strong><br />
our experiments clearly indicate that this is indeed <strong>the</strong> case. Our listeners<br />
had no difficulty in using <strong>the</strong> three response alternatives provided to <strong>the</strong>m.<br />
When <strong>the</strong> terminal pitch was lower than <strong>the</strong> preceding pivot point in <strong>the</strong><br />
contour <strong>the</strong> responses were almost unanimously for ‘comm<strong>and</strong>’. When <strong>the</strong><br />
IP-final pitch was higher than <strong>the</strong> preceding pivot point, <strong>the</strong> incidence <strong>of</strong><br />
‘continuation’ responses increased up to <strong>and</strong> including step 4, <strong>and</strong><br />
decreased for higher terminal pitches which were more readily identified as<br />
questions as <strong>the</strong> terminal pitch was higher. Although <strong>the</strong>re was always<br />
some ambiguity between <strong>the</strong> ‘continuation’ <strong>and</strong> ‘question’ alternatives, <strong>the</strong><br />
results clearly indicate that ‘continuation’ is signaled by mo<strong>der</strong>ate final<br />
pitch, <strong>and</strong> question by (extra) high pitch.<br />
The latter finding corresponds with our suggestion that asking a<br />
question involves a higher degree <strong>of</strong> appeal by <strong>the</strong> speaker to hearer than
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 55<br />
asking <strong>the</strong> listener’s continued attention. We may also note that our result<br />
clashes with Caspers (1998). She found that <strong>the</strong> intermediate final pitch (or<br />
high level pitch in her experiment) was unambiguously identified as<br />
continuation; extra high final pitch ambiguously coded ei<strong>the</strong>r continuation<br />
or question. Comparison <strong>of</strong> Caspers’ <strong>and</strong> our own results is hazardous since<br />
<strong>the</strong> utterance-final tone configurations differ, not so much at <strong>the</strong> un<strong>der</strong>lying<br />
tone level, but at <strong>the</strong> surface. It seems to me that <strong>the</strong> discrepancy between<br />
Caspers’ <strong>and</strong> our own findings can be resolved if we accept <strong>the</strong> possibility<br />
that Caspers’ extra high terminal pitch was simply not high enough to elicit<br />
<strong>the</strong> 80% ‘question’ responses that we got in our experiment.<br />
The results so far concur with van Heuven & Kirsner (2002). However,<br />
we may now go on to consi<strong>der</strong> <strong>the</strong> third, fourth <strong>and</strong> fifth question, which<br />
asked where <strong>the</strong> category boundaries are located along <strong>the</strong> final pitch<br />
continuum between ‘L%, ‘%’ <strong>and</strong> ‘H’, in <strong>the</strong> binary <strong>and</strong> ternary response<br />
tasks, <strong>and</strong> to what extent <strong>the</strong> boundaries coincide with a peak in <strong>the</strong><br />
discrimination function.<br />
The results obtained in <strong>the</strong> binary (‘comm<strong>and</strong>’ ~ ‘no comm<strong>and</strong>’) <strong>and</strong><br />
ternary (‘comm<strong>and</strong>’ ~ ‘continuation’ ~ ‘question’) identification tasks are<br />
virtually <strong>the</strong> same, yielding <strong>the</strong> same location <strong>of</strong> <strong>the</strong> boundary (at step 2.7)<br />
separating <strong>the</strong> ‘comm<strong>and</strong>’ category from <strong>the</strong> rest <strong>of</strong> <strong>the</strong> stimulus<br />
continuum. However, a very unstable boundary is found in <strong>the</strong> binary<br />
‘question’ ~ ‘no question’ task (at step 3.6), which is reflected in <strong>the</strong> poorly<br />
defined boundary separating <strong>the</strong> ‘continuation’ <strong>and</strong> ‘question’ categories in<br />
<strong>the</strong> ternary response task (at step 7.2). Moreover, we have seen that <strong>the</strong><br />
category boundary between ‘comm<strong>and</strong>’ <strong>and</strong> ‘no comm<strong>and</strong>’ coincides with<br />
a huge peak in <strong>the</strong> discrimination function. Although <strong>the</strong>re is a modest local<br />
maximum in <strong>the</strong> discrimination function that may be associated with a<br />
boundary between ‘continuation’ <strong>and</strong> ‘question’, this peak is not very<br />
convincing.<br />
I take <strong>the</strong>se findings as evidence that <strong>the</strong>re is a linguistic, or<br />
phonological, categorization <strong>of</strong> <strong>the</strong> IP-final boundary tone continuum in<br />
just two types, which is best characterized as low <strong>and</strong> non-low. The low<br />
boundary tone signals dominance or superiority on <strong>the</strong> part <strong>of</strong> <strong>the</strong> speaker.<br />
This is <strong>the</strong> boundary tone that is suited for issuing statements <strong>and</strong><br />
comm<strong>and</strong>s. The non-low boundary tone signals subservience <strong>of</strong> <strong>the</strong> speaker<br />
to <strong>the</strong> hearer; <strong>the</strong> speaker appeals to <strong>the</strong> hearer for his continued attention or<br />
for an answer to a question.<br />
The non-low part <strong>of</strong> <strong>the</strong> boundary opposition, however, represents a<br />
gradient, paralinguistic continuum between a mo<strong>der</strong>ate appeal (asking for
56 Vincent J. van Heuven<br />
<strong>the</strong> hearer’s continued attention) <strong>and</strong> a stronger appeal (asking <strong>the</strong> hearer<br />
for a verbal reply to a question). Here <strong>the</strong> lower terminal pitches are<br />
associated with weaker degrees <strong>of</strong> appeal (or subservience), <strong>and</strong> <strong>the</strong> higher<br />
levels with strong appeal, but in a continuous, gradient, non-phonological<br />
manner.<br />
Our results indicate that earlier findings reported by Remijsen & van<br />
Heuven (1999, 2003) are to be viewed with caution. We now know that <strong>the</strong><br />
proper task to be imposed on listeners should not be to decide whe<strong>the</strong>r <strong>the</strong><br />
stimulus is a statement (or a comm<strong>and</strong>) versus a question. If binary<br />
response alternatives are required, <strong>the</strong>n <strong>the</strong> categories should be ‘statement’<br />
versus ‘no statement’ but a better procedure would be to ask <strong>the</strong> listener to<br />
respond by choosing from three categories: ‘statement’ (equivalent to<br />
‘comm<strong>and</strong>’ in our experiments ~ ‘continuation’ ~ ‘question’. Had such<br />
precautions been taken by Remijsen & van Heuven, <strong>the</strong>ir category<br />
boundary would have been much better defined with less listener-individual<br />
variation.<br />
Methodologically, we argue that <strong>the</strong> classical identification-cumdiscrimination<br />
paradigm is a useful diagnostic tool in intonation research<br />
which allows linguists to decide experimentally whe<strong>the</strong>r a melodic contrast<br />
is categorical <strong>and</strong> <strong>the</strong>refore part <strong>of</strong> <strong>the</strong> phonology, or continuously gradient<br />
<strong>and</strong> <strong>the</strong>refore phonetic or even paralinguistic.<br />
Notes<br />
1 The experiments reported in this chapter were run by Susanne Strik <strong>and</strong> Josien<br />
Klink in partial fulfillment <strong>of</strong> <strong>the</strong> course requirements for <strong>the</strong> Experimental<br />
<strong>Phonetics</strong> Seminar taught by <strong>the</strong> Linguistics Programme at University <strong>of</strong><br />
Leiden.<br />
2 This commutation procedure is best viewed as a mental experiment; when <strong>the</strong><br />
exchange is implemented through actual digital tape splicing, <strong>the</strong> result is more<br />
<strong>of</strong>ten than not an uninterpretable stream <strong>of</strong> sound.<br />
3 The nature <strong>of</strong> <strong>the</strong> distinction between intonational categories is problematic for<br />
a fur<strong>the</strong>r reason: inter-listener agreement on <strong>the</strong> identity <strong>of</strong> intonational events<br />
is low (Pitrelli et al., 1994), particularly in comparison with <strong>the</strong> self-evident<br />
consensus on segmental distinctions. This lack <strong>of</strong> consistency has lead Taylor<br />
(1998) to reject a basic principle <strong>of</strong> (intonational) phonology, namely its<br />
categorical nature. With respect to methodology, researchers tend to act as<br />
expert listeners, linking contours that sound distinct to pragmatic meaning in an
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 57<br />
intuitive fashion. Accordingly, inter-researcher agreement may be low, too (e.g.<br />
Caspers, 1998).<br />
4 Never<strong>the</strong>less, large between-listener variability has been reported, for instance, in<br />
<strong>the</strong> cuing <strong>of</strong> <strong>the</strong> voiced/voiceless contrast by <strong>the</strong> duration <strong>of</strong> <strong>the</strong> pre-burst silent<br />
interval: <strong>the</strong> boundary was at 70 ms for subject #1 <strong>and</strong> over 100 ms for subject #7<br />
(Slis & Cohen, 1969). These results are commented on by Nooteboom & Cohen<br />
(1976: 84) as follows: ‘Although <strong>the</strong> cross-over from /d/ to /t/ proceeds ra<strong>the</strong>r<br />
gradually when averaged over all listeners, <strong>the</strong> boundary is quite sharply defined for<br />
individual listeners’ (my translation, VH).<br />
5 The ‘%’ sign following <strong>the</strong> tone letter (as in ‘L%’, ‘H%’) denotes a domainfinal<br />
boundary; domain-initial boundaries are coded by <strong>the</strong> ‘%’ sign preceding<br />
a tone letter (as in ‘%L’, ‘%H’). A ‘%’ sign unaccompanied by a tone letter<br />
may only occur in domain-final positions, where it is phonetically coded by a<br />
physical pause <strong>and</strong>/or pre-boundary leng<strong>the</strong>ning only.<br />
6 It has been argued by structuralists at least as far back as Merckens (1960) that<br />
V1 (‘verb first’) is directly opposed to V2 ('verb second') in signaling, for<br />
example, ‘non-assertion’ ra<strong>the</strong>r than ‘assertion’, since nei<strong>the</strong>r a comm<strong>and</strong> nor a<br />
question nor a condition expresses an ongoing state <strong>of</strong> affairs.<br />
7 A sequence like Neemt u de trein naar Wageningen might in addition be<br />
interpretable as a topic-drop-sentence (e.g. [Dan/Daar] neemt u de trein naar<br />
Wageningen ‘[Then/There] you take <strong>the</strong> train to Wageningen’, analogous to<br />
Doen we! ‘We'll do [it]’ or Weet ik! ‘[That] I know’. Although this added<br />
interpretation (with a ‘deleted’ element) is <strong>the</strong>oretically possible, we believe<br />
that it was highly unlikely un<strong>der</strong> <strong>the</strong> controlled conditions <strong>of</strong> <strong>the</strong> experiment.<br />
Fur<strong>the</strong>rmore, none <strong>of</strong> <strong>the</strong> experimental subjects volunteered <strong>the</strong> information<br />
that we had forgotten such an extra interpretation.<br />
8 This position does not exclude <strong>the</strong> possibility that statement <strong>and</strong> imperative are<br />
subtly different in <strong>the</strong>ir paralinguistic use <strong>of</strong> prosody. For instance, <strong>the</strong> overall<br />
pitch <strong>of</strong> <strong>the</strong> imperative may be lower, <strong>and</strong> it may be said with greater loudness<br />
<strong>and</strong> larger/higher pitch excursions on <strong>the</strong> accented syllables. This does not<br />
invalidate our claim that both statements <strong>and</strong> imperatives are coded by <strong>the</strong> ‘L%’<br />
terminal boundary.<br />
9 The ERB scale (Equivalent Rectangular B<strong>and</strong>width) is currently held to be <strong>the</strong><br />
most satisfactory psychophysical conversion for pitch intervals in human<br />
speech (Hermes & van Gestel, 1991; Ladd & Terken, 1995). The conversion<br />
from Hertz (f) to ERB (E) is achieved by a simple formula: E = 16.6 * log (1 + f<br />
/ 165.4).
58 Vincent J. van Heuven<br />
References<br />
Boersma, P. <strong>and</strong> Heuven, V.J. van (2001). Speak <strong>and</strong> unSpeak with Praat. Glot<br />
International, 5: 341-347.<br />
Boersma, P. <strong>and</strong> Weenink, D. (1996). Praat, a System for Doing <strong>Phonetics</strong> by<br />
Computer. Report <strong>of</strong> <strong>the</strong> Institute <strong>of</strong> Phonetic Sciences<br />
Amsterdam, 132.<br />
Caspers, J. (1998). ‘Who’s Next? The Melodic Marking <strong>of</strong> Question vs.<br />
Continuation in Dutch. Language <strong>and</strong> Speech, 41: 375-398.<br />
Gussenhoven, C., Rietveld, T. <strong>and</strong> Terken, J.M.B. (1999). Transcription <strong>of</strong><br />
Dutch Intonation. http://l<strong>and</strong>s.let.kun. nl/todi.<br />
Haan, J. (2002). Speaking <strong>of</strong> questions. An Exploration <strong>of</strong> Dutch Question<br />
Intonation. LOT Dissertation Series, nr. 52, Utrecht: LOT.<br />
Hart, J. 't, Collier, R. <strong>and</strong> Cohen, A. (1990). A Perceptual Study <strong>of</strong> Intonation.<br />
An Experimental-phonetic Approach to Speech Perception.<br />
Cambridge: Cambridge University Press.<br />
Hermes, D.J. <strong>and</strong> Gestel, J.C. van (1991). ‘The Frequency Scale <strong>of</strong> Speech<br />
Intonation. Journal <strong>of</strong> <strong>the</strong> Acoustical Society <strong>of</strong> America, 90: 97-<br />
102.<br />
Heuven, V.J. van (1986). Some acoustic characteristics <strong>and</strong> perceptual<br />
consequences <strong>of</strong> foreign accent in Dutch spoken by Turkish<br />
immigrant workers. In: J. van Oosten, J.F. Snapper (eds.) Dutch<br />
Linguistics at Berkeley, papers presented at <strong>the</strong> Dutch Linguistics<br />
Colloquium held at <strong>the</strong> University <strong>of</strong> California, Berkeley on<br />
November 9th, 1985, Berkeley: The Dutch Studies Program, U.C.<br />
Berkeley, 67-84.<br />
Heuven, V.J. van <strong>and</strong> Kirsner, R.S. (2002). Interaction <strong>of</strong> tone <strong>and</strong> particle in<br />
<strong>the</strong> signaling <strong>of</strong> clause type in Dutch. In: H. Broekhuis, P. Fikkert<br />
(eds.). Linguistics in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s 2002, Amsterdam<br />
/Philadelphia: John Benjamins, 73-84.<br />
Kirsner, R.S. <strong>and</strong> Heuven, V.J. van (1996). Boundary Tones <strong>and</strong> <strong>the</strong> Semantics<br />
<strong>of</strong> <strong>the</strong> Dutch Final Particles hè, hoor, zeg <strong>and</strong> joh. In: M. den<br />
Dikken, C. Cremers, eds., Linguistics in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s 1996,<br />
Amsterdam/Philadelphia: John Benjamins, 133-146.<br />
Kirsner, R.S., Heuven, V.J. van, <strong>and</strong> Caspers, J. (1998). From Request to<br />
Comm<strong>and</strong>: An Exploratory Experimental Study <strong>of</strong> Grammatical<br />
Form, Intonation, <strong>and</strong> Pragmatic Particle in Dutch Imperatives.<br />
In: R. van Bezooijen, R. Kager, eds., Linguistics in <strong>the</strong><br />
Ne<strong>the</strong>rl<strong>and</strong>s 1998. Amsterdam/Philadelphia: John Benjamins,<br />
135-148.
Boundary Tones in Dutch: Phonetic or Phonological Contrasts? 59<br />
Ladd, D.R. (1996). Intonational phonology. Cambridge: Cambridge University<br />
Press.<br />
Ladd, D.R. <strong>and</strong> Morton, R. (1997). The perception <strong>of</strong> intonational emphasis:<br />
continuous or categorical? Journal <strong>of</strong> <strong>Phonetics</strong>, 25: 313-342.<br />
Ladd, D.R. <strong>and</strong> Terken, J.M.B. (1995). Modelling intra- <strong>and</strong> inter-speaker pitch<br />
range variation. Proceedings <strong>of</strong> <strong>the</strong> 13 th International Congress <strong>of</strong><br />
Phonetic Sciences, Stockholm, 2: 386-389.<br />
Merckens, P.J. (1960). De plaats van de persoonsvorm: een verwaarloosd codeteken<br />
[The position <strong>of</strong> <strong>the</strong> finite verb: a neglected code sign]. De<br />
nieuwe taalgids, 53: 248-54.<br />
Moulines, E. <strong>and</strong> Verhelst, E. (1995). ‘Time-domain <strong>and</strong> frequency-domain<br />
techniques for prosodic modification <strong>of</strong> speech’. In: W.B. Kleijn<br />
<strong>and</strong> K.K. Paliwal, eds., Speech coding <strong>and</strong> syn<strong>the</strong>sis. Amsterdam:<br />
Elsevier Science, 519-555.<br />
Nooteboom, S.G. <strong>and</strong> Cohen, A. (1976). Spreken en verstaan. Een inleiding tot<br />
de experimentele fonetiek [Speaking <strong>and</strong> un<strong>der</strong>st<strong>and</strong>ing. An<br />
introduction to experimental phonetics], Assen: van Gorcum.<br />
Pitrelli, J.F., Beckman, M.E. <strong>and</strong> Hirschberg, J. (1994). Evaluation <strong>of</strong> prosodic<br />
transcription reliability in <strong>the</strong> ToBI framework. Proceedings <strong>of</strong><br />
<strong>the</strong> 3rd International Conference on Spoken Language<br />
Processing, Yokohama, 1: 123-126.<br />
Remijsen, A.C. <strong>and</strong> Heuven, V.J. van (1999). Gradient <strong>and</strong> categorical pitch<br />
dimensions in Dutch: Diagnostic test’. Proceedings <strong>of</strong> <strong>the</strong> 14th<br />
International Congress <strong>of</strong> Phonetic Sciences, San Francisco,<br />
1865-1868.<br />
Remijsen, A.C. <strong>and</strong> Heuven, V.J. van (2003). Linguistic versus paralinguistic<br />
status <strong>of</strong> prosodic contrasts, <strong>the</strong> case <strong>of</strong> high <strong>and</strong> low pitch in<br />
Dutch. In: J.M. van de Weijer, V.J. van Heuven, H.G. van <strong>der</strong><br />
Hulst (eds.): The phonological spectrum. Volume II:<br />
Suprasegmental structure. Current Issues in Linguistic Theory nr.<br />
235. Amsterdam/Philadelphia: John Benjamins, 225-246.<br />
Rietveld, A.C.M. <strong>and</strong> Heuven, V.J. van (2001). Algemene Fonetiek [General<br />
<strong>Phonetics</strong>]. Bussum: Coutinho.<br />
Slis, I.H. <strong>and</strong> Cohen, A. (1969). <strong>On</strong> <strong>the</strong> complex regulating <strong>the</strong> voicedvoiceless<br />
distinction, Language <strong>and</strong> Speech, 80-102: 137-155.<br />
Taylor, P. (1998). Analysis <strong>and</strong> syn<strong>the</strong>sis <strong>of</strong> intonation using <strong>the</strong> TILT model.<br />
Unpublished manuscript, Centre for Speech Technology<br />
Research, University <strong>of</strong> Edinburgh.
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language<br />
Area<br />
Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
1. Introduction<br />
Among <strong>the</strong> Germanic varieties <strong>the</strong> Frisian varieties in <strong>the</strong> Dutch province<br />
<strong>of</strong> Friesl<strong>and</strong> have <strong>the</strong>ir own position. The Frisians are proud <strong>of</strong> <strong>the</strong>ir<br />
language <strong>and</strong> more than 350,000 inhabitants <strong>of</strong> <strong>the</strong> province <strong>of</strong> Friesl<strong>and</strong><br />
speak Frisian every day. Heeringa (2004) shows that among <strong>the</strong> dialects in<br />
<strong>the</strong> Dutch language area <strong>the</strong> Frisian varieties are most distant with respect<br />
to st<strong>and</strong>ard Dutch. This may justify <strong>the</strong> fact that Frisian is recognized as a<br />
second <strong>of</strong>ficial language in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s. In addition to Frisian, in some<br />
towns <strong>and</strong> on some isl<strong>and</strong>s a mixed variety is used which is an intermediate<br />
form between Frisian <strong>and</strong> Dutch. The variety spoken in <strong>the</strong> Frisian towns is<br />
known as Town Frisian 1 .<br />
The Frisian language has existed for more than 2000 years. Genetically<br />
<strong>the</strong> Frisian dialects are most closely related to <strong>the</strong> English language.<br />
However, historical events have caused <strong>the</strong> English <strong>and</strong> <strong>the</strong> Frisian<br />
language to diverge, while Dutch <strong>and</strong> Frisian have converged. The<br />
linguistic distance to <strong>the</strong> o<strong>the</strong>r Germanic languages has also altered in <strong>the</strong><br />
course <strong>of</strong> history due to different degrees <strong>of</strong> linguistic contact. As a result<br />
traditional genetic trees do not give an up-to-date representation <strong>of</strong> <strong>the</strong><br />
distance between <strong>the</strong> mo<strong>der</strong>n Germanic languages.<br />
In <strong>the</strong> present investigation we measured linguistic distances between<br />
Frisian <strong>and</strong> <strong>the</strong> o<strong>the</strong>r Germanic languages in or<strong>der</strong> to get an impression <strong>of</strong><br />
<strong>the</strong> effect <strong>of</strong> genetic relationship <strong>and</strong> language contact for <strong>the</strong> position <strong>of</strong><br />
<strong>the</strong> mo<strong>der</strong>n Frisian language on <strong>the</strong> Germanic language map. We included<br />
six Frisian varieties <strong>and</strong> one Town Frisian variety in <strong>the</strong> investigation.<br />
Fur<strong>the</strong>rmore, eight Germanic st<strong>and</strong>ard languages were taken into account.<br />
Using this material, we firstly wished to obtain a hierarchical classification<br />
<strong>of</strong> <strong>the</strong> Germanic varieties. From this classification <strong>the</strong> position <strong>of</strong> (Town)
62 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
Frisian became clear. Secondly, we ranked all varieties with respect to each<br />
<strong>of</strong> <strong>the</strong> st<strong>and</strong>ard Germanic languages as well as to (Town) Frisian. The<br />
rankings showed <strong>the</strong> position <strong>of</strong> (Town) Frisian with respect to <strong>the</strong> st<strong>and</strong>ard<br />
languages <strong>and</strong> <strong>the</strong> position <strong>of</strong> <strong>the</strong> st<strong>and</strong>ard languages with respect to<br />
(Town) Frisian.<br />
In or<strong>der</strong> to obtain a classification <strong>of</strong> varieties <strong>and</strong> establish rankings, we<br />
needed a tool that can measure linguistic distances between <strong>the</strong> varieties.<br />
Bolognesi <strong>and</strong> Heeringa (2002) investigated <strong>the</strong> position <strong>of</strong> Sardinian<br />
dialects with respect to different Romance languages using <strong>the</strong> Levenshtein<br />
distance, an algorithm with which distances between word pronunciations<br />
are calculated. In our investigation we used <strong>the</strong> same methodology.<br />
In Section 2, we will present <strong>the</strong> traditional ideas about <strong>the</strong> genetic<br />
relationship between <strong>the</strong> Germanic languages <strong>and</strong> discuss <strong>the</strong> relationship<br />
between Frisian <strong>and</strong> <strong>the</strong> o<strong>the</strong>r Germanic languages. At <strong>the</strong> end <strong>of</strong> <strong>the</strong><br />
section we will discuss <strong>the</strong> expected outcome <strong>of</strong> <strong>the</strong> linguistic distance<br />
measurements between Frisian <strong>and</strong> <strong>the</strong> o<strong>the</strong>r Germanic languages. In<br />
Section 3 <strong>the</strong> data sources are described <strong>and</strong> in Section 4 <strong>the</strong> method for<br />
measuring linguistic distances between <strong>the</strong> language varieties is presented.<br />
The results are presented in Section 5, <strong>the</strong> discussion <strong>of</strong> which is presented<br />
in Section 6.<br />
2. Frisian <strong>and</strong> <strong>the</strong> Germanic languages<br />
2.1. History <strong>and</strong> classification <strong>of</strong> <strong>the</strong> Germanic languages 2<br />
The Germanic branch <strong>of</strong> <strong>the</strong> Indo-European languages has a large number<br />
<strong>of</strong> speakers, approximately 450 million native speakers, partly due to <strong>the</strong><br />
colonization <strong>of</strong> many parts <strong>of</strong> <strong>the</strong> world. However, <strong>the</strong> number <strong>of</strong> different<br />
languages within <strong>the</strong> Germanic group is ra<strong>the</strong>r limited. Depending on <strong>the</strong><br />
definition <strong>of</strong> what counts as a language <strong>the</strong>re are about 12 different<br />
languages. Traditionally, <strong>the</strong>y are divided into three subgroups: East<br />
Germanic (Gothic, which is no longer a living language), North Germanic<br />
(Icel<strong>and</strong>ic, Faeroese, Norwegian, Danish, <strong>and</strong> Swedish), <strong>and</strong> West<br />
Germanic (English, German, Dutch, Afrikaans, Yiddish, <strong>and</strong> Frisian).<br />
Some <strong>of</strong> <strong>the</strong>se languages are so similar that <strong>the</strong>y are only consi<strong>der</strong>ed<br />
independent languages because <strong>of</strong> <strong>the</strong>ir position as st<strong>and</strong>ardized languages
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 63<br />
spoken within <strong>the</strong> limits <strong>of</strong> a state. This goes for <strong>the</strong> languages <strong>of</strong> <strong>the</strong><br />
Sc<strong>and</strong>inavian countries, Swedish, Danish <strong>and</strong> Norwegian, which are<br />
mutually intelligible. O<strong>the</strong>r languages consist <strong>of</strong> dialects which are in fact<br />
so different that <strong>the</strong>y are no longer mutually intelligible but are still<br />
consi<strong>der</strong>ed one language because <strong>of</strong> st<strong>and</strong>ardization. Nor<strong>the</strong>rn <strong>and</strong> sou<strong>the</strong>rn<br />
German dialects are an example <strong>of</strong> this situation.<br />
Figure 3. The genetic tree <strong>of</strong> Germanic languages.<br />
In Figure 1, a traditional Germanic genetic tree is shown. We constructed<br />
this tree on <strong>the</strong> basis <strong>of</strong> data in <strong>the</strong> literature. The tree gives just a rough<br />
division, <strong>and</strong> linguistic distances should not be <strong>der</strong>ived from this tree. It is<br />
commonly assumed that <strong>the</strong> Germanic languages originate from <strong>the</strong><br />
sou<strong>the</strong>rn Sc<strong>and</strong>inavian <strong>and</strong> <strong>the</strong> nor<strong>the</strong>rn German region. After <strong>the</strong> migration<br />
<strong>of</strong> <strong>the</strong> Goths to <strong>the</strong> Balkans towards <strong>the</strong> end <strong>of</strong> <strong>the</strong> pre-Christian era, North-<br />
West Germanic remained uniform till <strong>the</strong> 5th century AD, after which a<br />
split between North <strong>and</strong> West Germanic occurred owing to dialectal<br />
variation <strong>and</strong> <strong>the</strong> departure <strong>of</strong> <strong>the</strong> Anglo-Saxons from <strong>the</strong> Continent <strong>and</strong> <strong>the</strong><br />
colonization <strong>of</strong> Jutl<strong>and</strong>.<br />
During <strong>the</strong> Viking Age, speakers <strong>of</strong> North Germanic settled in a large<br />
geographic area, which eventually led to <strong>the</strong> five mo<strong>der</strong>n languages (see<br />
above). Of <strong>the</strong>se languages, Icel<strong>and</strong>ic (<strong>and</strong> to a lesser degree Faeroese),<br />
which is based on <strong>the</strong> language <strong>of</strong> southwestern Norway where <strong>the</strong> settlers<br />
came from, can be consi<strong>der</strong>ed <strong>the</strong> most conservative language (S<strong>and</strong>øy,<br />
1994). Of <strong>the</strong> three mainl<strong>and</strong> Sc<strong>and</strong>inavian languages, Danish has moved
64 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
far<strong>the</strong>st away from <strong>the</strong> common Sc<strong>and</strong>inavian roots due to influences from<br />
<strong>the</strong> south.<br />
The parentage <strong>of</strong> <strong>the</strong> West Germanic languages is less clear. Different<br />
tribal groups representing different dialect groups spread across <strong>the</strong> area,<br />
which eventually resulted in <strong>the</strong> mo<strong>der</strong>n language situation. Historically<br />
Frisian <strong>and</strong> English both belong to <strong>the</strong> Ingwaeonic branch <strong>of</strong> <strong>the</strong> West<br />
Germanic language group. Originally <strong>the</strong> Frisian speech community<br />
extended from <strong>the</strong> present Danish-German bor<strong>der</strong> along <strong>the</strong> coast to <strong>the</strong><br />
French-Belgian bor<strong>der</strong> in <strong>the</strong> south. However, expansion from Saxons <strong>and</strong><br />
Franconians from <strong>the</strong> east <strong>and</strong> <strong>the</strong> south throughout <strong>the</strong> medieval period<br />
resulted in a loss <strong>of</strong> large Frisian areas <strong>and</strong> a division into three mutually<br />
intelligible varieties: West Frisian (spoken in <strong>the</strong> nor<strong>the</strong>rn Dutch province<br />
<strong>of</strong> Friesl<strong>and</strong> by more than 350,000 people), East Frisian or Saterl<strong>and</strong>ic<br />
(spoken by a thous<strong>and</strong> speakers in three villages west <strong>of</strong> Bremen) <strong>and</strong><br />
North Frisian (spoken by less than ten thous<strong>and</strong> people on <strong>the</strong> isl<strong>and</strong>s on<br />
<strong>the</strong> north-western coast <strong>of</strong> Germany).<br />
The English language came into being as a result <strong>of</strong> immigrations <strong>of</strong><br />
tribal Anglo-Saxon groups from <strong>the</strong> North Sea coast during <strong>the</strong> fifth <strong>and</strong><br />
sixth centuries. Whereas o<strong>the</strong>r insular Germanic varieties are in general<br />
ra<strong>the</strong>r conservative, <strong>the</strong> English insularity lacked this conservatism. English<br />
is consi<strong>der</strong>ed most closely related to Frisian on every linguistic level due to<br />
<strong>the</strong>ir common ancestorship <strong>and</strong> to continued language contact over <strong>the</strong><br />
North Sea.<br />
The German language is spoken in many European countries in a large<br />
number <strong>of</strong> dialects <strong>and</strong> varieties, which can be divided into Low German<br />
<strong>and</strong> High German. Yiddish, too, can be regarded as a German variety.<br />
Dutch is mainly based on <strong>the</strong> western varieties <strong>of</strong> <strong>the</strong> low Franconian area<br />
but low Saxon <strong>and</strong> Frisian elements are also found in this st<strong>and</strong>ard<br />
language. Scholars disagree about <strong>the</strong> precise position <strong>of</strong> Dutch <strong>and</strong> Low<br />
German in <strong>the</strong> language tree. They can be traced back to a common root<br />
<strong>of</strong>ten referred to as <strong>the</strong> Ingwaeonic language group, but are <strong>of</strong>ten grouped<br />
toge<strong>the</strong>r with High German as a separate West Germanic group. This<br />
grouping with High German might be <strong>the</strong> best representation <strong>of</strong> <strong>the</strong> mo<strong>der</strong>n<br />
language situation given that <strong>the</strong> individual dialects spoken in <strong>the</strong> area in<br />
fact form a dialect continuum. Afrikaans, finally, is a contemporary West<br />
Germanic language, developed from seventeenth century Dutch as a result<br />
<strong>of</strong> colonization, but with influences from African languages.
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 65<br />
2.2. The relationship between Frisian <strong>and</strong> <strong>the</strong> o<strong>the</strong>r Germanic languages.<br />
This short outline <strong>of</strong> <strong>the</strong> relationships among <strong>the</strong> Germanic languages<br />
shows that English is <strong>the</strong> language which is genetically closest to Frisian,<br />
<strong>and</strong> still today English is consi<strong>der</strong>ed to be most similar to Frisian. For<br />
example The Columbia Encyclopedia (2001) says: “Of all foreign<br />
languages, [Frisian] is most like English”. Pei (1966, p. 34) summarizes <strong>the</strong><br />
situation as follows: “Frisian, a variant <strong>of</strong> Dutch spoken along <strong>the</strong> Dutch<br />
<strong>and</strong> German North Sea coast, is <strong>the</strong> foreign speech that comes closest to<br />
mo<strong>der</strong>n English, as shown by <strong>the</strong> rhyme: ‘Good butter <strong>and</strong> good cheese is<br />
good English <strong>and</strong> good Fries’”. This rhyme refers to <strong>the</strong> fact that <strong>the</strong> words<br />
for butter <strong>and</strong> cheese are almost <strong>the</strong> same in <strong>the</strong> two languages. However,<br />
in <strong>the</strong> course <strong>of</strong> history, contact with o<strong>the</strong>r Germanic languages has caused<br />
Frisian to converge to <strong>the</strong>se languages. The Frisians have a long history <strong>of</strong><br />
trade <strong>and</strong> in early medieval times <strong>the</strong>y were one <strong>of</strong> <strong>the</strong> leading trading<br />
nations in Europe due to <strong>the</strong>ir strategic geographic position close to major<br />
trade routes along <strong>the</strong> rivers <strong>and</strong> <strong>the</strong> North Sea. Also, <strong>the</strong> Vikings <strong>and</strong> <strong>the</strong><br />
English were frequent visitors <strong>of</strong> <strong>the</strong> Frisian language area. This intensive<br />
contact with both English <strong>and</strong> <strong>the</strong> North Germanic languages, especially<br />
Danish, resulted in linguistic exchanges (see Feitsma, 1963; Miedema,<br />
1966; Wadstein, 1933). Later in history, <strong>the</strong> Frisian language was<br />
especially influenced by <strong>the</strong> Dutch language (which itself contains many<br />
Frisian elements). For a long period, Frisian was stigmatized as a peasant<br />
language <strong>and</strong> due to <strong>the</strong> weak social position <strong>of</strong> <strong>the</strong> Frisian language in <strong>the</strong><br />
Dutch community it was <strong>of</strong>ten suppressed, resulting in a strong Dutch<br />
impact on <strong>the</strong> Frisian language. Nowadays, Dutch as <strong>the</strong> language <strong>of</strong> <strong>the</strong><br />
administration still has a large influence on <strong>the</strong> media <strong>and</strong> <strong>the</strong>re has been<br />
substantial immigration <strong>of</strong> Dutch speaking people to Friesl<strong>and</strong>. However,<br />
<strong>the</strong> provincial government has decided to promote Frisian at all levels in<br />
<strong>the</strong> society.<br />
When investigating <strong>the</strong> position <strong>of</strong> <strong>the</strong> Frisian language within <strong>the</strong><br />
Germanic language group, <strong>the</strong>re are clearly two forces which should be<br />
taken into account. <strong>On</strong> <strong>the</strong> one h<strong>and</strong>, Frisian <strong>and</strong> English are genetically<br />
closely related <strong>and</strong> share sound changes which do not occur in <strong>the</strong> o<strong>the</strong>r<br />
Germanic languages. This yields <strong>the</strong> expectation that <strong>the</strong> linguistic distance<br />
between <strong>the</strong>se two languages is relatively small. <strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, <strong>the</strong><br />
close contact with Dutch makes it plausible that <strong>the</strong> Dutch <strong>and</strong> <strong>the</strong> Frisian<br />
languages have converged. Also <strong>the</strong> distance to Danish might be smaller<br />
than expected from <strong>the</strong> traditional division <strong>of</strong> Germanic into a North
66 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
Germanic <strong>and</strong> a West Germanic branch at an early stage because <strong>of</strong> <strong>the</strong><br />
intensive contacts in <strong>the</strong> past.<br />
3. Data sources<br />
In this section, we will first give a short characterization <strong>of</strong> <strong>the</strong> language<br />
varieties <strong>and</strong> <strong>the</strong> speakers who were recorded for our investigation. Next,<br />
we will present <strong>the</strong> nature <strong>of</strong> <strong>the</strong> recordings <strong>and</strong> <strong>the</strong> transcriptions which<br />
formed <strong>the</strong> basis for linguistic distance measurements.<br />
3.1. Language varieties<br />
Since our main interest was <strong>the</strong> Frisian language <strong>and</strong> its linguistic position<br />
within <strong>the</strong> Germanic language group we wished to represent this language<br />
as well as possible. For this reason, we included seven Frisian varieties,<br />
spread over <strong>the</strong> Frisian language area. Fur<strong>the</strong>rmore, our material contained<br />
eight Germanic st<strong>and</strong>ard languages. First, we will describe <strong>the</strong> Frisian<br />
varieties <strong>and</strong> next <strong>the</strong> st<strong>and</strong>ard languages.<br />
As far as <strong>the</strong> Frisian varieties are concerned, we chose varieties from<br />
different parts <strong>of</strong> <strong>the</strong> province, both from <strong>the</strong> coastal area <strong>and</strong> from <strong>the</strong><br />
inl<strong>and</strong>. The varieties are spoken in different dialect areas according to <strong>the</strong><br />
traditional classification (see below) <strong>and</strong> <strong>the</strong>y represent different stages <strong>of</strong><br />
conservatism. The precise choice <strong>of</strong> <strong>the</strong> seven varieties was determined by<br />
speaker availability for recordings in our vicinity <strong>and</strong> at <strong>the</strong> Fryske<br />
Akademy in Leeuwarden. In Figure 2, <strong>the</strong> geographical position <strong>of</strong> <strong>the</strong><br />
seven Frisian language varieties in <strong>the</strong> province <strong>of</strong> Friesl<strong>and</strong> is shown.<br />
Due to <strong>the</strong> absence <strong>of</strong> major geographical barriers, <strong>the</strong> Frisian language<br />
area is relatively uniform. The major dialectal distinctions are primarily<br />
phonological. Traditionally, three main dialect areas are distinguished (see<br />
e.g. H<strong>of</strong>, 1933; Visser, 1997): Klaaifrysk (clay Frisian) in <strong>the</strong> west,<br />
Wâldfrysk (forest Frisian) in <strong>the</strong> east <strong>and</strong> Súdwesthoeksk (southwest<br />
quarter) in <strong>the</strong> southwest. In our material Klaaifrysk is represented by <strong>the</strong><br />
dialects <strong>of</strong> Oosterbierum <strong>and</strong> Hijum, Wâldfrysk by Wetsens <strong>and</strong><br />
Westergeest, <strong>and</strong> Súdwesthoeksk by Tjerkgaast. Hindeloopen is in <strong>the</strong> area<br />
<strong>of</strong> Súdwesthoeksk. However, this dialect represents a highly conservative<br />
area. The phonological distance between Hindeloopen <strong>and</strong> <strong>the</strong> main<br />
dialects is substantial (van <strong>der</strong> Veen, 2001). Finally, our material contains
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 67<br />
<strong>the</strong> variety spoken in Leeuwarden (see note 1). This is an example <strong>of</strong> Town<br />
Frisian, which is also spoken in o<strong>the</strong>r cities <strong>of</strong> Friesl<strong>and</strong>. Town Frisian is a<br />
Dutch dialect strongly influenced by Frisian but stripped <strong>of</strong> <strong>the</strong> most<br />
characteristic Frisian elements (Goossens, 1977).<br />
Oosterbierum<br />
Hindeloopen<br />
Hijum<br />
Leeuwarden<br />
Tjerkgaast<br />
Wetsens<br />
Westergeest<br />
Figure 2. The geographical position <strong>of</strong> <strong>the</strong> seven Frisian language varieties in <strong>the</strong><br />
province <strong>of</strong> Friesl<strong>and</strong>.<br />
In addition to <strong>the</strong> Frisian dialects, <strong>the</strong> following eight st<strong>and</strong>ard languages<br />
were included: Icel<strong>and</strong>ic, Faroese, Norwegian, Swedish, Danish, English,<br />
Dutch, <strong>and</strong> German. We had meant to include all st<strong>and</strong>ard Germanic<br />
languages in our material. However, due to practical limitations a few<br />
smaller languages were not included.<br />
As for Norwegian, <strong>the</strong>re is no <strong>of</strong>ficial st<strong>and</strong>ard variety. The varieties<br />
spoken around <strong>the</strong> capital <strong>of</strong> Oslo in <strong>the</strong> sou<strong>the</strong>ast, however, are <strong>of</strong>ten<br />
consi<strong>der</strong>ed to represent <strong>the</strong> st<strong>and</strong>ard language. We based <strong>the</strong> present<br />
investigation on prior research on Norwegian dialects (see Heeringa <strong>and</strong><br />
Gooskens, 2003; Gooskens <strong>and</strong> Heeringa, submitted), <strong>and</strong> we chose <strong>the</strong><br />
recording which to Norwegians sounded most st<strong>and</strong>ard, namely <strong>the</strong><br />
Lillehammer recording 3 . It was our aim to select st<strong>and</strong>ard speakers from all<br />
countries, but it is possible that <strong>the</strong> speech <strong>of</strong> some speakers contains slight<br />
regional influences. The speakers from Icel<strong>and</strong>, <strong>the</strong> Faroe Isl<strong>and</strong>s <strong>and</strong>
68 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
Sweden spoke <strong>the</strong> st<strong>and</strong>ard varieties <strong>of</strong> <strong>the</strong> capitals. The Danish speaker<br />
came from Jutl<strong>and</strong>, <strong>the</strong> German speaker from Kiel, <strong>the</strong> English speaker<br />
from Birmingham <strong>and</strong> <strong>the</strong> Dutch speaker had lived at different places in <strong>the</strong><br />
Ne<strong>the</strong>rl<strong>and</strong>s, including a long period in <strong>the</strong> West during adolescence.<br />
3.2. Phonetic transcriptions<br />
The speakers all read aloud translations <strong>of</strong> <strong>the</strong> same text, namely <strong>the</strong> fable<br />
‘The North Wind <strong>and</strong> <strong>the</strong> Sun’. This text has <strong>of</strong>ten been used for phonetic<br />
investigations; see for example The International Phonetic Association<br />
(1949 <strong>and</strong> 1999) where <strong>the</strong> same text has been transcribed in a large<br />
number <strong>of</strong> different languages. A database <strong>of</strong> Norwegian transcriptions <strong>of</strong><br />
<strong>the</strong> same text has been compiled by J. Almberg (see note 3). As mentioned<br />
in <strong>the</strong> previous section, we only used <strong>the</strong> transcription <strong>of</strong> Lillehammer from<br />
this database. In future, we would like to investigate <strong>the</strong> relations between<br />
Norwegian <strong>and</strong> o<strong>the</strong>r Germanic varieties, using <strong>the</strong> greater part <strong>of</strong> <strong>the</strong><br />
transcriptions in this database. Therefore, our new transcriptions should be<br />
as comparable as possible with <strong>the</strong> existing Norwegian ones. To ensure<br />
this, our point <strong>of</strong> departure was <strong>the</strong> Norwegian text. This text consists <strong>of</strong> 91<br />
words (58 different words) which were used to calculate Levenshtein<br />
distances (see Section 4). The text was translated word for word from<br />
Norwegian into each <strong>of</strong> <strong>the</strong> Germanic language varieties. We are aware <strong>of</strong><br />
<strong>the</strong> fact that this may result in less natural speech: sentences were <strong>of</strong>ten<br />
syntactically wrong. However, it guarantees that for each <strong>of</strong> <strong>the</strong> 58 words a<br />
translation was obtained. The words were not recorded as a word list, but as<br />
sentences. Therefore in <strong>the</strong> new recordings words appear in a similar<br />
context as in <strong>the</strong> Norwegian varieties. This ensures that <strong>the</strong> influence <strong>of</strong><br />
assimilation phenomena on <strong>the</strong> results is as comparable as possible.<br />
Most new recordings were transcribed phonetically by one <strong>of</strong> <strong>the</strong><br />
authors. To ensure consistency with <strong>the</strong> existing Norwegian transcriptions,<br />
our new transcriptions were corrected by J. Almberg, <strong>the</strong> transcriber <strong>of</strong> <strong>the</strong><br />
Norwegian recordings. In most cases we incorporated <strong>the</strong> corrections. The<br />
transcription <strong>of</strong> <strong>the</strong> Faroese language was completely done by J. Almberg.<br />
The transcriptions were made in IPA as well as in X-SAMPA (eXtended<br />
Speech Assessment Methods Phonetic Alphabet). This is a machinereadable<br />
phonetic alphabet, which is also readable by people. Basically, it<br />
maps IPA-symbols to <strong>the</strong> 7 bit printable ASCII/ANSI characters 4 . The
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 69<br />
transcriptions were used to calculate <strong>the</strong> linguistic distances between<br />
varieties (see Section 4).<br />
4. Measuring distances between varieties<br />
In 1995 Kessler introduced <strong>the</strong> use <strong>of</strong> <strong>the</strong> Levenshtein distance as tool for<br />
measuring linguistic distances between language varieties. The Levenshtein<br />
distance is a string edit distance measure <strong>and</strong> Kessler applied this algorithm<br />
to <strong>the</strong> comparison <strong>of</strong> Irish dialects. Later on, this approach was applied by<br />
Nerbonne, Heeringa, Van den Hout, Van <strong>der</strong> Kooi, Otten, <strong>and</strong> Van de Vis<br />
(1996) to Dutch dialects. They assumed that distances between all possible<br />
pairs <strong>of</strong> segments are <strong>the</strong> same. E.g. <strong>the</strong> distance between an [�] <strong>and</strong> an [e]<br />
is <strong>the</strong> same as <strong>the</strong> distance between <strong>the</strong> [�] <strong>and</strong> [�]. Both Kessler (1995) <strong>and</strong><br />
Nerbonne <strong>and</strong> Heeringa (1997) also experimented with more refined<br />
versions <strong>of</strong> <strong>the</strong> Levenshtein algorithm in which gradual segment distances<br />
were used which were found on <strong>the</strong> basis <strong>of</strong> <strong>the</strong> feature systems <strong>of</strong><br />
Hoppenbrouwers (1988) <strong>and</strong> Vieregge et. al. (1984).<br />
In this paper we use an implementation <strong>of</strong> <strong>the</strong> Levenshtein distance in<br />
which sound distances are used which are found by comparing<br />
spectrograms. In Section 4.1 we account for <strong>the</strong> use <strong>of</strong> spectral distances<br />
<strong>and</strong> explain how we calculate <strong>the</strong>m. Comparisons are made on <strong>the</strong> basis <strong>of</strong><br />
<strong>the</strong> audiotape The Sounds <strong>of</strong> <strong>the</strong> International Phonetic Alphabet (Wells<br />
<strong>and</strong> House, 1995). In Section 4.2 we describe <strong>the</strong> Levenshtein distance <strong>and</strong><br />
explain how spectral distances can be used in this algorithm.<br />
4.1. Gradual segment distances<br />
When acquiring language, children learn to pronounce sounds by listening<br />
to <strong>the</strong> pronunciation <strong>of</strong> <strong>the</strong>ir parents or o<strong>the</strong>r people. The acoustic signal<br />
seems to be sufficient to find <strong>the</strong> articulation which is needed to realize <strong>the</strong><br />
sound. Acoustically, speech is just a series <strong>of</strong> changes in air pressure,<br />
quickly following each o<strong>the</strong>r. A spectrogram is a “graph with frequency on<br />
<strong>the</strong> vertical axis <strong>and</strong> time on <strong>the</strong> horizontal axis, with <strong>the</strong> darkness <strong>of</strong> <strong>the</strong><br />
graph at any point representing <strong>the</strong> intensity <strong>of</strong> <strong>the</strong> sound” (Trask, 1996, p.<br />
328).<br />
In this section we present <strong>the</strong> use <strong>of</strong> spectrograms for finding segment<br />
distances. Segment distances can also be found on <strong>the</strong> basis <strong>of</strong> phonological
70 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
or phonetic feature systems. However, we prefer <strong>the</strong> use <strong>of</strong> acoustic<br />
representations since <strong>the</strong>y are based on physical measurements. In Potter,<br />
Kopp <strong>and</strong> Green’s (1947) Visible Speech, spectrograms are shown for all<br />
common English sounds (see pp. 54-56). Looking at <strong>the</strong> spectrograms we<br />
already see which sounds are similar <strong>and</strong> which are not. We assume that<br />
visible (dis)similarity between spectrograms reflects perceptual<br />
(dis)similarity between segments to some extent. In Figure 3 <strong>the</strong><br />
spectrograms <strong>of</strong> some sounds are shown as pronounced by John Wells on<br />
<strong>the</strong> audiotape The Sounds <strong>of</strong> <strong>the</strong> International Phonetic Alphabet (Wells<br />
<strong>and</strong> House, 1995). The spectrograms are made with <strong>the</strong> computer program<br />
PRAAT 5 .<br />
Figure 3. Spectrograms <strong>of</strong> some sounds pronounced by John Wells. Upper <strong>the</strong> [i]<br />
(left) <strong>and</strong> <strong>the</strong> [e] (right) are shown, <strong>and</strong> lower <strong>the</strong> [p] (left) <strong>and</strong> <strong>the</strong> [s]<br />
(right) are visualized.<br />
4.1.1. Samples<br />
For finding spectrogram distances between all IPA segments we need<br />
samples <strong>of</strong> one or more speakers for each <strong>of</strong> <strong>the</strong>m. We found <strong>the</strong> samples<br />
on <strong>the</strong> tape The Sounds <strong>of</strong> <strong>the</strong> International Phonetic Alphabet on which all
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 71<br />
IPA sounds are pronounced by John Wells <strong>and</strong> Jill House. <strong>On</strong> <strong>the</strong> tape <strong>the</strong><br />
vowels are pronounced in isolation. The consonants are sometimes<br />
preceded, <strong>and</strong> always followed by an [a]. We cut out <strong>the</strong> part preceding <strong>the</strong><br />
[a], or <strong>the</strong> part between <strong>the</strong> [a]’s. We realize that <strong>the</strong> pronunciation <strong>of</strong><br />
sounds depends on <strong>the</strong>ir context. Since we use samples <strong>of</strong> vowels<br />
pronounced in isolation <strong>and</strong> samples <strong>of</strong> consonants selected from a limited<br />
context, our approach is a simplification <strong>of</strong> reality. However, Stevens<br />
(1998, p. 557) observes that<br />
“by limiting <strong>the</strong> context, it was possible to specify ra<strong>the</strong>r precisely <strong>the</strong><br />
articulatory aspects <strong>of</strong> <strong>the</strong> utterances <strong>and</strong> to develop models for estimating<br />
<strong>the</strong> acoustic patterns from <strong>the</strong> articulation”.<br />
The burst in a plosive <strong>of</strong> <strong>the</strong> IPA inventory is always preceded by a period<br />
<strong>of</strong> silence (voiceless plosives) or a period <strong>of</strong> murmur (voiced plosives).<br />
When a voiceless plosive is not preceded by an [a], it is not clear how long<br />
<strong>the</strong> period <strong>of</strong> silence which really belongs to <strong>the</strong> sounds lasts. Therefore we<br />
always cut out each plosive in such a way that <strong>the</strong> time span from <strong>the</strong><br />
beginning to <strong>the</strong> middle <strong>of</strong> <strong>the</strong> burst is equal to 90 ms. Among <strong>the</strong> plosives<br />
which were preceded by an [a] or which are voiced (so that <strong>the</strong> real time <strong>of</strong><br />
<strong>the</strong> start-up phase can be found) we found no sounds with a period <strong>of</strong><br />
silence or murmur which was clearly shorter than 90 ms.<br />
In voiceless plosives, <strong>the</strong> burst is followed by an [h]-like sound before<br />
<strong>the</strong> following vowel starts. A consequence <strong>of</strong> including this part in <strong>the</strong><br />
samples is that bursts <strong>of</strong>ten do not match when comparing two voiceless<br />
plosives. However, since aspiration is a characteristic property <strong>of</strong> voiceless<br />
sounds, we retained aspiration in <strong>the</strong> samples. In general, when comparing<br />
two voiced plosives, <strong>the</strong> bursts match. When comparing a voiceless plosive<br />
<strong>and</strong> a voiced plosive, <strong>the</strong> bursts do not match.<br />
To keep trills comparable to each o<strong>the</strong>r, we always cut three periods,<br />
even when <strong>the</strong> original samples contained more periods. When <strong>the</strong>re were<br />
more periods, <strong>the</strong> most regular looking sequence <strong>of</strong> three periods was cut.<br />
The Levenshtein algorithm also requires a definition <strong>of</strong> ‘silence’. To get<br />
a sample <strong>of</strong> ‘silence’ we cut a small silent part on <strong>the</strong> IPA tape. This<br />
assures that silence has approximately <strong>the</strong> same background noise as <strong>the</strong><br />
o<strong>the</strong>r sounds.<br />
To make <strong>the</strong> samples as comparable as possible, all vowel <strong>and</strong> extracted<br />
consonant samples are monotonized on <strong>the</strong> mean pitch <strong>of</strong> <strong>the</strong> 28<br />
concatenated vowels. The mean pitch <strong>of</strong> John Wells was 128 Hertz; <strong>the</strong>
72 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
mean pitch <strong>of</strong> Jill House was 192 Hertz. In or<strong>der</strong> to monotonize <strong>the</strong><br />
samples <strong>the</strong> pitch contours were changed to flat lines. The volume was not<br />
normalized because volume contains too much segment specific<br />
information. For example it is specific for <strong>the</strong> [v] that its volume is greater<br />
than that <strong>of</strong> <strong>the</strong> [f].<br />
4.1.2. Acoustic representation<br />
In <strong>the</strong> most common type <strong>of</strong> spectrogram <strong>the</strong> linear Hertz frequency scale is<br />
used. The difference between 100 Hz <strong>and</strong> 200 Hz is <strong>the</strong> same as <strong>the</strong><br />
difference between 1000 Hz <strong>and</strong> 1100 Hz. However, our perception <strong>of</strong><br />
frequency is non-linear. We hear <strong>the</strong> difference between 100 Hz <strong>and</strong> 200<br />
Hz as an octave interval, but also <strong>the</strong> difference between 1000 Hz <strong>and</strong> 2000<br />
Hz is perceived as an octave. Our ear evaluates frequency differences not<br />
absolutely, but relatively, namely in a logarithmic manner. Therefore, in <strong>the</strong><br />
Barkfilter, <strong>the</strong> Bark-scale is used which is roughly linear below 1000 Hz<br />
<strong>and</strong> roughly logarithmic above 1000 Hz (Zwicker <strong>and</strong> Feldtkeller, 1967).<br />
In <strong>the</strong> commonly used type <strong>of</strong> spectrogram <strong>the</strong> power spectral density is<br />
represented per frequency per time. The power spectral density is <strong>the</strong> power<br />
per unit <strong>of</strong> frequency as a function <strong>of</strong> <strong>the</strong> frequency. In <strong>the</strong> Barkfilter <strong>the</strong><br />
power spectral density is expressed in decibels (dB’s). “The decibel scale is<br />
a way <strong>of</strong> expressing sound amplitude that is better correlated with<br />
perceived loudness” (Johnson, 1997, p. 53). The decibel scale is a<br />
logarithmic scale. Multiplying <strong>the</strong> sound pressure ten times corresponds to<br />
an increase <strong>of</strong> 20 dB. <strong>On</strong> a decibel scale intensities are expressed relative to<br />
<strong>the</strong> auditory threshold. The auditory threshold <strong>of</strong> 0.00002 Pa corresponds<br />
with 0 dB (Rietveld <strong>and</strong> Van Heuven, 1997, p. 199).<br />
A Barkfilter is created from a sound by b<strong>and</strong> filtering in <strong>the</strong> frequency<br />
domain with a bank <strong>of</strong> filters. In PRAAT <strong>the</strong> lowest b<strong>and</strong> has a central<br />
frequency <strong>of</strong> 1 Bark per default, <strong>and</strong> each b<strong>and</strong> has a width <strong>of</strong> 1 Bark.<br />
There are 24 b<strong>and</strong>s, corresponding to <strong>the</strong> first 24 critical b<strong>and</strong>s <strong>of</strong> hearing<br />
as found along <strong>the</strong> basilar membrane (Zwicker <strong>and</strong> Fastl, 1990). A critical<br />
b<strong>and</strong> is an area within which two tones influence each o<strong>the</strong>r’s perceptibility<br />
(Rietveld <strong>and</strong> Van Heuven, 1997). Due to <strong>the</strong> Bark-scale <strong>the</strong> higher b<strong>and</strong>s<br />
summarize a wi<strong>der</strong> frequency range than <strong>the</strong> lower b<strong>and</strong>s.<br />
In PRAAT we used <strong>the</strong> default settings when using <strong>the</strong> Barkfilter. The<br />
sound signal is probed each 0.005 seconds with an analysis window <strong>of</strong><br />
0.015 seconds. O<strong>the</strong>r settings may give different results, but since it was
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 73<br />
not a priori obvious which results are optimal, we restricted ourselves to <strong>the</strong><br />
default settings. In Figure 4 Barkfilters for some segments are shown.<br />
Figure 4. Barkfilter spectrograms <strong>of</strong> some sounds pronounced by John Wells.<br />
Upper <strong>the</strong> [i] (left) <strong>and</strong> <strong>the</strong> [e] (right) are shown, <strong>and</strong> lower <strong>the</strong><br />
[p] (left) <strong>and</strong> <strong>the</strong> [s] (right) are visualized.<br />
4.1.3. Comparison<br />
In this section, we explain <strong>the</strong> comparison <strong>of</strong> segments in or<strong>der</strong> to get<br />
distances between segments that will be used in <strong>the</strong> Levenshtein distance<br />
measure. In a Barkfilter, <strong>the</strong> intensities <strong>of</strong> frequencies are given for a range<br />
<strong>of</strong> times. A spectrum contains <strong>the</strong> intensities <strong>of</strong> frequencies at one time.<br />
The smaller <strong>the</strong> time step, <strong>the</strong> more spectra <strong>the</strong>re are in <strong>the</strong> acoustic<br />
representation. We consistently used <strong>the</strong> same time step for all samples.<br />
It appears that <strong>the</strong> duration <strong>of</strong> <strong>the</strong> segment samples varies. This may be<br />
explained by variation in speech rate. Duration is also a sound-specific<br />
property. E.g., a plosive is shorter than a vowel. The result is that <strong>the</strong><br />
number <strong>of</strong> spectra per segment may vary, although for each segment <strong>the</strong><br />
same time step was used. Since we want to normalize <strong>the</strong> speech rate <strong>and</strong><br />
regard segments as linguistic units, we made sure that two segments get <strong>the</strong><br />
same number <strong>of</strong> spectra when <strong>the</strong>y are compared to each o<strong>the</strong>r.
74 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
When comparing one segment <strong>of</strong> m spectra with ano<strong>the</strong>r segment <strong>of</strong> n<br />
spectra, each <strong>of</strong> <strong>the</strong> m elements is duplicated n times, <strong>and</strong> each <strong>of</strong> <strong>the</strong> n<br />
elements is duplicated m times. So both segments get a length <strong>of</strong> m × n.<br />
In or<strong>der</strong> to find <strong>the</strong> distance between two sounds, <strong>the</strong> Euclidean distance<br />
is calculated between each pair <strong>of</strong> corresponding spectra, one from each <strong>of</strong><br />
<strong>the</strong> sounds. Assume a spectrum e1 <strong>and</strong> e2 with n frequencies, <strong>the</strong>n <strong>the</strong><br />
Euclidean distance is:<br />
Equation 1. Euclidean distance<br />
The distance between two segments is equal to <strong>the</strong> sum <strong>of</strong> <strong>the</strong> spectrum<br />
distances divided by <strong>the</strong> number <strong>of</strong> spectra. In this way we found that <strong>the</strong><br />
greatest distance occurs between <strong>the</strong> [a] <strong>and</strong> ‘silence’. We regard this<br />
maximum distance as 100%. O<strong>the</strong>r segment distances are divided by this<br />
maximum <strong>and</strong> multiplied by 100. This yields segment distances expressed<br />
in percentages. Word distances <strong>and</strong> distances between varieties which are<br />
based on <strong>the</strong>m may also be given in terms <strong>of</strong> percentages.<br />
In perception, small differences in pronunciation may play a relatively<br />
strong role in comparison with larger differences. Therefore we used<br />
logarithmic segment distances. The effect <strong>of</strong> using logarithmic distances is<br />
that small distances are weighed relatively more heavily than large<br />
distances. Since <strong>the</strong> logarithm <strong>of</strong> 0 is not defined, <strong>and</strong> <strong>the</strong> logarithm <strong>of</strong> 1 is<br />
0, distances are increased by 1 before <strong>the</strong> logarithm is calculated. To obtain<br />
percentages, we calculate ln(distance + 1) / ln(maximum distance + 1).<br />
4.1.4. Suprasegmentals <strong>and</strong> diacritics<br />
The sounds on <strong>the</strong> tape The Sounds <strong>of</strong> <strong>the</strong> International Phonetic Alphabet<br />
are pronounced without suprasegmentals <strong>and</strong> diacritics. However, a<br />
restricted set <strong>of</strong> suprasegmentals <strong>and</strong> diacritics can be processed in our<br />
system.<br />
Length marks <strong>and</strong> syllabification are processed by changing <strong>the</strong><br />
transcription beforeh<strong>and</strong>. In <strong>the</strong> X-SAMPA transcription, extra-short
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 75<br />
segments are kept unchanged, sounds with no length indication are<br />
doubled, half long sounds are trebled, <strong>and</strong> long sounds are quadrupled.<br />
Syllabic sounds are treated as long sounds, so <strong>the</strong>y are quadrupled.<br />
When processing <strong>the</strong> diacritics voiceless <strong>and</strong>/or voiced, we assume that a<br />
voiced voiceless segment (e.g. [��]) <strong>and</strong> a voiceless voiced segment (e.g. [d�])<br />
are intermediate pronunciations <strong>of</strong> a voiceless segment ([t]) <strong>and</strong> a voiced<br />
segment ([d]). Therefore we calculate <strong>the</strong> distance between a segment x <strong>and</strong><br />
a voiced segment y as <strong>the</strong> average <strong>of</strong> <strong>the</strong> distance between x <strong>and</strong> y <strong>and</strong> <strong>the</strong><br />
distance between x <strong>and</strong> <strong>the</strong> voiced counterpart <strong>of</strong> y. Similarly, <strong>the</strong> distance<br />
between a segment x <strong>and</strong> a voiceless segment y is calculated as <strong>the</strong> mean <strong>of</strong><br />
<strong>the</strong> distance between x <strong>and</strong> y <strong>and</strong> <strong>the</strong> distance between x <strong>and</strong> <strong>the</strong> voiceless<br />
counterpart <strong>of</strong> y. For voiced sounds which have no voiceless counterpart<br />
(<strong>the</strong> sonorants), or for voiceless sounds which have no voiced counterpart<br />
(<strong>the</strong> glottal stop) <strong>the</strong> sound itself is used.<br />
The diacritic apical is only processed for <strong>the</strong> [s] <strong>and</strong> <strong>the</strong> [z]. We<br />
calculate <strong>the</strong> distance between [s�] <strong>and</strong> e.g. [f] as <strong>the</strong> average <strong>of</strong> <strong>the</strong> distance<br />
between [s] <strong>and</strong> [f] <strong>and</strong> [�] <strong>and</strong> [f]. Similarly, <strong>the</strong> distance between [z�] <strong>and</strong><br />
e.g. [v] is calculated as <strong>the</strong> mean <strong>of</strong> [z] <strong>and</strong> [v] <strong>and</strong> [�] <strong>and</strong> [v].<br />
The thought behind <strong>the</strong> way in which <strong>the</strong> diacritic nasal is processed is<br />
that a nasal sound is more or less intermediate between its non-nasal<br />
version <strong>and</strong> <strong>the</strong> [n]. We calculate <strong>the</strong> distance between a segment x <strong>and</strong> a<br />
nasal segment y as <strong>the</strong> average <strong>of</strong> <strong>the</strong> distance between x <strong>and</strong> y <strong>and</strong> <strong>the</strong><br />
distance between x <strong>and</strong> [n].<br />
4.2. Levenshtein distance<br />
Using <strong>the</strong> Levenshtein distance, two dialects are compared by comparing<br />
<strong>the</strong> pronunciation <strong>of</strong> a word in <strong>the</strong> first dialect with <strong>the</strong> pronunciation <strong>of</strong> <strong>the</strong><br />
same word in <strong>the</strong> second. It is determined how one pronunciation is<br />
changed into <strong>the</strong> o<strong>the</strong>r by inserting, deleting or substituting sounds.<br />
Weights are assigned to <strong>the</strong>se three operations. In <strong>the</strong> simplest form <strong>of</strong> <strong>the</strong><br />
algorithm, all operations have <strong>the</strong> same cost, e.g. 1. Assume afternoon is<br />
pronounced as [����t��n��n] in <strong>the</strong> dialect <strong>of</strong> Savannah, Georgia, <strong>and</strong> as<br />
[�����������] in <strong>the</strong> dialect <strong>of</strong> Lancaster, Pennsylvania 6 . Changing one<br />
pronunciation into <strong>the</strong> o<strong>the</strong>r can be done as in table 1 (ignoring<br />
suprasegmentals <strong>and</strong> diacritics for this moment) 7 :
76 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
Table 1. Changing one pronunciation into ano<strong>the</strong>r using a minimal set <strong>of</strong><br />
operations.<br />
æ�ft�n�n delete � 1<br />
æft�n�n insert r 1<br />
æft�rn�n subst. �/u 1<br />
æft�rnun<br />
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯<br />
3<br />
In fact many sequence operations map [�æ�ft��n��n] to [�æft�r�nu�n]. The<br />
power <strong>of</strong> <strong>the</strong> Levenshtein algorithm is that it always finds <strong>the</strong> cost <strong>of</strong> <strong>the</strong><br />
cheapest mapping.<br />
Comparing pronunciations in this way, <strong>the</strong> distance between longer<br />
pronunciations will generally be greater than <strong>the</strong> distance between shorter<br />
pronunciations. The longer <strong>the</strong> pronunciation, <strong>the</strong> greater <strong>the</strong> chance for<br />
differences with respect to <strong>the</strong> corresponding pronunciation in ano<strong>the</strong>r<br />
variety. Because this does not accord with <strong>the</strong> idea that words are linguistic<br />
units, <strong>the</strong> sum <strong>of</strong> <strong>the</strong> operations is divided by <strong>the</strong> length <strong>of</strong> <strong>the</strong> longest<br />
alignment which gives <strong>the</strong> minimum cost. The longest alignment has <strong>the</strong><br />
greatest number <strong>of</strong> matches. In our example we have <strong>the</strong> following<br />
alignment:<br />
Table 2. Alignment which gives <strong>the</strong> minimal cost. The alignment corresponds<br />
with table 1.<br />
æ � f t � n � n<br />
æ f t � r n u n<br />
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯<br />
1 1 1<br />
The total cost <strong>of</strong> 3 (1+1+1) is now divided by <strong>the</strong> length <strong>of</strong> 9. This gives a<br />
word distance <strong>of</strong> 0.33 or 33%.<br />
In Section 3.1.3 we explained how distances between segments can be<br />
found using spectrograms. This makes it possible to refine our Levenshtein<br />
algorithm by using <strong>the</strong> spectrogram distances as operation weights. Now<br />
<strong>the</strong> cost <strong>of</strong> insertions, deletions <strong>and</strong> substitutions is not always equal to 1,<br />
but varies, i.e., it is equal to <strong>the</strong> spectrogram distance between <strong>the</strong> segment
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 77<br />
<strong>and</strong> ‘silence’ (insertions <strong>and</strong> deletions) or between two segments<br />
(substitution).<br />
To reckon with syllabification in words, <strong>the</strong> Levenshtein algorithm is<br />
adapted so that only a vowel may match with a vowel, a consonant with a<br />
consonant, <strong>the</strong> [j] or [w] with a vowel (or opposite), <strong>the</strong> [i] or [u] with a<br />
consonant (or opposite), <strong>and</strong> a central vowel (in our research only <strong>the</strong><br />
schwa) with a sonorant (or opposite). In this way unlikely matches (e.g. a<br />
[p] with a [a]) are prevented.<br />
In our research we used 58 different words. When a word occurred in<br />
<strong>the</strong> text more than once, <strong>the</strong> mean over <strong>the</strong> different pronunciations was<br />
used. So when comparing two dialects we get 58 Levenshtein distances.<br />
Now <strong>the</strong> dialect distance is equal to <strong>the</strong> sum <strong>of</strong> 58 Levenshtein distances<br />
divided by 58. When <strong>the</strong> word distances are presented in terms <strong>of</strong><br />
percentages, <strong>the</strong> dialect distance will also be presented in terms <strong>of</strong><br />
percentages. All distances between <strong>the</strong> 15 language varieties are arranged<br />
in a 15 × 15 matrix.<br />
5. Results<br />
The results <strong>of</strong> <strong>the</strong> Levenshtein distance measurements are analyzed in two<br />
ways. First, on <strong>the</strong> basis <strong>of</strong> <strong>the</strong> distance matrix we applied hierarchical<br />
cluster analysis (see Section 5.1). The goal <strong>of</strong> clustering is to identify <strong>the</strong><br />
main groups. The groups are called clusters. Clusters may consist <strong>of</strong><br />
subclusters, <strong>and</strong> subclusters may in turn consist <strong>of</strong> subsubclusters, etc. The<br />
result is a hierarchically structured tree in which <strong>the</strong> dialects are <strong>the</strong> leaves<br />
(Jain <strong>and</strong> Dubes, 1988). Several alternatives exist. We used <strong>the</strong> Unweighted<br />
Pair Group Method using Arithmetic averages (UPGMA), since<br />
dendrograms generated by this method reflected distances which correlated<br />
most strongly with <strong>the</strong> original Levenshtein distances (r=0.9832), see Sokal<br />
<strong>and</strong> Rohlf (1962).<br />
Second, we ranked all varieties in or<strong>der</strong> <strong>of</strong> relationship with <strong>the</strong> st<strong>and</strong>ard<br />
languages, Frisian <strong>and</strong> Town Frisian (see Section 5.2). When ranking with<br />
relation to Frisian, we looked at <strong>the</strong> average over all Frisian dialects. Since<br />
<strong>the</strong> ratings with respect to each <strong>of</strong> <strong>the</strong> Frisian varieties individually were<br />
very similar averaging was justified.
78 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
5.1. The classification <strong>of</strong> <strong>the</strong> Germanic languages<br />
Looking at <strong>the</strong> clusters <strong>of</strong> language varieties in Figure 5 we note that our<br />
results reflect <strong>the</strong> traditional classification <strong>of</strong> <strong>the</strong> Germanic languages to a<br />
large extent (see Figure 1). <strong>On</strong> <strong>the</strong> highest level <strong>the</strong>re is a division between<br />
English <strong>and</strong> <strong>the</strong> o<strong>the</strong>r Germanic languages. When we examine <strong>the</strong> group <strong>of</strong><br />
o<strong>the</strong>r Germanic languages, we find a clear division between <strong>the</strong> North<br />
Germanic languages <strong>and</strong> <strong>the</strong> West Germanic languages. Within <strong>the</strong> North<br />
Germanic group, we see a clear division between <strong>the</strong> Sc<strong>and</strong>inavian<br />
languages (Danish, Norwegian <strong>and</strong> Swedish) on <strong>the</strong> one h<strong>and</strong> <strong>and</strong> <strong>the</strong><br />
Faroese <strong>and</strong> Icel<strong>and</strong>ic on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>. In <strong>the</strong> genetic tree (see Figure 1),<br />
Norwegian is clustered with Icel<strong>and</strong>ic <strong>and</strong> Faroese. However, due to <strong>the</strong><br />
isolated position <strong>of</strong> Icel<strong>and</strong> <strong>and</strong> <strong>the</strong> Faroes <strong>and</strong> intensive language contact<br />
between Norway <strong>and</strong> <strong>the</strong> rest <strong>of</strong> Sc<strong>and</strong>inavia, mo<strong>der</strong>n Norwegian has<br />
become very similar to <strong>the</strong> mo<strong>der</strong>n languages <strong>of</strong> Denmark <strong>and</strong> Sweden. All<br />
varieties spoken in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s, including <strong>the</strong> Frisian varieties, cluster<br />
toge<strong>the</strong>r, <strong>and</strong> German clusters more closely to <strong>the</strong>se varieties than English.<br />
Figure 5. Dendrogram showing <strong>the</strong> clustering <strong>of</strong> <strong>the</strong> 14 language varieties in our<br />
study. The scale distance shows average Levenshtein distances in<br />
percentages.<br />
All Frisian dialects form a cluster. This clustering corresponds well with<br />
<strong>the</strong> traditional classification as sketched in Section 3.1. The dialects <strong>of</strong><br />
Hijum <strong>and</strong> Oosterbierum belong to Klaaifrysk <strong>and</strong> <strong>the</strong>se dialects form a<br />
cluster. The Wâldfrysk dialects <strong>of</strong> Westergeest <strong>and</strong> Wetsens also cluster<br />
toge<strong>the</strong>r. The Levenshtein distance between <strong>the</strong> four dialects is small,
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 79<br />
ranging from 19.6% between Hijum <strong>and</strong> Oosterbierum <strong>and</strong> 23.8% between<br />
Oosterbierum <strong>and</strong> Westergeest. Also <strong>the</strong> Súdwesthoeksk dialects,<br />
represented by <strong>the</strong> Tjerkgaast dialect, are ra<strong>the</strong>r close to <strong>the</strong> Klaaifrysk <strong>and</strong><br />
Wâldfrysk dialects (distances between 21.6% <strong>and</strong> 26.4%). The highly<br />
conservative dialect <strong>of</strong> Hindeloopen is more deviant from <strong>the</strong> o<strong>the</strong>r dialects<br />
(distances between 29.8% <strong>and</strong> 32.5%) <strong>and</strong> this is also <strong>the</strong> case for <strong>the</strong> Town<br />
Frisian dialect <strong>of</strong> Leeuwarden which is more similar to Dutch (20.3%) than<br />
to Frisian (between 32.3% <strong>and</strong> 35.8%) which confirms <strong>the</strong> characterization<br />
<strong>of</strong> Town Frisian by Kloeke (1927) as ‘Dutch in Frisian mouth’.<br />
5.2. The relationship between Frisian <strong>and</strong> <strong>the</strong> o<strong>the</strong>r Germanic languages<br />
From Table 3 <strong>and</strong> 4 it is possible to determine <strong>the</strong> distance between all<br />
Germanic st<strong>and</strong>ard languages. We are especially interested in <strong>the</strong> position<br />
<strong>of</strong> Frisian within <strong>the</strong> Germanic language group. For this purpose <strong>the</strong> mean<br />
distance over <strong>the</strong> 6 Frisian dialects (excluding <strong>the</strong> dialect <strong>of</strong> Leeuwarden<br />
which is consi<strong>der</strong>ed Dutch) has been added. This makes it possible to treat<br />
Frisian as one language. Examining <strong>the</strong> column which shows <strong>the</strong> ranking<br />
with respect to Frisian, we find that Dutch is most similar to Frisian (a<br />
mean distance <strong>of</strong> 38.7%). Clearly <strong>the</strong> intensive contact with Dutch during<br />
history has had a great impact on <strong>the</strong> distance between <strong>the</strong> two languages.<br />
Moreover, German appears to be closer to Frisian than any o<strong>the</strong>r language<br />
outside <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s. Looking at <strong>the</strong> ranking with respect to Dutch, it<br />
appears that Town Frisian is most similar (Leeuwarden 20.3%), followed<br />
by <strong>the</strong> Frisian varieties (average <strong>of</strong> 38.7%). Next, German is most similar,<br />
due to common historical roots <strong>and</strong> continuous contact (a distance <strong>of</strong><br />
53.3%).<br />
As discussed in <strong>the</strong> introduction, Friesl<strong>and</strong> has a long history <strong>of</strong><br />
language contact with <strong>the</strong> Sc<strong>and</strong>inavian countries, <strong>and</strong> traces <strong>of</strong><br />
Sc<strong>and</strong>inavian influences can be found in <strong>the</strong> Frisian language. The impact<br />
<strong>of</strong> this contact is reflected in our results only to a limited extent.<br />
Remarkably, <strong>the</strong> distances to <strong>the</strong> mainl<strong>and</strong> Sc<strong>and</strong>inavian languages<br />
(Danish, Norwegian <strong>and</strong> Swedish) are smaller (between 60.7% <strong>and</strong> 63.3%)<br />
than to English (65.3%) even though <strong>the</strong> Frisian language is genetically<br />
closer related to English than to Sc<strong>and</strong>inavian (see Section 2.1).
80 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
Table 3. Ranked Levenshtein distances in percentages between each <strong>of</strong> <strong>the</strong> five<br />
West Germanic languages <strong>and</strong> <strong>the</strong> o<strong>the</strong>r language varieties in <strong>the</strong><br />
investigation.<br />
Frisian Leeuwarden Dutch English German<br />
Dutch 20.3 Leeuw 20.3 Hindel 63.1 Dutch 53.3<br />
Wetsens 32.3 Hindel 37.5 Wetsens 64.4 Leeuw 54.2<br />
Westerg 32.7 Westerg 37.7 Dutch 64.7 Hindel 56.2<br />
Frisian 34.2 Wetsens 38.3 Swedish 64.9 Westerg 56.9<br />
Oosterb 34.3 Tjerkg 38.5 Leeuw 65.1 Oosterb 57.2<br />
Hindel 34.9 Frisian 38.7 Tjerkg 65.2 Tjerkg 57.3<br />
Leeuw 34.2 Tjerkg 35.3 Hijum 38.9 Frisian 65.3 Frisian 57.3<br />
Dutch 38.7 Hijum 35.8 Oosterb 41.3 Hijum 65.8 Hijum 57.5<br />
German 57.3 German 54.2 German 53.3 Westerg 65.8 Wetsens 58.6<br />
Swedish 60.7 Swedish 59.2 Swedish 60.9 Danish 66.7 Swedish 61.0<br />
Norweg 60.9 Norweg 60.0 Norweg 61.4 Faroese 67.1 Danish 63.5<br />
Danish 63.3 Danish 61.1 Danish 63.4 Oosterb 67.2 Norweg 64.0<br />
English 65.3 English 65.1 English 64.7 German 68.1 Faroese 67.1<br />
Faroese 67.7 Faroese 67.5 Faroese 66.1 Norweg 68.6 English 68.1<br />
Icel<strong>and</strong>ic 70.0 Icel<strong>and</strong>ic 69.6 Icel<strong>and</strong>ic 69.2 Icel<strong>and</strong>ic 69.1 Icel<strong>and</strong>ic 68.5<br />
Table 4. Ranked Levenshtein distances in percentages between each <strong>of</strong> <strong>the</strong> five<br />
North Germanic languages <strong>and</strong> <strong>the</strong> o<strong>the</strong>r language varieties in <strong>the</strong><br />
investigation.<br />
Danish Swedish Norwegian Icel<strong>and</strong>ic Faroese<br />
Norweg 43.8 Norweg 43.4 Swedish 43.4 Faroese 54.1 Swedish 53.6<br />
Swedish 47.0 Danish 47.0 Danish 43.8 Swedish 58.7 Icel<strong>and</strong>ic 54.1<br />
Faroese 58.5 Faroese 53.6 Faroese 57.2 Norweg 62.6 Norweg 57.2<br />
Leeuw 61.1 Icel<strong>and</strong>ic 58.7 Westerg 59.6 Danish 62.7 Danish 58.5<br />
Westerg 62.2 Hindel 59.2 Leeuw 60.0 German 68.5 Dutch 66.1<br />
Wetsens 62.3 Leeuw 59.2 Hindel 60.2 Tjerkg 69.1 Hindel 67.0<br />
Icel<strong>and</strong>ic 62.7 Westerg 59.6 Tjerkg 60.6 English 69.1 English 67.1<br />
Hijum 62.9 Tjerkg 60.0 Wetsens 60.7 Dutch 69.2 German 67.1<br />
Frisian 63.3 Frisian 60.7 Frisian 60.9 Leeuw 69.6 Westerg 67.4<br />
Hindel 63.4 Dutch 60.9 Dutch 61.4 Hijum 69.8 Leeuw 67.5<br />
Dutch 63.4 German 61.0 Oosterb 61.9 Frisian 70.0 Tjerkg 67.5<br />
German 63.5 Wetsens 61.1 Hijum 62.6 Wetsens 70.1 Frisian 67.5<br />
Tjerkg 63.8 Oosterb 61.4 Icel<strong>and</strong>ic 62.6 Hindel 70.1 Oosterb 67.7<br />
Oosterb 65.2 Hijum 62.7 German 64.0 Oosterb 70.3 Wetsens 68.1<br />
English 66.7 Icel<strong>and</strong>ic 64.9 English 68.6 Westerg 70.3 Hijum 68.2
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 81<br />
So, when looking at <strong>the</strong> results from a Frisian perspective, <strong>the</strong> close genetic<br />
relationship with English is not reflected in our results. Of <strong>the</strong> Germanic<br />
languages in our investigation, only Icel<strong>and</strong>ic <strong>and</strong> Faroese are less similar<br />
to Frisian than English. However, when looking at <strong>the</strong> results from an<br />
English perspective, we discover that <strong>of</strong> all Germanic language varieties in<br />
our material <strong>the</strong> Frisian dialect <strong>of</strong> Hindeloopen is most similar to English.<br />
As mentioned before, this dialect is highly conservative <strong>and</strong> fur<strong>the</strong>rmore it<br />
is spoken in a coastal place, which provides for easy contact with Engl<strong>and</strong>.<br />
Also <strong>the</strong> Frisian dialect <strong>of</strong> Wetsens is more similar to English than <strong>the</strong><br />
remaining Germanic languages. The o<strong>the</strong>r Frisian varieties are found<br />
elsewhere in <strong>the</strong> middle <strong>of</strong> <strong>the</strong> ranking. Among <strong>the</strong> non-Frisian varieties,<br />
Dutch appears to be most similar to English. However, all Germanic<br />
languages, including Frisian <strong>and</strong> Dutch, show a large linguistic distance to<br />
English, all distances being above 60%. The development <strong>of</strong> <strong>the</strong> English<br />
language has thus clearly taken place independently from <strong>the</strong> o<strong>the</strong>r<br />
Germanic languages, which can be explained by <strong>the</strong> strong influence from<br />
non-Germanic languages, especially French.<br />
Also Icel<strong>and</strong>ic shows a large distance to all o<strong>the</strong>r Germanic languages<br />
(from 54.1% to 70.0%), but in <strong>the</strong> Icel<strong>and</strong>ic case this is explained by <strong>the</strong><br />
conservative nature <strong>of</strong> this language ra<strong>the</strong>r than by language contact<br />
phenomena. Faroese is somewhat less conservative, but still shows ra<strong>the</strong>r<br />
large distances to <strong>the</strong> o<strong>the</strong>r languages (between 53.6% <strong>and</strong> 67.7%). The<br />
distances between <strong>the</strong> o<strong>the</strong>r Nordic languages are smaller (between 43.4%<br />
<strong>and</strong> 47%), as was expected given that <strong>the</strong> three Sc<strong>and</strong>inavian languages are<br />
mutually intelligible.<br />
6. Conclusions <strong>and</strong> discussion<br />
Overall, <strong>the</strong> classification <strong>of</strong> <strong>the</strong> Germanic languages resulting from our<br />
distance measurements supports our predictions. This goes for <strong>the</strong><br />
classification <strong>of</strong> <strong>the</strong> Frisian dialects <strong>and</strong> also for <strong>the</strong> rest <strong>of</strong> <strong>the</strong> Germanic<br />
languages. We interpret this as a confirmation <strong>of</strong> <strong>the</strong> suitability <strong>of</strong> our<br />
material showing that it is possible to measure Levenshtein distances on <strong>the</strong><br />
basis <strong>of</strong> whole texts with assimilation phenomena typical <strong>of</strong> connected<br />
speech <strong>and</strong> with a ra<strong>the</strong>r limited number <strong>of</strong> words.<br />
The aim <strong>of</strong> <strong>the</strong> present investigation was to get an impression <strong>of</strong> <strong>the</strong><br />
position <strong>of</strong> <strong>the</strong> Frisian language in <strong>the</strong> Germanic language area on <strong>the</strong> basis<br />
<strong>of</strong> quantitative data. The fact that Frisian is genetically most closely related
82 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
to English yields <strong>the</strong> expectation that <strong>the</strong>se two languages may still be<br />
linguistically similar. However, <strong>the</strong> distance between English <strong>and</strong> <strong>the</strong><br />
Frisian dialects is large. We can thus conclude that <strong>the</strong> close genetic<br />
relationship between English <strong>and</strong> Frisian is not reflected in <strong>the</strong> linguistic<br />
distances between <strong>the</strong> mo<strong>der</strong>n languages. Geographical <strong>and</strong> historical<br />
circumstances have caused <strong>the</strong> two languages to drift apart linguistically.<br />
Frisian has been strongly influenced by Dutch whereas English has been<br />
influenced by o<strong>the</strong>r languages, especially French.<br />
It would have been interesting to include <strong>the</strong>se languages in our<br />
material. This would have given an impression <strong>of</strong> <strong>the</strong>ir impact on <strong>the</strong><br />
English language. At <strong>the</strong> same time it would also have given us <strong>the</strong><br />
opportunity to test <strong>the</strong> Levenshtein method on a larger language family than<br />
<strong>the</strong> Germanic family with its relatively closely related languages. It would<br />
also be interesting to include Old English in our material since this would<br />
give us an impression <strong>of</strong> how mo<strong>der</strong>n Frisian is related to <strong>the</strong> English<br />
language at a time when it had only recently separated from <strong>the</strong> common<br />
Anglo-Saxon roots to which also Old Frisian belonged.<br />
For many centuries Frisian has been un<strong>der</strong> <strong>the</strong> strong influence from<br />
Dutch <strong>and</strong> <strong>the</strong> Frisian <strong>and</strong> Dutch language areas share a long common<br />
history. It <strong>the</strong>refore does not come as a surprise that Dutch is <strong>the</strong> Germanic<br />
language most similar to <strong>the</strong> language varieties spoken in Friesl<strong>and</strong>.<br />
It may be surprising that <strong>the</strong> linguistic distances between Dutch <strong>and</strong> <strong>the</strong><br />
Frisian dialects are smaller than <strong>the</strong> distances between <strong>the</strong> Sc<strong>and</strong>inavian<br />
languages (a mean difference <strong>of</strong> 6%). Sc<strong>and</strong>inavian languages are known to<br />
be mutually intelligible. This means that when, for example, a Swede <strong>and</strong> a<br />
Dane meet, <strong>the</strong>y mostly communicate each in <strong>the</strong>ir own language. This<br />
kind <strong>of</strong> communication, which is known as semi-communication (Haugen,<br />
1966), is not typical in <strong>the</strong> communication between Dutch-speaking <strong>and</strong><br />
Frisian-speaking citizens in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s. The two languages are<br />
consi<strong>der</strong>ed so different that it is not possible for a Dutch-speaking person to<br />
un<strong>der</strong>st<strong>and</strong> Frisian <strong>and</strong> consequently <strong>the</strong> Frisian interlocutor will have to<br />
speak Dutch to a non-Frisian person. Our results raise <strong>the</strong> question whe<strong>the</strong>r<br />
semi-communication would also be possible in a Dutch-Frisian situation. If<br />
this is not <strong>the</strong> case, we may explain this by linguistic <strong>and</strong> non-linguistic<br />
differences between <strong>the</strong> Frisian-Dutch situation <strong>and</strong> <strong>the</strong> Sc<strong>and</strong>inavian<br />
situation. The Levenshtein distance processes lexical, phonetic <strong>and</strong><br />
morphological differences. All three types are present in our transcription,<br />
since word lists are <strong>der</strong>ived from running texts. Syntactic characteristics are<br />
completely excluded from <strong>the</strong> analysis. It might be <strong>the</strong> case that certain
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 83<br />
characteristics play a larger role for <strong>the</strong> Levenshtein distances than<br />
desirable in <strong>the</strong> case <strong>of</strong> <strong>the</strong> Sc<strong>and</strong>inavian languages if we were to use <strong>the</strong><br />
method for <strong>the</strong> explaining mutual intelligibility. For example, it is wellknown<br />
among <strong>the</strong> speakers <strong>of</strong> Sc<strong>and</strong>inavian languages that many words<br />
end in an ‘a’ in Swedish while ending in an ‘e’ in Danish. Probably people<br />
use this knowledge in an inter-Sc<strong>and</strong>inavian situation. However, this<br />
difference is included in <strong>the</strong> Levenshtein distances between Swedish <strong>and</strong><br />
Danish. It is possible that Frisian-Dutch differences are less predictable or<br />
less well-known by speakers <strong>of</strong> <strong>the</strong> two languages. It is also possible that<br />
<strong>the</strong> difference in communication in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s <strong>and</strong> in Sc<strong>and</strong>inavia<br />
should be sought at <strong>the</strong> extra-linguistic level. Sc<strong>and</strong>inavian research on<br />
semi-communication has shown that <strong>the</strong> willingness to un<strong>der</strong>st<strong>and</strong> <strong>and</strong> <strong>the</strong><br />
belief that it is possible to communicate play a large role for mutual<br />
intelligibility between speakers <strong>of</strong> closely related languages.<br />
Staying with <strong>the</strong> Sc<strong>and</strong>inavian languages, it should be noted that <strong>the</strong><br />
mainl<strong>and</strong> Sc<strong>and</strong>inavian languages are in fact closer to Frisian than English,<br />
even though <strong>the</strong> Sc<strong>and</strong>inavian languages belong genetically to ano<strong>the</strong>r<br />
Germanic branch than English <strong>and</strong> Frisian. This can probably be explained<br />
by intensive contacts between Frisians <strong>and</strong> Sc<strong>and</strong>inavians for many<br />
centuries. However, <strong>the</strong> common idea among some speakers <strong>of</strong> Frisian <strong>and</strong><br />
Sc<strong>and</strong>inavian that <strong>the</strong> two languages are so close that <strong>the</strong>y are almost<br />
mutually intelligible is not confirmed by our results, at least not as far as<br />
<strong>the</strong> st<strong>and</strong>ard Sc<strong>and</strong>inavian languages are concerned. Probably this popular<br />
idea is built on <strong>the</strong> fact that a few frequent words are identical in Frisian<br />
<strong>and</strong> Sc<strong>and</strong>inavian. It is possible, however, that this picture would change if<br />
we would include more Danish dialects in our material. For example, it<br />
seems to be relatively easy for fishermen from Friesl<strong>and</strong> to speak to <strong>the</strong>ir<br />
colleagues from <strong>the</strong> west coast <strong>of</strong> Denmark. Part <strong>of</strong> <strong>the</strong> explanation might<br />
also be that fishermen share a common vocabulary <strong>of</strong> pr<strong>of</strong>essional terms.<br />
Also <strong>the</strong> frequent contact <strong>and</strong> a strong motivation to communicate<br />
successfully are likely to be important factors.<br />
As we mentioned in <strong>the</strong> introduction, among dialects in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s<br />
<strong>and</strong> Fl<strong>and</strong>ers, <strong>the</strong> Frisian varieties are most deviant from St<strong>and</strong>ard Dutch.<br />
However, among <strong>the</strong> varieties which are recognized as languages in <strong>the</strong><br />
Germanic language area, Frisian is most similar to Dutch. The smallest<br />
distance between two languages, apart from Frisian, was found between<br />
Norwegian <strong>and</strong> Swedish: 43.4%. The distance between Frisian <strong>and</strong> Dutch is<br />
smaller: 38.7%. The Town Frisian variety <strong>of</strong> <strong>the</strong> capital <strong>of</strong> Friesl<strong>and</strong><br />
(Leeuwarden) has a distance <strong>of</strong> only 20.3% to Dutch. Although <strong>the</strong>
84 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
recognition <strong>of</strong> Frisian as second <strong>of</strong>ficial language in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s is right<br />
in our opinion, we found that <strong>the</strong> current linguistic position <strong>of</strong> Frisian<br />
provide too little foundation for becoming independent from <strong>the</strong><br />
Ne<strong>the</strong>rl<strong>and</strong>s, as some Frisians may wish 8 .<br />
Acknowledgements<br />
This research would have been impossible without informants who were<br />
willing to translate <strong>the</strong> story <strong>of</strong> ‘<strong>the</strong> Northwind <strong>and</strong> <strong>the</strong> Sun’. We wish to<br />
thank G. Blom (Hindeloopen), J. Spoelstra (Hijum) <strong>and</strong> W. Visser<br />
(Oosterbierum). All <strong>of</strong> <strong>the</strong>m are affiliated with <strong>the</strong> Fryske Akademy in<br />
Leeuwarden. We also thank S. van Dellen (Wetsens), T. de Graaf<br />
(Leeuwarden), F. Postma (Tjerkgaast) <strong>and</strong> O. Vries (Westergeest), all <strong>of</strong><br />
<strong>the</strong>m employees <strong>of</strong> <strong>the</strong> University <strong>of</strong> Groningen. We thank J. Allen<br />
(Engl<strong>and</strong>), A. Mikaelsdóttir (Icel<strong>and</strong>), Vigdis Petersen (<strong>the</strong> Faroes), R.<br />
Kraayenbrink (<strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s), K. Sjöberg (Sweden) <strong>and</strong> R. Schmidt<br />
(Germany). We are also very grateful to Jørn Almberg for making available<br />
<strong>the</strong> recording <strong>of</strong> Lillehammer (Norway). The recordings <strong>and</strong> transcriptions<br />
<strong>of</strong> <strong>the</strong> Frisian transcriptions are made by <strong>the</strong> second author, <strong>and</strong> those <strong>of</strong> <strong>the</strong><br />
st<strong>and</strong>ard languages (except Norway <strong>and</strong> <strong>the</strong> Faroes) by <strong>the</strong> first author. The<br />
transcriptions subsequently were checked by Jørn Almberg who we thank<br />
gratefully for correcting our transcriptions. Fur<strong>the</strong>rmore, we wish to<br />
express our gratitude to Peter Kleiweg for his s<strong>of</strong>tware for creating <strong>the</strong> map<br />
(Figure 2) <strong>and</strong> visualizing <strong>the</strong> dendrogram (Figure 5). Finally we thank<br />
Maartje Schreu<strong>der</strong> for reading an earlier version <strong>of</strong> this article <strong>and</strong> giving<br />
useful comments <strong>and</strong> Angeliek van Hout for reviewing our English.<br />
Notes<br />
1 Dr. Tjeerd de Graaf, <strong>the</strong> central figure in this volume, was born in Leeuwarden,<br />
<strong>the</strong> capital <strong>of</strong> Friesl<strong>and</strong>. Leeuwarden is one <strong>of</strong> <strong>the</strong> places where Town Frisian is<br />
spoken. Tjeerd de Graaf is a native speaker <strong>of</strong> this dialect, but later on he also<br />
learned (st<strong>and</strong>ard) Frisian. The Leeuwarden speaker in <strong>the</strong> present investigation<br />
was Tjeerd de Graaf (see Section 3.1).<br />
2 Most <strong>of</strong> this section is based on König <strong>and</strong> Van <strong>der</strong> Auwera (1994).<br />
3 The Lillehammer recording can be found at http://www.ling.hf.ntnu.no/nos/<br />
toge<strong>the</strong>r with 52 recordings <strong>of</strong> o<strong>the</strong>r Norwegian dialects.
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 85<br />
4 Since our material included two toneme languages, Swedish <strong>and</strong> Norwegian,<br />
also <strong>the</strong> two tonemes I <strong>and</strong> II were transcribed. For <strong>the</strong> o<strong>the</strong>r varieties primary<br />
stress was noted. Stress <strong>and</strong> tonemes were, however, not included for<br />
calculation <strong>of</strong> linguistic distances.<br />
5 The program PRAAT is a free public-domain program developed by Paul<br />
Boersma <strong>and</strong> David Weenink at <strong>the</strong> Institute <strong>of</strong> Phonetic Sciences <strong>of</strong> <strong>the</strong><br />
University <strong>of</strong> Amsterdam <strong>and</strong> available at http://www.fon.hum.uva.nl/praat.<br />
6 The data is taken from <strong>the</strong> Linguistic Atlas <strong>of</strong> <strong>the</strong> Middle <strong>and</strong> South Atlantic<br />
States (LAMSAS) <strong>and</strong> available via: http://hyde.park.uga.edu/lamsas/.<br />
7 The example should not be interpreted as a historical reconstruction <strong>of</strong> <strong>the</strong> way<br />
in which one pronunciation changed into ano<strong>the</strong>r. From that point <strong>of</strong> view it<br />
may be more obvious to show how [�����������] changed into [����t��n��n]. We<br />
just show that <strong>the</strong> distance between two arbitrary pronunciations is found on <strong>the</strong><br />
basis <strong>of</strong> <strong>the</strong> least costly set <strong>of</strong> operations mapping one pronunciation into<br />
ano<strong>the</strong>r.<br />
8 Tjeerd de Graaf has never taken such an extreme position. Possibly speakers <strong>of</strong><br />
Town Frisian have a more mo<strong>der</strong>ate opinion towards this issue since Town<br />
Frisian is more closely related to st<strong>and</strong>ard Dutch, as appeared in Figure 5 <strong>and</strong><br />
Table 3.<br />
References<br />
Bolognesi, R. <strong>and</strong> W. Heeringa (2002). De invloed van dominante talen op het<br />
lexicon en de fonologie van Sardische dialecten. In: D. Bakker,<br />
T. S<strong>and</strong>ers, R. Schoonen <strong>and</strong> Per van <strong>der</strong> Wijst (eds.).<br />
Gramma/TTT: tijdschrift voor taalwetenschap. Nijmegen<br />
University Press, Nijmegen, 9 (1): 45-84.<br />
Feitsma, T. (1963). Sproglige berøringer mellem Frisl<strong>and</strong> og Sk<strong>and</strong>inavien.<br />
Sprog og kultur, 23: 97-121.<br />
Gooskens, Ch. <strong>and</strong> W. Heeringa (submitted). Perceptive Evaluation <strong>of</strong><br />
Levenshtein Dialect Distance Measurements Using Norwegian<br />
Dialect Data. (submitted to Language Variation <strong>and</strong> Change).<br />
Goossens, J. (1977). Inleiding tot de Ne<strong>der</strong>l<strong>and</strong>se Dialectologie. Wolters-<br />
Noordh<strong>of</strong>f, Groningen.<br />
Haugen, E. (1966). Semicommunication: The Language Gap in Sc<strong>and</strong>inavia.<br />
Sociological Inquiry, 36 (2): 280-297.<br />
Heeringa, W. (2004). Measuring Dialect Pronunciation Differences using<br />
Levenshtein Distance. Doctoral dissertation. University <strong>of</strong><br />
Groningen.
86 Charlotte Gooskens <strong>and</strong> Wilbert Heeringa<br />
Heeringa, W. <strong>and</strong> C. Gooskens (2003). Norwegian Dialects Examined<br />
Perceptually <strong>and</strong> Acoustically. In: J. Nerbonne <strong>and</strong> W.<br />
Kretzschmar (eds.). Computers <strong>and</strong> <strong>the</strong> Humanities. Kluwer<br />
Academic Publishers, Dordrecht, 37 (3): 293-315.<br />
H<strong>of</strong>, J. J. (1933). Friesche Dialectgeographie. ‘s Gravenhage (Noord- en Zuid-<br />
Ne<strong>der</strong>l<strong>and</strong>se Dialectbiblio<strong>the</strong>ek 3).<br />
Hoppenbrouwers, C <strong>and</strong> G. Hoppenbrouwers (1988). De featurefrequentie<br />
methode en de classificatie van Ne<strong>der</strong>l<strong>and</strong>se dialecten. TABU:<br />
Bulletin voor Taalwetenschap, 18 (2): 51-92.<br />
Jain, A.K. <strong>and</strong> R.C. Dubes (1988). Algorithms for Clustering Data. Prentice<br />
Hall, Englewood Cliffs, New Yersey.<br />
Johnson, K. (1997). Acoustic <strong>and</strong> Auditory <strong>Phonetics</strong>. Blackwell Publishers,<br />
Cambridge etc..<br />
Kessler, B. (1995). Computational dialectology in Irish Gaelic. In: Proceedings<br />
<strong>of</strong> <strong>the</strong> 7 th Conference <strong>of</strong> <strong>the</strong> European Chapter <strong>of</strong> <strong>the</strong> Association<br />
for Computational Linguistics. EACL, Dublin, 60-67.<br />
Kloeke, G. G. (1927). De Holl<strong>and</strong>sche expansie in de zestiende en zeventiende<br />
eeuw en haar weerspiegeling in de hedendaagsche<br />
Ne<strong>der</strong>l<strong>and</strong>sche dialecten. Nijh<strong>of</strong>f, ‘s-Gravenhage.<br />
König, E. <strong>and</strong> J. van <strong>der</strong> Auwera (1994). eds. The Germanic Languages.<br />
Routledge, London.<br />
Miedema, H.T.J. (1966). Van York naar Jorwerd. Enkele problemen uit de<br />
Friese taalgeschiedenis. J.B. Wolters, Groningen.<br />
Nerbonne, J., W. Heeringa, E. van den Hout, P. van <strong>der</strong> Kooi, S. Otten, <strong>and</strong> W.<br />
van de Vis, (1996). Phonetic Distance between Dutch dialects.<br />
In: G. Durieux, W. Daelemans, <strong>and</strong> S. Gillis (eds.). CLIN VI,<br />
Papers from <strong>the</strong> sixth CLIN meeting. Antwerpen. University <strong>of</strong><br />
Antwerp, Center for Dutch Language <strong>and</strong> Speech, 185-202.<br />
Nerbonne, J. <strong>and</strong> W. Heeringa (1997). Measuring dialect distances<br />
phonetically. In: J. Coleman (ed.). Workshop on Computational<br />
<strong>Phonology</strong>. Madrid, 11-18.<br />
Pei, M. (1966). The story <strong>of</strong> language. Allen & Unwin, London.<br />
Potter, R.K., G.A. Kopp <strong>and</strong> H.C. Green (1947). Visible Speech. The Bell<br />
Telephone Laboratories Series. Van Nostr<strong>and</strong>, New York.<br />
Rietveld, A.C.M. <strong>and</strong> V.J. Van Heuven (1997). Algemene fonetiek. Coutinho,<br />
Bussum.<br />
S<strong>and</strong>øy, H. (1994). Utan kontakt og endring? In: U.-B. Kotsinas <strong>and</strong> J.<br />
Helg<strong>and</strong>er (eds.). Dialektkontakt, språkkontakt och<br />
språkförändring i Norden. Almqvist & Wiksell International,<br />
Stockholm, 38-51.<br />
Sokal, R.R. <strong>and</strong> F.J. Rohlf (1962). The comparison <strong>of</strong> dendrograms by<br />
objective methods. Taxon, 11: 33-40.
The Position <strong>of</strong> Frisian in <strong>the</strong> Germanic Language Area 87<br />
Stevens, K.N. (1998). Acoustic <strong>Phonetics</strong>. MIT Press, Cambridge.<br />
The Columbia Encyclopedia (2001). www.bartleby.com/65/fr/Frisianl.html<br />
The International Phonetic Association (1949). The principles <strong>of</strong> <strong>the</strong><br />
International Phonetic Association: being a description <strong>of</strong> <strong>the</strong><br />
International Phonetic Alphabet <strong>and</strong> <strong>the</strong> manner <strong>of</strong> using it,<br />
illustrated by texts in 51 languages. International Phonetic<br />
Association, London.<br />
The International Phonetic Association (1999). H<strong>and</strong>book <strong>of</strong> <strong>the</strong> International<br />
Phonetic Association: a guide to <strong>the</strong> use <strong>of</strong> <strong>the</strong> International<br />
Phonetic Alphabet. Cambridge University Press, Cambridge.<br />
Trask, R.L. (1996). A Dictionary <strong>of</strong> <strong>Phonetics</strong> <strong>and</strong> <strong>Phonology</strong>. Routledge,<br />
London <strong>and</strong> New York.<br />
Van <strong>der</strong> Veen, K. F. (2001). West Frisian Dialectology <strong>and</strong> Dialects. In: H. H.<br />
Munske (ed.). H<strong>and</strong>book <strong>of</strong> Frisian Studies. Niemeyer,<br />
Tübingen, 83-98.<br />
Vieregge, W. H., A.C.M. Rietveld <strong>and</strong> C. Jansen (1984). A distinctive feature<br />
based system for <strong>the</strong> evaluation <strong>of</strong> segmental transcription in<br />
Dutch. In: M.P.R. van den Broecke <strong>and</strong> A. Cohen. Proceedings<br />
<strong>of</strong> <strong>the</strong> 10 th International Congress <strong>of</strong> Phonetic Sciences. Foris<br />
Publications, Dordrecht <strong>and</strong> Cinnaminson, 654-659.<br />
Visser, W. (1997). The syllable in Frisian. Holl<strong>and</strong> Academic Graphics, The<br />
Hague.<br />
Wadstein, E. (1933). <strong>On</strong> <strong>the</strong> Relations between Sc<strong>and</strong>inavians <strong>and</strong> Frisians in<br />
Early Times. University <strong>of</strong> London, London.<br />
Wells, J. <strong>and</strong> J. House (1995). The sounds <strong>of</strong> <strong>the</strong> International Phonetic<br />
Alphabet. UCL, London.<br />
Zwicker, E. <strong>and</strong> H. Fastl (1990). Psychoacoustics <strong>and</strong> Models. Springer Verlag,<br />
Berlin.<br />
Zwicker, E. <strong>and</strong> R. Feldtkeller (1967). Das Ohr als Nachrichtemfänger.<br />
Monographien <strong>der</strong> elektrischen Nachrichtentechnik. 19, 2 nd<br />
revised edition. Hirzel, Stuttgart.
Learning Phonotactics with Simple Processors<br />
John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
Abstract<br />
This paper explores <strong>the</strong> learning <strong>of</strong> phonotactics in neural networks.<br />
Experiments are conducted on <strong>the</strong> complete set <strong>of</strong> over 5,000 Dutch<br />
monosyllables extracted from CELEX, <strong>and</strong> <strong>the</strong> results are shown to be<br />
accurate within 5% error. Extensive comparisons to human phonotactic<br />
learning conclude <strong>the</strong> paper. We focus on whe<strong>the</strong>r phonotactics can be<br />
effectively learned <strong>and</strong> how <strong>the</strong> learning which is induced compares to<br />
human behavior.<br />
1. Introduction 1<br />
Phonotactics concerns <strong>the</strong> organization <strong>of</strong> <strong>the</strong> phonemes in words <strong>and</strong><br />
syllables. The phonotactic rules constrain how phonemes combine in or<strong>der</strong><br />
to form larger linguistic units (syllables <strong>and</strong> words) in that language (Laver,<br />
1994). For example, Cohen, Ebeling & van Holk (1972) describe <strong>the</strong><br />
phoneme combinations possible in Dutch, which will be <strong>the</strong> language in<br />
focus in this study.<br />
Phonotactic rules are implicit in natural languages so that humans<br />
require no explicit instruction about which combinations are allowed <strong>and</strong><br />
which are not. An explicit phonotactic grammar can <strong>of</strong> course be abstracted<br />
from <strong>the</strong> words in a language, but this is an activity linguists engage in, not<br />
language learners in general. Children normally learn a language's<br />
phonotactics in <strong>the</strong>ir early language development <strong>and</strong> probably update it<br />
only slightly once <strong>the</strong>y have mastered <strong>the</strong> language.<br />
Most work on language acquisition has arisen in linguistics <strong>and</strong><br />
psychology, <strong>and</strong> that work employs mechanisms that have been developed<br />
for language, typically, discrete, symbol-manipulation systems.<br />
Phonotactics in particular has been modeled with n-gram models, Finite
90 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
State Machines, Inductive Logic Programming, etc. (Tjong Kim Sang,<br />
1998; Konstantopoulos, 2003). Such approaches are effective, but a<br />
cognitive scientist may ask whe<strong>the</strong>r <strong>the</strong> same success could be possible<br />
using less custom-made tools. The brain, viewed as a computational<br />
machine, exploits o<strong>the</strong>r principles, which have been modeled in <strong>the</strong><br />
approach known as Parallel Distributed Processing (PDP), which was<br />
thoroughly described in <strong>the</strong> seminal work <strong>of</strong> Rumelhart & McClell<strong>and</strong><br />
(1986). Computational models inspired by <strong>the</strong> brain structure <strong>and</strong> neural<br />
processing principles are Neural Networks (NNs), also known as<br />
connectionist models.<br />
Learning phonotactic grammars is not an easy problem, especially when<br />
one restricts one's attention to cognitively plausible models. Since<br />
languages are experienced <strong>and</strong> produced dynamically, we need to focus on<br />
<strong>the</strong> processing <strong>of</strong> sequences, which complicates <strong>the</strong> learning task. The<br />
history <strong>of</strong> research in connectionist language learning shows both successes<br />
<strong>and</strong> failures even when one concentrates on simpler structures, such as<br />
phonotactics (Stoianov, Nerbonne & Bouma, 1998; Stoianov & Nerbonne,<br />
2000; Tjong Kim Sang, 1995; Tjong Kim Sang & Nerbonne, 1999; Pacton,<br />
Perruchet, Fayol & Cleeremans, 2001).<br />
This paper will attack phonotactics learning with models that have no<br />
specifically linguistic knowledge encoded a priori. The models naturally do<br />
have a bias, viz., toward extracting local conditioning factors for<br />
phonotactics, but we maintain that this is a natural bias for many sorts <strong>of</strong><br />
sequential behavior, not only linguistic processing. A first-or<strong>der</strong> Discrete<br />
Time Recurrent Neural Network (DTRNN) (Carrasco, Forcada & Neco,<br />
1999; Tsoi & Back, 1997) will be used: <strong>the</strong> Simple Recurrent Network<br />
(SRNs) (Elman, 1988). SRNs have been applied to different language<br />
problems (Elman, 1991; Christiansen & Chater, 1999; Lawrence, Giles &<br />
Fong, 1995), including learning phonotactics (Shillcock, Levy, Lindsey,<br />
Cairns & Chater, 1993; Shillcock, Cairns, Chater & Levy, 1997). With<br />
respect to phonotactics, we have also contributed reports (Stoianov et al.,<br />
1998; Stoianov & Nerbonne, 2000; Stoianov, 1998).<br />
SRNs have been shown capable <strong>of</strong> representing regular languages<br />
(Omlin & Giles, 1996; Carrasco et al., 1999). Kaplan & Kay (1994)<br />
demonstrated that <strong>the</strong> apparently context-sensitive rules that are st<strong>and</strong>ardly<br />
found in phonological descriptions can in fact be expressed within <strong>the</strong> more<br />
restrictive formalism <strong>of</strong> regular relations. We begin thus with a device<br />
which is in principle capable <strong>of</strong> representing <strong>the</strong> needed patterns.
Learning Phonotactics with Simple Processors 91<br />
We <strong>the</strong>n simulate <strong>the</strong> language learning task by training networks to<br />
produce context-dependent predictions. We also show how <strong>the</strong> continuous<br />
predictions <strong>of</strong> trained SRNs - likelihoods that a particular token can follow<br />
<strong>the</strong> current context - can be transformed into more useful discrete<br />
predictions, or, alternatively, string recognitions.<br />
In spite <strong>of</strong> <strong>the</strong> above claims about representability, <strong>the</strong> Back-<br />
Propagation (BP) <strong>and</strong> Back-Propagation Through Time (BPTT) learning<br />
algorithms used to train SRNs do not always find optimal solutions - SRNs<br />
that produce only correct context-dependent successors or recognize only<br />
strings from <strong>the</strong> training language. Hence, section 3 focuses on <strong>the</strong> practical<br />
demonstration that a realistic language learning task may be simulated by<br />
an SRN. We evaluate <strong>the</strong> network learning from different perspectives -<br />
grammar learning, phonotactics learning, <strong>and</strong> language recognition. The<br />
last two methods need one language-specific parameter - a threshold - that<br />
distinguishes successors/words allowed in <strong>the</strong> training language. This<br />
threshold is found with a post-training procedure, but it could also be<br />
sought interactively during training.<br />
Finally, section 4 assesses <strong>the</strong> networks from linguistic <strong>and</strong><br />
psycholinguistic perspectives: a static analysis extracts acquired linguistic<br />
knowledge from network weights, <strong>and</strong> <strong>the</strong> network performance is<br />
compared to humans' in a lexical decision task. The network performance,<br />
in particular <strong>the</strong> distribution <strong>of</strong> errors as a function <strong>of</strong> string position, will<br />
be compared to alternative construals <strong>of</strong> Dutch syllabic structure -<br />
following a suggestion from discussions <strong>of</strong> psycholinguistic experiments<br />
about English syllables (Kessler & Treiman, 1997).<br />
1.1. Motivations for a Phonotactic Device<br />
This section will review st<strong>and</strong>ard arguments that demonstrate <strong>the</strong> cognitive<br />
<strong>and</strong> practical importance <strong>of</strong> phonotactics. English phonotactic rules such as:<br />
‘/s/ may precede, but not follow /t/ syllable-initially’<br />
(ignoring loanwords such as `tsar' <strong>and</strong> `tse-tse') may be adduced by judging<br />
<strong>the</strong> well-formedness <strong>of</strong> sequences <strong>of</strong> letters/phonemes, taken as words in<br />
<strong>the</strong> language, e.g. /st�p/ vs. */ts�p/. There may also be cases judged to be <strong>of</strong><br />
intermediate acceptability. So, even if all <strong>of</strong> <strong>the</strong> following are English<br />
words:
92 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
/m����/ `mo<strong>the</strong>r', /f����/ `fa<strong>the</strong>r', /s�st��/ `sister'<br />
None <strong>of</strong> <strong>the</strong> following are, however:<br />
*/m���/, */f����/, */tss���/<br />
None <strong>of</strong> <strong>the</strong>se sound like English words. However, <strong>the</strong> following<br />
sequences:<br />
/m����/, /fu���/, /s�nt��/<br />
"sound" much more like English, even if <strong>the</strong>y mean nothing <strong>and</strong> <strong>the</strong>refore<br />
are not genuine English words. We suspect that, e.g., /s�nt��/ 'santer', could<br />
be used to name a new object or a concept.<br />
This simple example shows that we have a feeling for word structure,<br />
even if no explicit knowledge. Given <strong>the</strong> huge variety <strong>of</strong> words, it is more<br />
efficient to put this knowledge into a compact form - a set <strong>of</strong> phonotactic<br />
rules. These rules would state which phonemic sequences sound correct <strong>and</strong><br />
which do not. In this same vein, second language learners experience a<br />
period when <strong>the</strong>y recognize that certain phonemic combinations (words)<br />
belong to <strong>the</strong> language <strong>the</strong>y learn without knowing <strong>the</strong> meaning <strong>of</strong> <strong>the</strong>se<br />
words.<br />
Convincing psycholinguistic evidence that we make use <strong>of</strong> phonotactics<br />
comes from studying <strong>the</strong> information sources used in word segmentation<br />
(McQueen, 1998). In a variety <strong>of</strong> experiments, this author shows that word<br />
boundary locations are likely to be signaled by phonotactics. The author<br />
rules out <strong>the</strong> possibility that o<strong>the</strong>r sources <strong>of</strong> information, such as prosodic<br />
cues, syllabic structure <strong>and</strong> lexemes, are sufficient for segmentation.<br />
Similarly, Treiman & Zukowski (1990) had shown earlier that phonotactics<br />
play an important role in <strong>the</strong> syllabification process. According to<br />
McQueen (1998), phonotactic <strong>and</strong> metrical cues play complementary roles<br />
in <strong>the</strong> segmentation process. In accordance with this, some researchers have<br />
elaborated on a model for word segmentation: <strong>the</strong> Possible Word<br />
Constraints Model (Norris, McQueen, Cutler & Butterfield, 1997), in which<br />
likely word-boundary locations are marked by phonotactics, metrical cues,<br />
etc., <strong>and</strong> in which <strong>the</strong>y are fur<strong>the</strong>r fixed by using lexicon-specific<br />
knowledge.
Learning Phonotactics with Simple Processors 93<br />
Exploiting <strong>the</strong> specific phonotactics <strong>of</strong> Japanese, Dupoux, Pallier,<br />
Kakehi & Mehler (2001) conducted an experiment with Japanese listeners<br />
who heard stimuli that contained illegal consonant clusters. The listeners<br />
tended to hear an acoustically absent vowel that brought <strong>the</strong>ir perception<br />
into line with Japanese phonotactics. The authors were able to rule out<br />
lexical influences as a putative source for <strong>the</strong> perception <strong>of</strong> <strong>the</strong> illusory<br />
vowel, which suggests that speech perception must use phonotactic<br />
information directly.<br />
Fur<strong>the</strong>r justification for <strong>the</strong> postulation <strong>of</strong> a neurobiological device that<br />
encodes phonotactics comes from neurolinguistic <strong>and</strong> neuroimaging<br />
studies. It is widely accepted that <strong>the</strong> neuronal structure <strong>of</strong> Broca’s area (in<br />
<strong>the</strong> brain's left frontal lobe) is used for language processing, <strong>and</strong> more<br />
specially that it represents a general sequential device (Stowe, Wijers,<br />
Willemsen, Reul<strong>and</strong>, Paans & Vaalburg, 1994; Reilly, 2002). A general<br />
sequential processor capable <strong>of</strong> working at <strong>the</strong> phonemic level would be a<br />
plausible realization <strong>of</strong> a neuronal phonotactic device.<br />
Besides cognitive modeling, <strong>the</strong>re are also a number <strong>of</strong> practical<br />
problems that would benefit from effective phonotactic processing. In<br />
speech recognition, for example, a number <strong>of</strong> hypo<strong>the</strong>ses that explain <strong>the</strong><br />
speech signal are created, from which <strong>the</strong> impossible sound combinations<br />
have to be filtered out before fur<strong>the</strong>r processing. This exemplifies a lexical<br />
decision task, in which a model is trained on a language L <strong>and</strong> <strong>the</strong>n tests<br />
whe<strong>the</strong>r a given string belongs to L. In such a task a phonotactic device<br />
would be <strong>of</strong> use. Ano<strong>the</strong>r important problem in speech recognition is word<br />
segmentation. Speech is continuous, but we divide it into psychologically<br />
significant units such as words <strong>and</strong> syllables. As noted above, <strong>the</strong>re are a<br />
number <strong>of</strong> cues that we can use to distinguish <strong>the</strong>se elements - prosodic<br />
markers, context, but also phonotactics. Similarly to <strong>the</strong> former problem, an<br />
intuitive strategy here is to split <strong>the</strong> phonetic/phonemic stream at <strong>the</strong> points<br />
<strong>of</strong> violation <strong>of</strong> phonotactic constraints (see Shillcock et al. (1997) <strong>and</strong><br />
Cairns, Shillcock, Chater & Levy (1997) for connectionist modeling).<br />
Similarly, <strong>the</strong> constraints <strong>of</strong> <strong>the</strong> letters forming words in written languages<br />
(graphotactics) are useful in word processing applications, for example,<br />
spell-checking.<br />
There is ano<strong>the</strong>r, more speculative aspect to investigating phonotactics.<br />
Searching for an explanation <strong>of</strong> <strong>the</strong> structure <strong>of</strong> <strong>the</strong> natural languages,<br />
Carstairs-McCarthy presented in his recent book (1999) an analogy<br />
between syllable structure <strong>and</strong> sentence structure. He argues that sentences<br />
<strong>and</strong> syllables have a similar type <strong>of</strong> structure. Therefore, if we find a proper
94 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
mechanism for learning <strong>the</strong> syllabic structures, we might apply a similar<br />
mechanism to learning syntax as well. Of course, syntax is much more<br />
complex <strong>and</strong> more challenging, but if Carstairs-McCarthy is right, <strong>the</strong> basic<br />
principles <strong>of</strong> both devices might be <strong>the</strong> same.<br />
2. Simple Recurrent Networks<br />
This section will briefly present Simple Recurrent Networks (Elman, 1988;<br />
Robinson & Fallside, 1988) <strong>and</strong> will review earlier studies <strong>of</strong> sequential,<br />
especially phonotactic learning. Detailed descriptions <strong>of</strong> <strong>the</strong> SRN<br />
processing mechanisms <strong>and</strong> <strong>the</strong> Back-Propagation Through Time learning<br />
algorithm that is used to train <strong>the</strong> model are available elsewhere (Stoianov,<br />
2001; Haykin, 1994), <strong>and</strong> will be reviewed only superficially.<br />
Figure 1. Learning phonotactics with <strong>the</strong> SRNs. If <strong>the</strong> training data set contains <strong>the</strong><br />
words /n�t#/, /n�ts#/ <strong>and</strong> /n�t��rk#/ <strong>the</strong>n after <strong>the</strong> network has processed<br />
a left context /n�/, <strong>the</strong> reaction to an input token /t/ will be active neurons<br />
corresponding to <strong>the</strong> symbol '#' <strong>and</strong> <strong>the</strong> phonemes /s/, <strong>and</strong> /�/.<br />
Simple Recurrent Networks (SRNs) were invented to encode simple<br />
artificial grammars, as an extension <strong>of</strong> <strong>the</strong> Multilayer Perceptron<br />
(Rumelhart, Hinton & Williams, 1986) with an extra input - a context layer<br />
that holds <strong>the</strong> hidden layer activations at <strong>the</strong> previous processing cycle.<br />
After training, Elman (1988) conducted investigations on how context
Learning Phonotactics with Simple Processors 95<br />
evolves in time. The analysis showed graded encoding <strong>of</strong> <strong>the</strong> input<br />
sequence: similar items presented to <strong>the</strong> input were clustered at close, but<br />
different, shifting positions. That is, <strong>the</strong> network discovered <strong>and</strong> implicitly<br />
represented in a distributed way <strong>the</strong> rules <strong>of</strong> <strong>the</strong> grammar generating <strong>the</strong><br />
training sequences. This is noteworthy, because <strong>the</strong> rules for context were<br />
not encoded, but ra<strong>the</strong>r acquired through experience. The capacity <strong>of</strong> SRNs<br />
to learn simple artificial languages was fur<strong>the</strong>r explored in a number <strong>of</strong><br />
studies (Cleeremans, Servan-Schreiber & McClell<strong>and</strong>, 1989; Gasser, 1992).<br />
SRNs have <strong>the</strong> structure shown in Figure 1. They operate as follows:<br />
Input sequences S I are presented to <strong>the</strong> input layer, one element S I (t) at a<br />
time. The purpose <strong>of</strong> <strong>the</strong> input layer is just to transfer activation to <strong>the</strong><br />
hidden layer through a weight matrix. The hidden layer in turn copies its<br />
activations after every step to <strong>the</strong> context layer, which provides an<br />
additional input to <strong>the</strong> hidden layer - i.e., information about <strong>the</strong> past, after a<br />
brief delay. Finally, <strong>the</strong> hidden layer neurons output <strong>the</strong>ir signal through a<br />
second weight matrix to <strong>the</strong> output layer neurons. The activation <strong>of</strong> <strong>the</strong><br />
latter is interpreted as <strong>the</strong> product <strong>of</strong> <strong>the</strong> network. Since <strong>the</strong> activation <strong>of</strong><br />
<strong>the</strong> hidden layer depends both on its previous state (<strong>the</strong> context) <strong>and</strong> on <strong>the</strong><br />
current input, SRNs have <strong>the</strong> <strong>the</strong>oretical capacity to be sensitive to <strong>the</strong><br />
entire history <strong>of</strong> <strong>the</strong> input sequence. However, practical limitations restrict<br />
<strong>the</strong> time span <strong>of</strong> <strong>the</strong> context information to maximally 10-15 steps<br />
(Christiansen & Chater, 1999). The size <strong>of</strong> <strong>the</strong> layers does not restrict <strong>the</strong><br />
range <strong>of</strong> temporal sensitivity.<br />
The network operates in two working regimens - supervised training <strong>and</strong><br />
network use. In <strong>the</strong> latter, <strong>the</strong> network is presented <strong>the</strong> sequential input data<br />
S I (t) <strong>and</strong> computes <strong>the</strong> output N(t) using contextual information. The<br />
training regimen involves <strong>the</strong> same sort <strong>of</strong> processing as network use <strong>and</strong><br />
also includes a second, training step, which compares <strong>the</strong> network reaction<br />
N(t) to <strong>the</strong> desired one S T (t), <strong>and</strong> which uses <strong>the</strong> difference to adjust <strong>the</strong><br />
network behavior in a way that improves future network performance on<br />
<strong>the</strong> same data.<br />
The two most popular supervised learning algorithms used to train<br />
SRNs are <strong>the</strong> st<strong>and</strong>ard Back-Propagation algorithm (Rumelhart et al., 1986)<br />
<strong>and</strong> <strong>the</strong> Back-Propagation Through Time algorithm (Haykin, 1994). While<br />
<strong>the</strong> earlier is simpler because it uses information from one previous time<br />
step only (<strong>the</strong> context activation, <strong>the</strong> current network activations, <strong>and</strong><br />
error), <strong>the</strong> latter trains <strong>the</strong> network faster, because it collects errors from all<br />
time steps during which <strong>the</strong> network processes <strong>the</strong> current sequence <strong>and</strong><br />
<strong>the</strong>refore it adjusts <strong>the</strong> weights more precisely. However, <strong>the</strong> BPTT
96 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
learning algorithm is also cognitively less plausible, since <strong>the</strong> collection <strong>of</strong><br />
<strong>the</strong> time-spanning information requires mechanisms specific for <strong>the</strong><br />
symbolic methods. Never<strong>the</strong>less, this compromise allows more extensive<br />
research, <strong>and</strong> without it <strong>the</strong> problems discussed below would require much<br />
longer training time when using st<strong>and</strong>ard computers for simulations.<br />
Therefore, in <strong>the</strong> experiments reported here <strong>the</strong> BPTT learning algorithm<br />
will be used. In brief, it works in <strong>the</strong> following way: <strong>the</strong> network reaction to<br />
a given input sequence is compared to <strong>the</strong> desired target sequence at every<br />
time step <strong>and</strong> an error is computed. The network activation <strong>and</strong> error at<br />
each step are kept in a stack. When <strong>the</strong> whole sequence is processed, <strong>the</strong><br />
error is propagated back through space (<strong>the</strong> layers) <strong>and</strong> time, <strong>and</strong> weightupdating<br />
values are computed. Then, <strong>the</strong> network weights are adjusted with<br />
<strong>the</strong> values computed in this way.<br />
2.1. Learning Phonotactics with SRNs<br />
Dell, Juliano & Govindjee (1993) showed that words could be described<br />
not only with symbolic approaches, using word structure <strong>and</strong> content, but<br />
also by a connectionist approach. In this early study <strong>of</strong> learning word<br />
structure with neural nets (NNs), <strong>the</strong> authors trained SRNs to predict <strong>the</strong><br />
phoneme that follows <strong>the</strong> current input phoneme, given context<br />
information. The data sets contained 100 - 500 English words. An<br />
important issue in <strong>the</strong>ir paper is <strong>the</strong> analysis <strong>and</strong> modeling <strong>of</strong> a number <strong>of</strong><br />
speech-error phenomena, which were taken as strong support for parallel<br />
distributed processing (PDP) models, in particular SRNs. Some <strong>of</strong> <strong>the</strong>se<br />
phenomena were: phonological movement errors (reading list - leading<br />
list), manner errors (department - jepartment), phonotactic regularity<br />
violations (dorm - dlorm), consonant-vowel category confusions <strong>and</strong> initial<br />
consonant omissions (cluster-initial consonants dropping as when `stop' is<br />
mispronounced [t�p]).<br />
Aiming at segmentation <strong>of</strong> continuous phonetic input, Shillcock et al.<br />
(1997) <strong>and</strong> Cairns et al. (1997) trained SRNs with a version <strong>of</strong> <strong>the</strong> BPTT<br />
learning algorithm on English phonotactics. They used 2 million<br />
phonological segments <strong>der</strong>ived from a transcribed speech corpus <strong>and</strong><br />
encoded with a vector containing nine phonological features. The neural<br />
network was presented a single phoneme at a time <strong>and</strong> was trained to<br />
produce <strong>the</strong> previous, <strong>the</strong> current <strong>and</strong> <strong>the</strong> next phonemes. The output<br />
corresponding to <strong>the</strong> predicted phoneme was matched against <strong>the</strong> following
Learning Phonotactics with Simple Processors 97<br />
phoneme, measuring cross-entropy; this produced a varying error signal<br />
with occasional peaks corresponding to word boundaries. The SRN<br />
reportedly learned to reproduce <strong>the</strong> current phoneme <strong>and</strong> <strong>the</strong> previous one,<br />
but was poor at predicting <strong>the</strong> following phoneme. Correspondingly, <strong>the</strong><br />
segmentation performance was quite modest, predicting only about onefifth<br />
<strong>of</strong> <strong>the</strong> word boundaries correctly, but it was more successful in<br />
predicting syllable boundaries. It was significantly improved by adding<br />
o<strong>the</strong>r cues such as prosodic information. This means that phonotactics<br />
might be used alone for syllable detection, but polysyllabic word detection<br />
needs extra cues.<br />
In ano<strong>the</strong>r connectionist study on phonological regularities, Rodd (1997)<br />
trained SRNs on 602 Turkish words; <strong>the</strong> networks were trained to predict<br />
<strong>the</strong> following phonemes. Analyzing <strong>the</strong> hidden layer representations<br />
developed during <strong>the</strong> training, <strong>the</strong> author found that hidden units came to<br />
correspond to graded detectors for natural phonological classes such as<br />
vowels, consonants, voiced stops <strong>and</strong> front <strong>and</strong> back vowels. This is fur<strong>the</strong>r<br />
evidence that NN models can capture important properties <strong>of</strong> <strong>the</strong> data <strong>the</strong>y<br />
have been trained on without any prior knowledge, based only on statistical<br />
co-occurrences.<br />
Learning <strong>the</strong> graphotactics <strong>and</strong> phonotactics <strong>of</strong> Dutch monosyllables<br />
with connectionist models was first explored by Tjong Kim Sang (1995)<br />
<strong>and</strong> Tjong Kim Sang & Nerbonne (1999), who trained SRNs to predict<br />
graphemes/phonemes based on preceding segments. The data was<br />
orthogonally encoded, that is, for each phoneme or grapheme <strong>the</strong>re was<br />
exactly one neuron activated at <strong>the</strong> input <strong>and</strong> output layers (see below 3.1).<br />
To test <strong>the</strong> knowledge learned by <strong>the</strong> network, Tjong Kim Sang <strong>and</strong><br />
Nerbonne tested whe<strong>the</strong>r <strong>the</strong> activation <strong>of</strong> <strong>the</strong> neurons corresponding to <strong>the</strong><br />
expected symbols are greater than a threshold determined as <strong>the</strong> lowest<br />
activation for some correct sequence encountered during <strong>the</strong> training data.<br />
This resulted in almost perfect acceptance <strong>of</strong> unseen Dutch words<br />
(generalization), but also in negligible discrimination with respect to (illformed)<br />
r<strong>and</strong>om strings. The authors concluded that “SRNs are unfit for<br />
processing our data set” (Tjong Kim Sang & Nerbonne, 1999).<br />
These early works on learning phonotactics with SRNs prompted <strong>the</strong><br />
work reported here. First, Stoianov et al. (1998) demonstrated that <strong>the</strong><br />
SRNs in Tjong Kim Sang <strong>and</strong> Nerbonne's work were learning phonotactics<br />
ra<strong>the</strong>r better than those authors had realized. By analyzing <strong>the</strong> error as a<br />
function <strong>of</strong> <strong>the</strong> acceptance threshold, Stoianov et al. (1998) were able to<br />
demonstrate <strong>the</strong> existence <strong>of</strong> thresholds successful at both <strong>the</strong> acceptance <strong>of</strong>
98 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
well-formed data <strong>and</strong> <strong>the</strong> rejection <strong>of</strong> ill-formed data (see below 3.6.2 for a<br />
description <strong>of</strong> how we determine such thresholds). The interval <strong>of</strong> highperforming<br />
thresholds is narrow, which is why earlier work had not<br />
identified it (see Figure 2 on how narrow <strong>the</strong> window is). More recently,<br />
Stoianov & Nerbonne (2000) have studied <strong>the</strong> performance <strong>of</strong> SRNs from a<br />
cognitive perspective, attending to <strong>the</strong> errors produced by <strong>the</strong> network <strong>and</strong><br />
to what extent it correlates with <strong>the</strong> performance <strong>of</strong> humans on related<br />
lexical decision tasks. The current article ties <strong>the</strong>se two str<strong>and</strong>s <strong>of</strong> work <strong>and</strong><br />
presents it systematically.<br />
3. Experiments<br />
The challenge in connectionist modeling is not only developing <strong>the</strong>oretical<br />
frameworks, but also obtaining <strong>the</strong> most from <strong>the</strong> network models during<br />
experimentation. This section focuses on experiments on learning <strong>the</strong><br />
phonotactics <strong>of</strong> Dutch syllables with Simple Recurrent Networks <strong>and</strong><br />
discusses a number <strong>of</strong> related problems. It will be followed by a study on<br />
<strong>the</strong> network behavior from a linguistic point <strong>of</strong> view.<br />
3.1. Some implementation decisions<br />
SRNs were presented in section 2. A first implementation decision<br />
concerns how sounds are to be represented. A simple orthogonal strategy is<br />
to choose a vector <strong>of</strong> n neurons to represent n phonemes, to assign each<br />
phoneme (e.g. /�/) to a neuron (e.g., neuron 5 in a sequence <strong>of</strong> 45), <strong>and</strong> <strong>the</strong>n<br />
to activate that one neuron <strong>and</strong> deactivate all <strong>the</strong> o<strong>the</strong>rs whenever <strong>the</strong><br />
phoneme is to be represented (so a /�/ is represented by four deactivated<br />
neurons, a single activated one, <strong>and</strong> <strong>the</strong>n forty more deactivated neurons).<br />
This orthogonal strategy makes no assumptions about phonemes being<br />
naturally grouped into classes on <strong>the</strong> basis <strong>of</strong> linguistic features such as<br />
consonant/vowel status, voicing, place <strong>of</strong> articulation, etc. An alternative<br />
strategy exploits such features by assigning each feature to a neuron <strong>and</strong><br />
<strong>the</strong>n representing a phoneme via a translation <strong>of</strong> its feature description into<br />
a sequence <strong>of</strong> corresponding neural activations.<br />
In phonotactics learning, <strong>the</strong> input encoding method might be featurebased<br />
or orthogonal, but <strong>the</strong> output decoding should be orthogonal in or<strong>der</strong><br />
to obtain a simple prediction <strong>of</strong> successors, <strong>and</strong> to avoid a bias induced
Learning Phonotactics with Simple Processors 99<br />
from <strong>the</strong> peculiarities <strong>of</strong> <strong>the</strong> feature encoding scheme used. The input<br />
encoding chosen was also orthogonal, which also requires <strong>the</strong> network<br />
discover natural classes <strong>of</strong> phonemes by itself.<br />
The orthogonal encoding implies that we need as many neurons as we<br />
have phonemes, plus one for <strong>the</strong> end-<strong>of</strong>-word '#' symbol. That is, <strong>the</strong> input<br />
<strong>and</strong> output layers will have 45 neurons. However, it is usually difficult to<br />
choose <strong>the</strong> right size <strong>of</strong> <strong>the</strong> hidden layer for a particular learning problem.<br />
That size is ra<strong>the</strong>r indirectly related to <strong>the</strong> learning task <strong>and</strong> encoding<br />
chosen (as a subcomponent <strong>of</strong> <strong>the</strong> learning task). A linguistic bias in <strong>the</strong><br />
encoding scheme, e.g., feature-based encoding, would simplify <strong>the</strong> learning<br />
task <strong>and</strong> decrease <strong>the</strong> number <strong>of</strong> <strong>the</strong> hidden neurons required learning it<br />
(Stoianov, 2001). Intuition tells us that hidden layers that are too small lead<br />
to an overly crude representation <strong>of</strong> <strong>the</strong> problem <strong>and</strong> larger error. Larger<br />
hidden layers, on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, increase <strong>the</strong> chance that <strong>the</strong> network<br />
w<strong>and</strong>ers aimlessly because <strong>the</strong> space <strong>of</strong> possibilities it needs to traverse is<br />
too large. Therefore, we sought an effective size in a pragmatic fashion.<br />
Starting with a plausible size, we compared its performance to nets with<br />
double <strong>and</strong> half <strong>the</strong> number <strong>of</strong> neurons in <strong>the</strong> hidden layer. We repeated in<br />
<strong>the</strong> direction <strong>of</strong> <strong>the</strong> better behavior, keeping track <strong>of</strong> earlier bounds in or<strong>der</strong><br />
to home in on an appropriate size. In this way we settled on a range <strong>of</strong> 20-<br />
80 neurons in <strong>the</strong> hidden layer, <strong>and</strong> we continued experimentation on<br />
phonotactic learning using only nets <strong>of</strong> this size.<br />
However, even given <strong>the</strong> right size <strong>of</strong> <strong>the</strong> hidden layer, <strong>the</strong> training will<br />
not always result in an optimal weight set W* since <strong>the</strong> network learning is<br />
nondeterministic - each network training process depends on a number <strong>of</strong><br />
stochastic variables, e.g., initial network weights <strong>and</strong> an or<strong>der</strong> <strong>of</strong><br />
presentation <strong>of</strong> examples. Therefore, in or<strong>der</strong> to produce more successful<br />
learning, several SRNs with different initial weights were trained in a pool<br />
(group).<br />
The back-propagation learning algorithm is controlled by two main<br />
parameters - a learning coefficient � <strong>and</strong> a smoothing parameter �. The first<br />
one controls <strong>the</strong> speed <strong>of</strong> <strong>the</strong> learning <strong>and</strong> is usually set within <strong>the</strong> range<br />
(0.1…0.3). It is advisable to choose a smaller value when <strong>the</strong> hidden layer<br />
is larger. Also, this parameter may vary in time, starting with a larger initial<br />
value that decreases progressively in time (as suggested in Kuan, Hornik &<br />
White (1994) for <strong>the</strong> learning algorithm to improve its chances at attaining<br />
a global minimum in error). Intuitively, such a schedule helps <strong>the</strong> network<br />
approximately to locate <strong>the</strong> region with <strong>the</strong> global minima <strong>and</strong> later to<br />
make more precise steps in searching for that minimum (Haykin, 1994;
100 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
Reed & Marks, II 1999). The smoothing parameter � will be set to 0.7,<br />
which also allows <strong>the</strong> network to escape from local minima during <strong>the</strong><br />
search walk over <strong>the</strong> error surface.<br />
The training process also depends on <strong>the</strong> initial values <strong>of</strong> <strong>the</strong> weights.<br />
They are set to r<strong>and</strong>om values drawn from a region (-r...+r). It is also<br />
important to find a proper value for r, since large initial weight values will<br />
produce chaotic network behavior, impeding <strong>the</strong> training. We used r = 0.1.<br />
The SRNs used for this problem are schematically represented in Fig. 1,<br />
where <strong>the</strong> SRN reaction to an input sequence /n�/ after training on an<br />
exemplary set containing <strong>the</strong> sequences /n�t#/, /n�ts#/, /n�t��rk#/ is given.<br />
For this particular database, <strong>the</strong> network has experienced <strong>the</strong> tokens '#', /s/<br />
<strong>and</strong> /�/ as possible successors to /n��/ during training <strong>and</strong> <strong>the</strong>refore it will<br />
activate <strong>the</strong>m in response to this input sequence.<br />
3.2. Linguistic Data - Dutch syllables<br />
A data base L M <strong>of</strong> all Dutch monosyllables - 5,580 words - was extracted<br />
from <strong>the</strong> CELEX (1993) lexical database. CELEX is a difficult data source<br />
because it contains many rare <strong>and</strong> foreign words among its approximately<br />
350,000 Dutch lexical entries, which additionally complicate <strong>the</strong> learning<br />
task. Filtering out non-typical words is a formidable task <strong>and</strong> one which<br />
might introduce experimenter prejudice, <strong>and</strong> <strong>the</strong>refore all monosyllables<br />
were used. The monosyllables have a mean length <strong>of</strong> 4.1(� = 0.94; min = 2;<br />
max = 8) tokens <strong>and</strong> are built from a set <strong>of</strong> 44 phonemes plus one extra<br />
symbol representing space (#) used as a filler specifying end-<strong>of</strong>-word.<br />
The main dataset is split into a training (L 1 ) <strong>and</strong> a testing (L 2 ) database<br />
in proportion approximately 85% to 15%. The training database will be<br />
used to train a Simple Recurrent Network <strong>and</strong> <strong>the</strong> testing one will be used<br />
for evaluating <strong>the</strong> success <strong>of</strong> word recognition. Negative data also will be<br />
created for test purposes. The complete database L M will be used for some<br />
parts <strong>of</strong> evaluation.<br />
In language modeling it is important to explore <strong>the</strong> frequencies <strong>of</strong> word<br />
occurrences which naturally bias humans' linguistic performance. If a<br />
model is trained on data in proportion to its empirical frequency, this<br />
focuses <strong>the</strong> learning on <strong>the</strong> more frequent words <strong>and</strong> thus improves <strong>the</strong><br />
performance <strong>of</strong> <strong>the</strong> model. This also makes feasible a comparison <strong>of</strong> <strong>the</strong><br />
model's performance with that <strong>of</strong> humans performing various linguistic<br />
tasks, such as a lexical decision task. For <strong>the</strong>se reasons, we used <strong>the</strong> word
Learning Phonotactics with Simple Processors 101<br />
frequencies given in <strong>the</strong> CELEX database. Because <strong>the</strong> frequencies vary<br />
greatly ([0...100,000]), we presented training data items in proportion with<br />
<strong>the</strong> natural logarithms <strong>of</strong> <strong>the</strong>ir frequencies, in accordance with st<strong>and</strong>ard<br />
practice (Plaut, McClell<strong>and</strong>, Seidenberg & Patterson, 1996). This approach<br />
resulted in frequencies in a range <strong>of</strong> [1...12].<br />
3.3. Difficulty<br />
<strong>On</strong>e way to characterize <strong>the</strong> complexity <strong>of</strong> <strong>the</strong> training set is to compute <strong>the</strong><br />
entropy <strong>of</strong> <strong>the</strong> distribution <strong>of</strong> successors, for every available left context.<br />
The entropy <strong>of</strong> a language L viewed as a stochastic process measures <strong>the</strong><br />
average surprise value associated with each element (Mitchell, 1997). In<br />
our case, <strong>the</strong> language is a set <strong>of</strong> words <strong>and</strong> <strong>the</strong> elements are phonemes,<br />
hence <strong>the</strong> appropriate entropy measures <strong>the</strong> average surprise value for<br />
phonemes c preceded by a context s. Entropy is measured for a given<br />
distribution, which in our case is <strong>the</strong> set <strong>of</strong> all possible successors. We<br />
compute entropy Entr(s) for a given context s with (1):<br />
Equation 1. Entropy<br />
where � is <strong>the</strong> alphabet <strong>of</strong> segment symbols, <strong>and</strong> p(c) <strong>the</strong> probability <strong>of</strong> a<br />
given context. Then <strong>the</strong> average entropy for all available contexts s∈L,<br />
weighted with <strong>the</strong>ir frequencies, will be <strong>the</strong> measure <strong>of</strong> <strong>the</strong> complexity <strong>of</strong><br />
<strong>the</strong> words. The smaller this measure, <strong>the</strong> less difficult are <strong>the</strong> words. The<br />
maximal possible value for one context would be log2(45), that is, 5.49, <strong>and</strong><br />
this would only obtain for <strong>the</strong> unlikely case that each phoneme was equally<br />
likely in that context. The actual average value <strong>of</strong> <strong>the</strong> entropy measured for<br />
<strong>the</strong> Dutch monosyllables, is 2.24, � = 1.32. The minimal value was 0.0, <strong>and</strong><br />
<strong>the</strong> maximal value was 3.96. These values may be interpreted as follows:<br />
The minimal value <strong>of</strong> 0.0 means that <strong>the</strong>re are left contexts with only one<br />
possible successor (log2(1) = 0). A maximal value <strong>of</strong> 3.96 means that <strong>the</strong>re<br />
is one context which is as unpredictable as one in which 2 3.96 = 16<br />
successors were equally likely. The mean entropy is 2.24, which is to say<br />
that in average 4.7 phonemes follow a given left context.
102 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
3.4. Negative Data<br />
We noted above that negative data is also necessary for evaluation. Since<br />
we are interested in models that discriminate more precisely <strong>the</strong> strings<br />
from L (<strong>the</strong> Dutch syllables), <strong>the</strong> negative data for <strong>the</strong> following<br />
experiments will be biased toward L.<br />
Three negative testing sets were generated <strong>and</strong> used: First, a set RM containing strings with syllabic form [C] 0...3 V[C] 0...4 , based on <strong>the</strong> empirical<br />
observation that <strong>the</strong> Dutch mono-syllables have up to three onset (word<br />
initial) consonants <strong>and</strong> up to four coda (word final) consonants. The second<br />
group consists <strong>of</strong> three sub-sets <strong>of</strong> R M: {R 1 M , R 2 M , R M<br />
3 + }, with fixed distances <strong>of</strong><br />
<strong>the</strong> r<strong>and</strong>om strings to any existing Dutch word at 1, 2, <strong>and</strong> 3+ phonemes,<br />
respectively (measured by edit distance (Nerbonne, Heeringa & Kleiweg,<br />
1999)). Controlling for <strong>the</strong> distance to any training word allows us to assess<br />
more precisely <strong>the</strong> performance <strong>of</strong> <strong>the</strong> model. And finally, a third group:<br />
r<strong>and</strong>om strings built <strong>of</strong> concatenations <strong>of</strong> n-grams picked r<strong>and</strong>omly from<br />
Dutch monosyllables. In particular, two sets - R 2 N <strong>and</strong> R 3 N - were r<strong>and</strong>omly<br />
developed, based on bigrams <strong>and</strong> trigrams, correspondingly.<br />
The latter groups are <strong>the</strong> most "difficult" ones, <strong>and</strong> especially R 3 N ,<br />
because it consists <strong>of</strong> strings that are closest to Dutch. They are also useful<br />
for <strong>the</strong> comparison <strong>of</strong> SRN methods to n-gram modeling. The<br />
corresponding n-gram models will always wrongly recognize <strong>the</strong>se r<strong>and</strong>om<br />
strings as words from <strong>the</strong> language. Where <strong>the</strong> connectionist predictor<br />
recognizes <strong>the</strong>m as non-words, it outperforms <strong>the</strong> corresponding n-gram<br />
models, which are consi<strong>der</strong>ed as benchmark models for prediction tasks<br />
such as phonotactics learning.<br />
3.5. Training<br />
This section reports on network training. We will add a few more details<br />
about <strong>the</strong> training procedure, <strong>the</strong>n we will present pilot experiments aimed<br />
at determining <strong>the</strong> hidden layer size. The later parts will analyze <strong>the</strong><br />
network performance.
3.5.1. Procedure<br />
Learning Phonotactics with Simple Processors 103<br />
The networks were trained in a pool on <strong>the</strong> same problem, <strong>and</strong><br />
independently <strong>of</strong> each o<strong>the</strong>r, with <strong>the</strong> BPTT learning algorithm. The<br />
training <strong>of</strong> each individual network was organized in epochs, in <strong>the</strong> course<br />
<strong>of</strong> which <strong>the</strong> whole training data set is presented in accordance with <strong>the</strong><br />
word frequencies. The total <strong>of</strong> <strong>the</strong> logarithm <strong>of</strong> <strong>the</strong> frequencies in <strong>the</strong><br />
training data base L 1 M is about 11,000, which is also <strong>the</strong> number <strong>of</strong><br />
presentations <strong>of</strong> sequences per epoch, drawn in a r<strong>and</strong>om or<strong>der</strong>. Next, for<br />
each word, <strong>the</strong> corresponding sequence <strong>of</strong> phonemes is presented to <strong>the</strong><br />
input, one at a time, followed by <strong>the</strong> end-<strong>of</strong>-sequence marker `#'. Each time<br />
step is completed by copying <strong>the</strong> hidden layer activations to <strong>the</strong> context<br />
layer, which is used in <strong>the</strong> following step.<br />
The parameters <strong>of</strong> <strong>the</strong> learning algorithm were as follows: <strong>the</strong> learning<br />
coefficient � started at 0.3 <strong>and</strong> dropped by 30% each epoch, finishing at<br />
0.001; <strong>the</strong> momentum (smoothing) term � = 0.7. The networks required 30<br />
epochs to complete training. After this point, very little improvement is<br />
noted.<br />
3.5.2. Pilot experiments<br />
Pilot experiments aiming at searching for <strong>the</strong> most appropriate hidden layer<br />
size were done with 20, 40 <strong>and</strong> 80 hidden neurons. In or<strong>der</strong> to avoid o<strong>the</strong>r<br />
nondeterminism which comes from <strong>the</strong> r<strong>and</strong>om selection <strong>of</strong> negative data,<br />
during <strong>the</strong> pilot experiments <strong>the</strong> network was tested solely on its ability to<br />
distinguish admissible from inadmissible successors. Those experiments<br />
were done with a small pool <strong>of</strong> three networks, each <strong>of</strong> <strong>the</strong>m trained for 30<br />
epochs, which resulted in approximately 330,000 word presentations or<br />
1,300,000 segments. The total number <strong>of</strong> individual word presentations<br />
ranged from 30 to 300, according to <strong>the</strong> individual word frequencies. The<br />
results <strong>of</strong> <strong>the</strong> training are given in Table 1, un<strong>der</strong> <strong>the</strong> group <strong>of</strong> columns<br />
"Optimal phonotactics". In <strong>the</strong> course <strong>of</strong> <strong>the</strong> training, <strong>the</strong> networks typically<br />
started with a sharp error drop to about 13%, which soon turned into a very<br />
slow decrease (see Table 2, left 3 columns).<br />
The training <strong>of</strong> <strong>the</strong> three pools with hidden layer size 20, 40 <strong>and</strong> 80,<br />
resulted in networks with similar performance, with <strong>the</strong> largest network<br />
performing best. Additional experiments with SRNs with 100 hidden<br />
neurons resulted in larger errors than a network with 80 hidden neurons, so
104 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
that we settled experimentally on 80 hidden neurons as <strong>the</strong> likely optimal<br />
size. It is clear that this procedure is rough, <strong>and</strong> that one needs to be on<br />
guard against premature concentration on one size model.<br />
Table 1. Results <strong>of</strong> a pilot study on phonotactics learning by SRNs with 20, 40,<br />
<strong>and</strong> 80 (rows) hidden neurons. Each network is independently trained on<br />
language LM three times (columns). The performance is measured (left 3<br />
columns) using <strong>the</strong> error in predicting <strong>the</strong> next phoneme, <strong>and</strong> (right 3<br />
columns) using L2 (semi-Euclidean) distance between <strong>the</strong> empirical<br />
context-dependent predictions <strong>and</strong> <strong>the</strong> network predictions for each<br />
context in <strong>the</strong> tree. Those two methods do not depend on r<strong>and</strong>omly<br />
chosen negative data.<br />
Optimal Phonotactics ||SRN L , T L ||L2<br />
Hidd Layer Size SRN1 SRN2 SRN3 SRN1 SRN2 SRN3<br />
20 10.57% 10.65% 10.57% 0.0643 0.0642 0.0642<br />
40 10.44% 10.51% 10.44% 0.0637 0.0637 0.0637<br />
80 10.00% 9.97% 10.02% 0.0634 0.0634 0.0632<br />
Table 2. A typical shape <strong>of</strong> <strong>the</strong> SRN error during training. The error drops sharply<br />
in <strong>the</strong> beginning <strong>and</strong> <strong>the</strong>n slowly decreases to convergence.<br />
Epoch 1 2-4 5-10 11-15 16-30<br />
Error (%) 15.0 12.0 10.8 10.7 10.5<br />
3.6. Evaluation<br />
The performance <strong>of</strong> a neural predictor trained on phonotactics may be<br />
evaluated with different methods, depending on <strong>the</strong> particular task <strong>the</strong><br />
network is applied to. In this section we evaluate <strong>the</strong> neural networks<br />
performing best during <strong>the</strong> pilot studies.<br />
3.6.1. Likelihoods<br />
The direct outcome <strong>of</strong> training <strong>the</strong> sequential prediction task is learning <strong>the</strong><br />
successors' distribution. This will <strong>the</strong>refore be used as a basic evaluation<br />
method: <strong>the</strong> empirical context-dependent successor distribution Ps L (C) will<br />
be matched against <strong>the</strong> network context dependent predictions NPs L (C). For
Learning Phonotactics with Simple Processors 105<br />
this purpose, <strong>the</strong> output <strong>of</strong> <strong>the</strong> network will be normalized <strong>and</strong> matched<br />
against <strong>the</strong> distribution in <strong>the</strong> language data. This procedure resulted in a<br />
mean L2 (semi-Euclidean) distance <strong>of</strong> 0.063 - 0.064, where <strong>the</strong> optimal<br />
value would be zero (see Table 1, right 3 columns). 2 These values are close<br />
to optimal but baseline models (completely r<strong>and</strong>om networks) also result in<br />
approximately 0.085 L2 distance.<br />
3.6.2. Phonotactic Constraints<br />
To evaluate <strong>the</strong> network's success in becoming sensitive to phonotactic<br />
constraints, we first need to judge how well it predicts individual<br />
phonemes. For this purpose we seek a threshold above which phonemes are<br />
predicted to be admissible <strong>and</strong> below which <strong>the</strong>y are predicted to be<br />
inadmissible. This is done empirically - we perform a binary search for an<br />
optimal threshold, i.e. <strong>the</strong> threshold �� that minimizes <strong>the</strong> network error<br />
E(�). The classification obtained in this fashion constitutes <strong>the</strong> network's<br />
predictions about phonotactics.<br />
We now turn to evaluating <strong>the</strong> network's predictions: <strong>the</strong> method to<br />
evaluate <strong>the</strong> network from this point <strong>of</strong> view compares <strong>the</strong> contextdependent<br />
network predictions with <strong>the</strong> corresponding empirical<br />
distributions. For this purpose, <strong>the</strong> method described by Stoianov (2001)<br />
will be used. The algorithm traverses a trie (Aho, Hopcr<strong>of</strong>t & Ullman,<br />
1983: 163-169), which is a tree representing <strong>the</strong> vocabulary where initial<br />
segments are <strong>the</strong> first branches. Words are paths through this data structure.<br />
The algorithm computes <strong>the</strong> performance at <strong>the</strong> optimal threshold<br />
determined using <strong>the</strong> procedure described in <strong>the</strong> last paragraph, i.e., at <strong>the</strong><br />
threshold which determines which phonemes are admissible <strong>and</strong> which<br />
inadmissible (see also 2.1). This approach compares <strong>the</strong> actual distribution<br />
with <strong>the</strong> learned distribution, <strong>and</strong> we normally use <strong>the</strong> complete database L M<br />
for training <strong>and</strong> testing.<br />
Figure 2 shows <strong>the</strong> error <strong>of</strong> SRN1 8 0 at different values <strong>of</strong> <strong>the</strong> threshold.<br />
The optimal threshold searching procedure resulted in 6.0% erroneous<br />
phoneme prediction at a threshold <strong>of</strong> 0.0175. This means that if we want to<br />
predict phonemes with this SRN, <strong>the</strong>y would be accepted as allowed<br />
successors if <strong>the</strong> activation <strong>of</strong> <strong>the</strong> correspondent neurons are higher than<br />
0.0175.
106 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
3.6.3. Word Recognition<br />
Using an SRN trained on phoneme prediction as a word recognizing device<br />
shifts <strong>the</strong> focus from phoneme prediction to sequence classification. We<br />
wish to see whe<strong>the</strong>r it can classify sequences <strong>of</strong> phonemes into well-formed<br />
words on <strong>the</strong> one h<strong>and</strong> <strong>and</strong> ill-formed non-words on <strong>the</strong> o<strong>the</strong>r. To do this<br />
we need to translate <strong>the</strong> phoneme (prediction) values into sequence values.<br />
We do this by taking <strong>the</strong> sum <strong>of</strong> <strong>the</strong> phoneme error values for <strong>the</strong> sequence<br />
<strong>of</strong> phonemes in <strong>the</strong> string, normalized to correct for length effects. But to<br />
translate this sum into a classification, we again need to determine an<br />
acceptability threshold, <strong>and</strong> we use a variant <strong>of</strong> <strong>the</strong> same empirical<br />
optimization described above. The threshold arrived at for this purpose is<br />
slightly lower than <strong>the</strong> optimal threshold from <strong>the</strong> previous algorithm. This<br />
means that <strong>the</strong> network accepts more phonemes, which, however, is<br />
compensated for by <strong>the</strong> fact that a string is accepted only if all its phonemes<br />
are predicted. In string recognition it is better to increase <strong>the</strong> phoneme<br />
acceptance rate, because <strong>the</strong> chance to detect a non-word is larger when<br />
more tokens are tested.<br />
Figure 2. SRN error (in %) as a function <strong>of</strong> <strong>the</strong> threshold �� The False Negative<br />
Error increases as <strong>the</strong> threshold increases because more <strong>and</strong> more<br />
admissible phonemes are incorrectly rejected. At <strong>the</strong> same time, <strong>the</strong> False<br />
Positive Error decreases because fewer unwanted successors are falsely<br />
accepted. The mean <strong>of</strong> those two errors is <strong>the</strong> network error, which finds<br />
its minimum 6.0% at threshold �� = 0.0175. Notice that <strong>the</strong> optimal<br />
threshold is limited to a small range. This illustrates how critical <strong>the</strong><br />
exact setting <strong>of</strong> threshold is for good performance.
Learning Phonotactics with Simple Processors 107<br />
Since <strong>the</strong> performance measure here is <strong>the</strong> mean percentage <strong>of</strong> correctly<br />
recognized monosyllables <strong>and</strong> correctly rejected r<strong>and</strong>om strings, we<br />
incorporate both in seeking <strong>the</strong> optimal threshold. The negative data is as<br />
described above in 3.4. Concerning <strong>the</strong> positive data, this approach allows<br />
us to test <strong>the</strong> generalization capacity <strong>of</strong> <strong>the</strong> model, so that <strong>the</strong> training L 1 M<br />
<strong>and</strong> testing L 2 M subsets may be used here - <strong>the</strong> first for training <strong>the</strong> model<br />
<strong>and</strong> evaluating it during training, <strong>and</strong> <strong>the</strong> second to test <strong>the</strong> generalization<br />
capacity <strong>of</strong> <strong>the</strong> trained network.<br />
<strong>On</strong>ce we determine <strong>the</strong> optimal sequence-acceptance threshold (0.016),<br />
we obtain 5% error on <strong>the</strong> positive training dataset L 1 M <strong>and</strong> <strong>the</strong> negative<br />
strings from RM , where <strong>the</strong> error varied 0.5% depending on <strong>the</strong> r<strong>and</strong>om data<br />
set generated.<br />
The model was tested fur<strong>the</strong>r on <strong>the</strong> second group <strong>of</strong> negative data sets.<br />
As expected, strings which are more unlike Dutch resulted in smaller error.<br />
Performance on r<strong>and</strong>om strings from R N<br />
3 + is almost perfect. In <strong>the</strong> opposite<br />
case, <strong>the</strong> strings close to real words (from R 1 N ) resulted in larger error.<br />
The generalization capabilities <strong>of</strong> <strong>the</strong> network were tested on <strong>the</strong> L 2 M<br />
positive data, unseen during training. The error on this test set was about<br />
6%. An explanation <strong>of</strong> <strong>the</strong> increase <strong>of</strong> <strong>the</strong> error will be presented later,<br />
when <strong>the</strong> error will be studied by varying its properties.<br />
Ano<strong>the</strong>r interesting issue is how SRN performance compares to o<strong>the</strong>r<br />
known models, e.g. n-grams. The trained SRN definitely outperformed<br />
bigrams <strong>and</strong> trigrams, which was shown by testing <strong>the</strong> trained SRNs on <strong>the</strong><br />
non-words from R 2 N <strong>and</strong> R 3 N sets, yielding 19% <strong>and</strong> 35% error, respectively.<br />
This means that <strong>the</strong> SRN correctly rejected four out <strong>of</strong> five non-word<br />
strings composed <strong>of</strong> correct bigrams <strong>and</strong> two out <strong>of</strong> three non-word strings<br />
made <strong>of</strong> trigrams. To clarify, note that bigram models would have 100%<br />
error on R 2 N , <strong>and</strong> trigram models 100% error on R 3 N .<br />
4. Network Analysis<br />
The distributed representations in Neural Networks prevent <strong>the</strong> analysis <strong>of</strong><br />
generalizations in trained models by simple observation, which symbolic<br />
learning methods allow. Smaller NNs may be analyzed to some extent by<br />
examination, but for larger networks this is practically impossible.<br />
It is possible, however, to analyze trained networks to extract abstract<br />
knowledge about <strong>the</strong>ir behavior. Elman (1988), for example, trained an<br />
SRN to learn sentences <strong>and</strong> <strong>the</strong>n analyzed <strong>the</strong> hidden layer activations <strong>of</strong>
108 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
that SRN in various contexts, from which he showed that <strong>the</strong> network had<br />
internally developed syntactical categories. Similarly, we trained SRNs on<br />
phonotactics (Stoianov et al., 1998), <strong>and</strong> <strong>the</strong>n analyzed <strong>the</strong> network<br />
statically, by viewing <strong>the</strong> weight vectors <strong>of</strong> each neuron as pattern<br />
classifiers. We showed that <strong>the</strong> SRN had induced generalizations about<br />
phonetic categories. We follow that earlier work in or<strong>der</strong> to study network<br />
behavior, <strong>and</strong> we present <strong>the</strong> results <strong>of</strong> this study in <strong>the</strong> first subsection.<br />
Ano<strong>the</strong>r approach to <strong>the</strong> analysis <strong>of</strong> connectionist models assumes that<br />
<strong>the</strong>y are black boxes <strong>and</strong> examines <strong>the</strong> variation <strong>of</strong> network performance<br />
while varying some properties <strong>of</strong> <strong>the</strong> data (Plaut et al., 1996; Stoianov,<br />
Stowe & Nerbonne, 1999). For example, one can vary word frequency,<br />
length, etc., <strong>and</strong> study <strong>the</strong> network error. When modeling human cognitive<br />
functions with this approach one can compare <strong>the</strong> behavior <strong>of</strong> <strong>the</strong> cognitive<br />
system <strong>and</strong> its artificial models. For example, in phonotactic modeling, one<br />
can compare results from psycholinguistic studies <strong>of</strong> a lexical decision task<br />
with <strong>the</strong> network reaction. This will be subject <strong>of</strong> study in <strong>the</strong> rest <strong>of</strong> <strong>the</strong><br />
section.<br />
4.1. Weight Analysis<br />
The neurons <strong>of</strong> a neural network act as pattern classifiers. The inputs<br />
selectively activate one or ano<strong>the</strong>r neuron, depending on <strong>the</strong> weight<br />
vectors. This means that information about network structure may be<br />
extracted from <strong>the</strong> weight vectors.<br />
In this section we will present a cluster analysis <strong>of</strong> <strong>the</strong> neurons in <strong>the</strong><br />
output layer. For that purpose, <strong>the</strong> mean weight vectors <strong>of</strong> <strong>the</strong> output layer<br />
<strong>of</strong> one <strong>of</strong> <strong>the</strong> networks - SRN2 4 0 (from Table 1) - were clustered using a<br />
minimum variance (Ward's) method, <strong>and</strong> each vector in <strong>the</strong> resulting<br />
dendrogram was labeled with <strong>the</strong> phoneme it corresponds to. 3 The resulting<br />
diagram is shown in Figure 3.
Learning Phonotactics with Simple Processors 109<br />
Figure 3. Cluster analysis <strong>of</strong> <strong>the</strong> vector <strong>of</strong> <strong>the</strong> output neurons, labeled with <strong>the</strong><br />
phonemes <strong>the</strong>y correspond to. The weight vectors are split into clusters<br />
which roughly correspond to existing phonetic categories.
110 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
We can see that <strong>the</strong> weight vectors (<strong>and</strong> correspondingly, <strong>the</strong> phonemes)<br />
cluster into some well-known major natural classes - vowels (in <strong>the</strong> bottom)<br />
<strong>and</strong> consonants (<strong>the</strong> upper part). The vowels are split into two major<br />
categories: low vowels <strong>and</strong> semi-low, front vowels (/��, �, a, e/), <strong>and</strong> high,<br />
back ones. The latter, in turn, are clustered into round+ <strong>and</strong> round- classes.<br />
Consonants appear to be categorized in a way less congruent with<br />
phonetics. But here, too, some established groups are distinguished. The<br />
first subgroup contains non-coronal consonants (/f, k, m, p, x/) with <strong>the</strong><br />
exceptions <strong>of</strong> /l/ <strong>and</strong> /n/. Ano<strong>the</strong>r subgroup contains voiced obstruents (/�,<br />
d, �, ���/). The delimiter '#' is also clustered as a consonant, in a group with<br />
/t/, which is also natural. The upper part <strong>of</strong> <strong>the</strong> figure seems to contain<br />
phonemes from different groups, but we can recognize that most <strong>of</strong> <strong>the</strong>se<br />
phonemes are quite rare in Dutch monosyllables, e.g., /�/, perhaps because<br />
<strong>the</strong>y have been 'loaned' from o<strong>the</strong>r languages, e.g. /g/.<br />
4.2. Functional analysis<br />
We may also study NNs by examining <strong>the</strong>ir performance as a function <strong>of</strong><br />
factors such as word frequency, similarity neighborhood, <strong>and</strong> word length.<br />
Such an analysis relates computational language modeling to<br />
psycholinguistics, <strong>and</strong> we submit that it is useful to compare <strong>the</strong> models'<br />
performance with humans'. In this section we introduce several factors<br />
which have played a role in psycholinguistic <strong>the</strong>orizing. We <strong>the</strong>n examine<br />
<strong>the</strong> performance <strong>of</strong> our model as a function <strong>of</strong> <strong>the</strong>se factors.<br />
4.2.1. Psycholinguistic Factors<br />
Frequency is one <strong>of</strong> <strong>the</strong> most thoroughly investigated characteristics <strong>of</strong><br />
words that affect performance. Numerous previous studies have<br />
demonstrated that <strong>the</strong> ease <strong>and</strong> <strong>the</strong> time with which spoken words are<br />
recognized are monotonically related to <strong>the</strong> experienced frequency <strong>of</strong><br />
words in <strong>the</strong> language environment (Luce, Pisoni & Goldinger, 1990; Plaut<br />
et al., 1996). The general tendencies found are that <strong>the</strong> more frequent words<br />
are, <strong>the</strong> faster <strong>and</strong> <strong>the</strong> more precise <strong>the</strong>y are recognized.<br />
Our perception <strong>of</strong> a word is likewise known to depend on its similarity<br />
to o<strong>the</strong>r words. The similarity neighborhood <strong>of</strong> a word is defined as <strong>the</strong><br />
collection <strong>of</strong> words that are phonetically similar to it. Some neighborhoods
Learning Phonotactics with Simple Processors 111<br />
are dense with many phonetically similar words while o<strong>the</strong>rs are sparse<br />
with few.<br />
The so-called Col<strong>the</strong>arth-N measure <strong>of</strong> a word w counts <strong>the</strong> number <strong>of</strong><br />
words that might be produced by replacing a single letter <strong>of</strong> w with some<br />
o<strong>the</strong>r. We modify this concept slightly to make it sensitive to similarity <strong>of</strong><br />
sub-syllabic elements, so that we regard words as similar when <strong>the</strong>y share<br />
two <strong>of</strong> <strong>the</strong> subsyllabic elements - onset, nucleus <strong>and</strong> coda. Empty onsets or<br />
codas are counted as <strong>the</strong> same. The word neighborhood is computed by<br />
counting <strong>the</strong> number <strong>of</strong> <strong>the</strong> similar words. If implemented precisely, <strong>the</strong><br />
complexity <strong>of</strong> <strong>the</strong> measuring process just explained is high, so we reduce it<br />
by probing for sub-syllables ra<strong>the</strong>r than for units <strong>of</strong> variable size, starting<br />
from a single phoneme. This simplifies <strong>and</strong> speeds up processing. The<br />
neighborhood size <strong>of</strong> <strong>the</strong> corpus we used ranged from 0 to 77 <strong>and</strong> had mean<br />
value <strong>of</strong> �= 30; � = 13.<br />
For example, <strong>the</strong> phonological neighborhood <strong>of</strong> <strong>the</strong> Dutch word broeds<br />
/bruts/ is given below. Note that <strong>the</strong> neighborhood contains only Dutch<br />
words.<br />
/br�ts/, /brots/, /bruj/, /brujt/, /bruk/, /brur/, /brus/, /brut/, /buts/, /kuts/,<br />
/puts/, /tuts/<br />
These represent <strong>the</strong> pronunciations <strong>of</strong> Brits `British', broods `bread'<br />
(gen.sg.), broei `brew', broeit `brew' (3rd. sg.), broek `pants', broer<br />
`bro<strong>the</strong>r', broes `spray nozzle', broed `brood', boots `boots' (Eng. loan),<br />
koets `coach', poets `clean' <strong>and</strong> toets `test'. Among <strong>the</strong> words with very poor<br />
neighborhood are /����/ schwung, /b�rts/ boards, /�����jnt/ joint, <strong>and</strong><br />
/sk���rs/ squares, all <strong>of</strong> which are <strong>of</strong> foreign origin. Words such as /h�k/<br />
hek, /b�s/ bas, /l�xt/ lacht, <strong>and</strong> /b�kt/ bakt have large neighborhoods.<br />
It is still controversial how similarity neighborhood influences cognitive<br />
processes (Balota, Paul & Spieler, 1999). Intuitively, it seems likely that<br />
words with larger neighborhoods are easier to access due to many similar<br />
items, but from ano<strong>the</strong>r perspective <strong>the</strong>se words might be more difficult to<br />
access due to <strong>the</strong> nearby competitors <strong>and</strong> longer selection process.<br />
However, in <strong>the</strong> more specific lexical decision task, <strong>the</strong> overall activity <strong>of</strong><br />
many c<strong>and</strong>idates has been shown to facilitate lexical decisions, so we will<br />
look for <strong>the</strong> same effect here.<br />
The property word length might affect performance in <strong>the</strong> lexical<br />
decision task in two different ways. <strong>On</strong> one h<strong>and</strong>, longer words provide<br />
more evidence since more phonemes are available to decide whe<strong>the</strong>r <strong>the</strong>
112 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
input sequence is a word so that we expect higher precision for longer<br />
words, <strong>and</strong> lower precision for particularly short words. <strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>,<br />
network error accumulating in iteration increases <strong>the</strong> error in phoneme<br />
predictions at later positions, which in turn will increase <strong>the</strong> overall error<br />
for longer words. For <strong>the</strong>se reasons we expect U-shaped patterns <strong>of</strong> error as<br />
word length increases. Such a pattern was observed in a study on modeling<br />
grapheme-to-phoneme conversion with SRNs (Stoianov et al., 1999). Static<br />
NNs are less likely (than dynamic models such as SRNs) to produce such<br />
patterns.<br />
So far we have presented three main characteristics <strong>of</strong> <strong>the</strong> individual<br />
words, which we expect to affect <strong>the</strong> performance <strong>of</strong> <strong>the</strong> model. However, a<br />
statistical correlation analysis (bivariate Spearman test) showed that <strong>the</strong>y<br />
are not independent, which means that an analysis <strong>of</strong> <strong>the</strong> influence <strong>of</strong> any<br />
single factor should control for <strong>the</strong> rest. In particular, <strong>the</strong>re is high negative<br />
correlation between word neighborhood <strong>and</strong> word length (r = -0.476),<br />
smaller positive correlation between neighborhood <strong>and</strong> frequency (r =<br />
0.223), <strong>and</strong> very small negative correlation between frequency <strong>and</strong> word<br />
length (r = -0.107). Because <strong>of</strong> <strong>the</strong> large amount <strong>of</strong> data all <strong>the</strong>se<br />
coefficients are significant at <strong>the</strong> 0.001 level.<br />
Finally, it will be useful to seek a correlate in <strong>the</strong> simulation for reaction<br />
time, which psycholinguists are particularly fond <strong>of</strong> using as a probe to<br />
un<strong>der</strong>st<strong>and</strong>ing linguistic structure. Perhaps we can find an SRN correlate to<br />
Reaction Time (RT) for <strong>the</strong> lexical decision task in network confidence,<br />
i.e., <strong>the</strong> amount <strong>of</strong> evidence that <strong>the</strong> test string is a word from <strong>the</strong> training<br />
language. The less confident <strong>the</strong> network, <strong>the</strong> slower <strong>the</strong> reaction, which<br />
can be implemented with a lateral inhibition (Haykin, 1994; Plaut et al.,<br />
1996). The network confidence for a given word might be expressed as <strong>the</strong><br />
product <strong>of</strong> <strong>the</strong> activations <strong>of</strong> <strong>the</strong> neurons corresponding to <strong>the</strong> phonemes <strong>of</strong><br />
that word. A similar measure, which we call uncertainty U, is <strong>the</strong> negative<br />
sum <strong>of</strong> (output) neuron activation logarithms, normalized with respect to<br />
word length |w| (2). Note that U varies inversely with confidence. Less<br />
certain sequences get higher (positive) scores.
Equation 2.<br />
Learning Phonotactics with Simple Processors 113<br />
To analyze <strong>the</strong> influence <strong>of</strong> <strong>the</strong>se parameters, <strong>the</strong> network scores <strong>and</strong> Uvalues<br />
were recorded for each monosyllabic word at <strong>the</strong> optimal threshold<br />
��= 0.016. The data was <strong>the</strong>n submitted to <strong>the</strong> statistical package SPSS for<br />
analysis <strong>of</strong> variance using SPSS's General Linear Model (GLM). When<br />
analyzing network score, <strong>the</strong> analysis revealed main effects <strong>of</strong> all three<br />
parameters discussed above: word neighborhood size (F = 18.4; p <<br />
0.0001), word frequency (F = 19.2; p < 0.0001), word length (F = 11.5; p <<br />
0.0001). There was also interaction between neighborhood size <strong>and</strong> <strong>the</strong><br />
o<strong>the</strong>r parameters: <strong>the</strong> interaction with word frequency had an F -score 6.6<br />
<strong>and</strong> <strong>the</strong> interaction <strong>of</strong> <strong>the</strong> neighborhood with word length had an F-score <strong>of</strong><br />
4.9, both significant at 0.0001 level. Table 3 summarizes <strong>the</strong> findings. Error<br />
decreases both as neighborhood size <strong>and</strong> as frequency increases, <strong>and</strong> error<br />
dependent on length shows <strong>the</strong> predicted U-shaped form (Table 3c).<br />
Table 3. Effect <strong>of</strong> (a) frequency, (b) neighborhood density <strong>and</strong> (c) length on word<br />
uncertainty U <strong>and</strong> word error.<br />
a.<br />
Frequency Low Mid High<br />
U 2.30 2.20 2.18<br />
Error (%) 8.6 4.1 1.5<br />
b.<br />
Neighb. size Low Mid High<br />
U 2.62 2.30 2.21<br />
Error (%) 12.7 3.9 0.8<br />
c.<br />
Length Low Mid High<br />
U 2.63 2.20 2.13<br />
Error (%) 5.2 4.4 13.1
114 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
Analysis <strong>of</strong> variance on <strong>the</strong> U-values revealed similar dependencies. There<br />
were main effects <strong>of</strong> word neighborhood size (F = 58.2; p < 0.0001), word<br />
frequency (F = 45.9; p < 0.0001), word length (F = 137.5; p < 0.0001), as<br />
well as <strong>the</strong> earlier observed interactions between neighborhood density <strong>and</strong><br />
<strong>the</strong> o<strong>the</strong>r two variables: word length (F = 10.4; p < 0.001) <strong>and</strong> frequency (F<br />
= 5.235; p < 0.005).<br />
The frequency pattern <strong>of</strong> error <strong>and</strong> uncertainty variance was expected,<br />
given <strong>the</strong> increased evidence to <strong>the</strong> network for more frequent words. The<br />
displayed length effect showed that <strong>the</strong> influence <strong>of</strong> error gained in<br />
recursion is weaker than <strong>the</strong> effect <strong>of</strong> stronger evidence for longer words.<br />
Also, <strong>the</strong> pattern <strong>of</strong> performance when varying neighborhood density<br />
confirmed <strong>the</strong> hypo<strong>the</strong>sis <strong>of</strong> <strong>the</strong> lexical decision literature that larger<br />
neighborhoods makes it easier for words to be recognized as such.<br />
4.3. Syllabic structure<br />
Phonotactic constraints might hint at how <strong>the</strong> stream <strong>of</strong> phonemes is<br />
organized in <strong>the</strong> language processing system. The popular phoneme,<br />
syllable <strong>and</strong> word entities may not be <strong>the</strong> only units that we use for lexical<br />
access <strong>and</strong> production. There are suggestions that in addition, some subsyllabic<br />
elements are involved in those processes, that is, <strong>the</strong> syllables<br />
might have not linear structure, but more complex representations (Kessler<br />
& Treiman, 1997). For that purpose, we will analyze how <strong>the</strong> phoneme<br />
prediction error at a threshold <strong>of</strong> 0.016 - where <strong>the</strong> network resulted in best<br />
word recognition - is located within words with respect to <strong>the</strong> following<br />
sub-syllabic elements - onset, nucleus <strong>and</strong> coda. The particular hypo<strong>the</strong>sis<br />
that will be tested is whe<strong>the</strong>r Dutch monosyllables follow <strong>the</strong> structure<br />
below that was found in English as well (Kessler & Treiman, 1997).<br />
( <strong>On</strong>set - Rhyme (Nucleus - Coda) )<br />
The distribution <strong>of</strong> phoneme error within words (Table 4a) shows that <strong>the</strong><br />
network makes more mistakes at <strong>the</strong> beginning than at <strong>the</strong> end <strong>of</strong> words,<br />
where SRN becomes more confident in its decision. This could be<br />
explained with increasing contextual information that more severely<br />
restricts possible phonemic combinations. A more precise analysis <strong>of</strong> <strong>the</strong><br />
error position in <strong>the</strong> onset, <strong>the</strong> nucleus <strong>and</strong> <strong>the</strong> coda fur<strong>the</strong>r reveals o<strong>the</strong>r<br />
interesting phenomena (Table 4b).
Learning Phonotactics with Simple Processors 115<br />
Table 4. Distribution <strong>of</strong> phoneme prediction error at a threshold <strong>of</strong> 0.016 by (a)<br />
phoneme position within words <strong>and</strong> (b) phoneme position within subsyllables.<br />
Word <strong>and</strong> <strong>On</strong>set positions start from 2, because <strong>the</strong> prediction<br />
starts after <strong>the</strong> first phoneme.<br />
a.<br />
Word Position 2 3 4 5 6 7 8<br />
Error (%) 4.3 1.7 1.4 0.6 0.3 0.3 0.00<br />
b.<br />
Sub-syllabes <strong>On</strong>set Nucleus Coda<br />
Relative Position 2 3 1 1 2 3<br />
Error (%)<br />
2.6 0.0 4.5 1.0 1.5 2.0<br />
4<br />
2.6<br />
First, error within <strong>the</strong> coda increases at <strong>the</strong> coda's end. We attribute this to<br />
error accumulated toward <strong>the</strong> end <strong>of</strong> <strong>the</strong> words, as was predicted earlier.<br />
The mean entropy in <strong>the</strong> coda (1.32; � = 0.87) is smaller than <strong>the</strong> mean<br />
entropy in <strong>the</strong> onset (1.53; � = 0.78), where we do not observe such effects.<br />
So looser constraints are not <strong>the</strong> reason for <strong>the</strong> relatively greater error in <strong>the</strong><br />
coda. Next, <strong>the</strong> error at <strong>the</strong> transition onset-nucleus is much higher than <strong>the</strong><br />
error at <strong>the</strong> surrounding positions, which means that <strong>the</strong> break between<br />
onset <strong>and</strong> rhyme (<strong>the</strong> conjunction nucleus-coda) is significant. This<br />
distribution is also consistent with <strong>the</strong> statistical finding that <strong>the</strong> entropy is<br />
larger in <strong>the</strong> body (<strong>the</strong> transition point onset-nucleus) (3.45; � = 0.39), than<br />
in <strong>the</strong> rhyme (1.94; � = 1.21). All this data support <strong>the</strong> hypo<strong>the</strong>sis that onset<br />
<strong>and</strong> rhyme play significant roles in lexical access <strong>and</strong> that <strong>the</strong> syllabic<br />
structure confirmed for English by Kessler & Treiman (1997) is valid for<br />
Dutch, too.<br />
5. Conclusions<br />
Phonotactic constraints restrict <strong>the</strong> way phonemes combine in or<strong>der</strong> to form<br />
words. These constraints are empirical <strong>and</strong> can be abstracted from <strong>the</strong><br />
lexicon - ei<strong>the</strong>r by extracting rules directly, or via models <strong>of</strong> that lexicon.<br />
Existing language models are usually based on abstract symbolic methods,<br />
which provide good tools for studying such knowledge. But linguistic<br />
research from a connectionist perspective can provide a fresh perspective<br />
about language because <strong>the</strong> brain <strong>and</strong> artificial neural networks share<br />
principles <strong>of</strong> computations <strong>and</strong> data representations.
116 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
Connectionist language modeling, however, is a challenging task.<br />
Neural networks use distributed processing <strong>and</strong> continuous computations,<br />
while languages have a discrete, symbolic nature. This means that some<br />
special tools are necessary if one is to model linguistic problems with<br />
connectionist models. The research reported in this paper attempted to<br />
provide answers to two basic questions: first, whe<strong>the</strong>r phonotactic learning<br />
is possible at all in connectionist systems, which had been doubted earlier<br />
(Tjong Kim Sang, 1995; Tjong Kim Sang, 1998). In <strong>the</strong> case <strong>of</strong> a positive<br />
answer, <strong>the</strong> second question is how NN performance compares to human<br />
ability. In or<strong>der</strong> to draw this comparison, we needed to extract <strong>the</strong><br />
phonotactic knowledge from a network which has learned <strong>the</strong> sequential<br />
structure. We proposed several ways <strong>of</strong> doing this.<br />
Section 3 studied <strong>the</strong> first question. Even if <strong>the</strong>re are <strong>the</strong>oretical results<br />
demonstrating that NNs have <strong>the</strong> needed finite-state capacity for<br />
phonotactic processing, <strong>the</strong>re are practical limitations, so that we needed<br />
experimental support to demonstrate <strong>the</strong> practical capability <strong>of</strong> SRNs to<br />
learn phonotactics. A key to solving <strong>the</strong> problems <strong>of</strong> earlier investigators<br />
was to focus on finding a threshold that optimally discriminated <strong>the</strong><br />
continuous neuron activations with respect to phoneme acceptance <strong>and</strong><br />
rejection simultaneously. The threshold range at which <strong>the</strong> network<br />
achieves good discrimination is very small (see Figure 2), which illustrates<br />
how critical <strong>the</strong> exact setting <strong>of</strong> <strong>the</strong> threshold is. We also suggested that this<br />
threshold might be computed interactively, after processing each symbol,<br />
which is cognitively plausible, but we postpone a demonstration <strong>of</strong> this to<br />
ano<strong>the</strong>r paper.<br />
The network performance on word recognition - word acceptance rate <strong>of</strong><br />
95% <strong>and</strong> r<strong>and</strong>om string rejection rate <strong>of</strong> 95% at a threshold <strong>of</strong> 0.016 -<br />
competes with <strong>the</strong> scores <strong>of</strong> symbolic techniques such as Inductive Logic<br />
Programming <strong>and</strong> Hidden Markov Models (Tjong Kim Sang, 1998), both<br />
<strong>of</strong> which reflect low-level human processing architecture with less fidelity.<br />
Section 4 addressed <strong>the</strong> second question <strong>of</strong> how o<strong>the</strong>r linguistic<br />
knowledge encoded into <strong>the</strong> networks can be extracted. Two approaches<br />
were used. Section 4.1 clustered <strong>the</strong> weights <strong>of</strong> <strong>the</strong> network, revealing that<br />
<strong>the</strong> network has independently become sensitive to established phonetic<br />
categories.<br />
We went on to analyze how various factors which have been shown to<br />
play a role in human performance find <strong>the</strong>ir counterparts in <strong>the</strong> network's<br />
performance. Psycholinguistics has shown, for example, <strong>the</strong> ease <strong>and</strong> <strong>the</strong><br />
time with which spoken words are recognized are monotonically related to
Learning Phonotactics with Simple Processors 117<br />
<strong>the</strong> frequency <strong>of</strong> words in language experience (Luce et al., 1990). The<br />
model likewise reflected <strong>the</strong> importance <strong>of</strong> neighborhood density in<br />
facilitating word recognition, which we speculated stems from <strong>the</strong><br />
supportive evidence which more similar patterns lend to <strong>the</strong> words in <strong>the</strong>ir<br />
neighborhood. Whenever network <strong>and</strong> human subjects exhibit a similar<br />
sensitivity to well-established parameters, we see a confirmation <strong>of</strong> <strong>the</strong><br />
plausibility <strong>of</strong> <strong>the</strong> architecture chosen.<br />
Finally, <strong>the</strong> distribution <strong>of</strong> <strong>the</strong> errors within <strong>the</strong> words showed ano<strong>the</strong>r<br />
linguistically interesting result. In particular, <strong>the</strong> network tended to err<br />
more <strong>of</strong>ten at <strong>the</strong> transition onset-nucleus - which is also typical for<br />
transitions between adjacent words in <strong>the</strong> speech stream <strong>and</strong> used for<br />
speech segmentation. Analogically, we can conclude from this that <strong>the</strong><br />
nucleus-coda unit - <strong>the</strong> rhyme - is a significant linguistic unit for <strong>the</strong> Dutch<br />
language, a result suggested earlier for English (Kessler & Treiman, 1997).<br />
We wind up this conclusion with one disclaimer <strong>and</strong> a repetition <strong>of</strong> <strong>the</strong><br />
central claim. We have not claimed that SRNs are <strong>the</strong> only (connectionist)<br />
model capable <strong>of</strong> dynamic processing, nor that <strong>the</strong>y are biologically <strong>the</strong><br />
most plausible neural network. Our central claim is to have demonstrated<br />
that relatively simple connectionist mechanisms have <strong>the</strong> capacity to model<br />
<strong>and</strong> learn phonotactic structure.<br />
Notes<br />
1<br />
The authors are particularly pleased to <strong>of</strong>fer this piece to a Festschrift honoring<br />
Dr. Dr. h.c. Tjeerd de Graaf, who graciously agreed to cooperate in <strong>the</strong><br />
supervision <strong>of</strong> Stoianov's Ph.D. project 1997-2001 at <strong>the</strong> University <strong>of</strong><br />
Groningen. Even if Tjeerd is best known for his more recent work on<br />
descriptive linguistics, minority languages <strong>and</strong> language documentation, his<br />
early training in physics <strong>and</strong> earlier research on acoustic phonetics made him<br />
one <strong>of</strong> <strong>the</strong> best-suited supervisors for projects such as <strong>the</strong> one reported on here<br />
involving advanced learning algorithms. Tjeerd's sympathy with Eastern<br />
European languages <strong>and</strong> cultures is visceral <strong>and</strong> might have led him to agree in<br />
any case, but we particularly appreciated his phonetic acumen.<br />
2<br />
The distance is related to Euclidean, but more exactly <strong>the</strong> distance between <strong>the</strong><br />
two n-dimensional vectors is
118 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
3 The cluster analysis in Figure 3 was produced by programs written by Peter<br />
Kleiweg, available at http://www.let.rug.nl/alfa.<br />
References<br />
Aho, Alfred, John Hopcr<strong>of</strong>t & Jeffrey Ullman (1983). Data Structures <strong>and</strong><br />
Algorithms. Addison Wesley.<br />
Balota, David, Stephen Paul & Daniel Spieler (1999). Attentional control <strong>of</strong><br />
lexical processing pathways during word recognition <strong>and</strong><br />
reading. In: S. Garrod & M. Pickering (eds). Studies in cognition:<br />
Language processing. UCL Press, London, Engl<strong>and</strong>, 15-57.<br />
Cairns, Paul, R. Shillcock, Nick Chater & Joe Levy (1997). Bootstrapping word<br />
boundaries: A bottom-up corpus-based approach to speech<br />
segmentation. Cognitive Psychology, 33(2): 111-153.<br />
Carrasco, Rafael, Mikel Forcada & Ramon Neco (1999). Stable encoding <strong>of</strong><br />
finite-state machines in discrete-time recurrent neural networks<br />
with sigmoid units. Neural Computation, 12(9): 2129-2174.<br />
Carstairs-McCarthy, Andrew (1999). The Origins <strong>of</strong> Complex Language.<br />
Oxford Univ Press.<br />
CELEX (1993). The CELEX Lexical Data Base (cd-rom), Linguistic Data<br />
Consortium. http://www.kun.nl/celex.<br />
Christiansen, Morton H. & Nick Chater (1999). Toward a connectionist model<br />
<strong>of</strong> recursion in human linguistic performance. Cognitive Science,<br />
23: 157-205.<br />
Cleeremans, A., D. Servan-Schreiber & J.L. McClell<strong>and</strong> (1989). Finite state<br />
automata <strong>and</strong> simple recurrent networks. Neural Computation,<br />
1(3): 372-381.<br />
Cohen, A., C. Ebeling & A.G.F. van Holk (1972). Fonologie van het<br />
Ne<strong>der</strong>l<strong>and</strong>s en het Fries. Martinus Nijh<strong>of</strong>f, The Hague.<br />
Dell, Gary, Cornell Juliano & Anita Govindjee (1993). Structure <strong>and</strong> content in<br />
language production: A <strong>the</strong>ory <strong>of</strong> frame constraints in<br />
phonological speech errors. Cognitive Science, 17: 145-195.<br />
Dupoux, Emmanuel, Christophe Pallier, Kazuhiko Kakehi & Jacques Mehler<br />
(2001). New evidence for prelexical phonological processing in<br />
word recognition. Language <strong>and</strong> Cognitive Processes, 5(16):<br />
491-505.<br />
Elman, Jeffrey L. (1988). Finding structure in time. Technical Report 9901,<br />
Center for Research in Language, UCSD, CA.
Learning Phonotactics with Simple Processors 119<br />
Elman, Jeffrey L. (1991). Distributed representations, simple recurrent<br />
networks, <strong>and</strong> grammatical structure. Machine Learning, 7(2/3):<br />
195-226.<br />
Gasser, Michael (1992). Learning distributed representations for syllables. In:<br />
Proc. <strong>of</strong> 14th Annual Conference <strong>of</strong> Cognitive Science Society,<br />
396- 401.<br />
Haykin, Simon (1994). Neural Networks. Macmillian Publ, NJ.<br />
Kaplan, Ronald & Martin Kay (1994). Regular models <strong>of</strong> phonological rule<br />
systems. Computational Linguistics, 20/3: 331-378.<br />
Kessler, Brett & Rebecca Treiman (1997). Syllable structure <strong>and</strong> <strong>the</strong><br />
distribution <strong>of</strong> phonemes in English syllables. Journal <strong>of</strong> Memory<br />
<strong>and</strong> Language, 37: 295-311.<br />
Konstantopoulos, Stasinos (2003). Using Inductive Logic Programming to<br />
Learn Local Linguistic Structures. PhD <strong>the</strong>sis, Rijksuniversiteit<br />
Groningen.<br />
Kuan, Chung-Ming, Kurt Hornik & Halbert White (1994). A convergence<br />
result for learning in recurrent neural networks. Neural<br />
Computation, 6: 420-440.<br />
Laver, John (1994). Principles <strong>of</strong> <strong>Phonetics</strong>. Cambridge University Press,<br />
Cambridge.<br />
Lawrence, Steve, C. Lee Giles & S. Fong (1995). <strong>On</strong> <strong>the</strong> applicability <strong>of</strong> neural<br />
networks <strong>and</strong> machine learning methodologies to natural<br />
language processing. Technical report, Univ. <strong>of</strong> Maryl<strong>and</strong>.<br />
Luce, Paul L., David B. Pisoni & Steven D. Goldinger (1990). Similarity<br />
neighborhoods <strong>of</strong> spoken words. In: G. T. M. Altmann (ed.).<br />
Cognitive Models <strong>of</strong> Speech Processing. A Bradford Book,<br />
Cambridge, Massachusetts, USA, 122-147.<br />
McQueen, James (1998). Segmentation <strong>of</strong> continuous speech using<br />
phonotactics. Journal <strong>of</strong> Memory <strong>and</strong> Language, 39: 21-46.<br />
Mitchell, Thomas (1997). Machine Learning. McGraw Hill College.<br />
Nerbonne, John, Wilbert Heeringa & Peter Kleiweg (1999). Edit distance <strong>and</strong><br />
dialect proximity. In: D. Sank<strong>of</strong>f & J. Kruskal (eds). Time Warps,<br />
String Edits <strong>and</strong> Macromolecules: The Theory <strong>and</strong> Practice <strong>of</strong><br />
Sequence Comparison, 2nd ed.. CSLI, Stanford, CA, v-xv.<br />
Norris, D., J.M. McQueen, A. Cutler & S. Butterfield (1997). The possibleword<br />
constraint in <strong>the</strong> segmentation <strong>of</strong> continuous speech.<br />
Cognitive Psychology, 34: 191-243.<br />
Omlin, Christian W. & C. Lee Giles (1996). Constructing deterministic finitestate<br />
automata in recurrent neural networks. Journal <strong>of</strong> <strong>the</strong> ACM,<br />
43(6): 937-972.
120 John Nerbonne <strong>and</strong> Ivilin Stoianov<br />
Pacton, S., P. Perruchet, M. Fayol & A. Cleeremans (2001). Implicit learning in<br />
real world context: The case <strong>of</strong> orthographic regularities. Journal<br />
<strong>of</strong> Experimental Psychology: General, 130(3): 401-426.<br />
Plaut, D.C., J. McClell<strong>and</strong>, M. Seidenberg & K. Patterson (1996).<br />
Un<strong>der</strong>st<strong>and</strong>ing normal <strong>and</strong> impaired word reading:<br />
Computational principles in quasi-regular domains.<br />
Psychological Review, 103: 56-115.<br />
Reed, Russell D. & Robert J. Marks II (1999). Neural Smithing. MIT Press,<br />
Cambridge, MA.<br />
Reilly, Ronan (2002). The relationship between object manipulation <strong>and</strong><br />
language development in Broca's area: A connectionist<br />
simulation <strong>of</strong> Greenfield's hypo<strong>the</strong>sis. Behavioral <strong>and</strong> Brain<br />
Sciences, 25: 145-153.<br />
Robinson, A. J. & F. Fallside (1988). Static <strong>and</strong> dynamic error propagation<br />
networks with application to speech coding. In: D. Z. An<strong>der</strong>son<br />
(ed.). Neural Information Processing Systems. American Institute<br />
<strong>of</strong> Physics, NY.<br />
Rodd, Jennifer (1997). Recurrent neural-network learning <strong>of</strong> phonological<br />
regularities in Turkish. In: Proc. <strong>of</strong> Int. Conf. on Computational<br />
Natural Language Learning. Madrid, 97-106.<br />
Rumelhart, David E. & James A. McClell<strong>and</strong> (1986). Parallel Distributed<br />
Processing: Explorations <strong>of</strong> <strong>the</strong> Microstructure <strong>of</strong> Cognition. The<br />
MIT Press, Cambridge, MA.<br />
Rumelhart, D.E., G.E. Hinton & R.J. Williams (1986). Learning internal<br />
representations by error propagation. In: D. E. Rumelhart & J. A.<br />
McClell<strong>and</strong> (eds.). Parallel Distributed Processing: Explorations<br />
<strong>of</strong> <strong>the</strong> Microstructure <strong>of</strong> Cognition, Volume 1, Foundations . The<br />
MIT Press, Cambridge, MA, 318-363.<br />
Shillcock, Richard, Paul Cairns, Nick Chater & Joe Levy (1997). Statistical <strong>and</strong><br />
connectionist modelling <strong>of</strong> <strong>the</strong> development <strong>of</strong> speech<br />
segmentation. In: Broe<strong>der</strong> & Murre (eds.). Models <strong>of</strong> Language<br />
Learning. MIT Press.<br />
Shillcock, Richard, Joe Levy, Ge<strong>of</strong>f Lindsey, Paul Cairns & Nick Chater<br />
(1993). Connectionist modelling <strong>of</strong> phonological space In: T. M.<br />
Ellison & J. Scobbie (eds.). Computational <strong>Phonology</strong>.<br />
Edinburgh Working Papers in Cognitive Science, Edinburgh, 8:<br />
179-195<br />
Stoianov, Ivilin Peev (1998). Tree-based analysis <strong>of</strong> simple recurrent network<br />
learning. In: 36 Annual Meeting <strong>of</strong> <strong>the</strong> Association for<br />
Computational Linguistics <strong>and</strong> 17 Int. Conference on<br />
Compuational Linguistics. Vol. 2, Montreal, Canada, 1502-1504.
Learning Phonotactics with Simple Processors 121<br />
Stoianov, Ivilin Peev (2001). Connectionist Lexical Modelling. PhD <strong>the</strong>sis,<br />
Rijksuniversiteit Groningen.<br />
Stoianov, Ivilin Peev & John Nerbonne (2000). Exploring phonotactics with<br />
simple recurrent networks. In: F. van Eynde, I. Schuurman & N.<br />
Schelkens (eds.). Computational Linguistics in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s,<br />
1998. Rodopi, Amsterdam, NL, 51-68.<br />
Stoianov, Ivilin Peev, John Nerbonne & Huub Bouma (1998). Modelling <strong>the</strong><br />
phonotactic structure <strong>of</strong> natural language words with simple<br />
recurrent networks. In: P.-A. Coppen, H. van Halteren & L.<br />
Teunissen (eds.). Computational Linguistics in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s,<br />
1997. Rodopi, Amsterdam, NL, 77-96.<br />
Stoianov, Ivilin Peev, Laurie Stowe & John Nerbonne (1999). Connectionist<br />
learning to read aloud <strong>and</strong> correlation to human data. In: 21st<br />
Annual Meeting <strong>of</strong> <strong>the</strong> Cognitive Science Society, Vancouver,<br />
Canada. Lawrence Erlbaum Ass., London, 706-711.<br />
Stowe, Laurie, Anton Wijers, A. Willemsen, Eric Reul<strong>and</strong>, A. Paans & Wim<br />
Vaalburg (1994). Pet studies <strong>of</strong> language: An assessment <strong>of</strong> <strong>the</strong><br />
reliability <strong>of</strong> <strong>the</strong> technique. Journal <strong>of</strong> Psycholinguistic Research,<br />
23(6): 499-527.<br />
Tjong Kim Sang, Erick (1995). The limitations <strong>of</strong> modeling finite state<br />
grammars with simple recurrent networks. In: Proceedings <strong>of</strong> <strong>the</strong><br />
5th Computational Linguistics in The Ne<strong>the</strong>rl<strong>and</strong>s, 133-143.<br />
Tjong Kim Sang, Erick (1998). Machine Learning <strong>of</strong> Phonotactics. PhD <strong>the</strong>sis,<br />
Rijksuniversiteit Groningen.<br />
Tjong Kim Sang, Erik & John Nerbonne (1999). Learning simple phonotactics.<br />
In: Proceedings <strong>of</strong> <strong>the</strong> Workshop on Neural, Symbolic, <strong>and</strong><br />
Reinforcement Methods for Sequence Processing, Machine<br />
Learning Workshop at IJCAI '99, 41-46.<br />
Treiman, R. & A. Zukowski (1990). Toward an un<strong>der</strong>st<strong>and</strong>ing <strong>of</strong> English<br />
syllabification. Journal <strong>of</strong> Memory <strong>and</strong> Language, 34: 66-85.<br />
Tsoi, Ah Chung & Andrew Back (1997). Discrete time recurrent neural<br />
network architectures: A unifying review. Neurocomputing, 15:<br />
183-223.
Weak Interactions<br />
Yiddish influence in Hungarian, Esperanto <strong>and</strong><br />
Mo<strong>der</strong>n Hebrew<br />
Tamás Bíró<br />
When I arrived in Groningen, I was introduced to Tjeerd de Graaf as<br />
somebody speaking Hungarian. Then it turned out that both <strong>of</strong> us were<br />
interested in Yiddish. Fur<strong>the</strong>rmore, we shared <strong>the</strong> fact that we started our<br />
scientific life within physics, although, unlike Tjeerd, I have not worked as<br />
a physicist since my graduation. Never<strong>the</strong>less, as a second year physics<br />
student I received a research question from <strong>the</strong> late leading Hungarian<br />
physicist George Marx that was also somehow related to Tjeerd’s earlier<br />
research topic, neutrino astrophysics.<br />
Neutrinos are funny particles. They are extremely light, if <strong>the</strong>y have any<br />
mass, at all. 1 Therefore, <strong>the</strong>y cannot interact through gravitation. Because<br />
<strong>the</strong>y do not have any electrical charge ei<strong>the</strong>r, electromagnetic interaction is<br />
also unknown to <strong>the</strong>m. The only way <strong>the</strong>y can interact with <strong>the</strong> universe is<br />
<strong>the</strong> so-called weak interaction, one <strong>of</strong> <strong>the</strong> four fundamental forces. 2<br />
Nowadays physicists spend an inconceivable amount <strong>of</strong> budget building<br />
gigantic, un<strong>der</strong>ground basins containing millions <strong>of</strong> liters <strong>of</strong> heavy water<br />
just to try to detect a few neutrinos per year out <strong>of</strong> <strong>the</strong> very intense stream<br />
<strong>of</strong> neutrinos flowing constantly from <strong>the</strong> Sun <strong>and</strong> going through <strong>the</strong> Earth,<br />
that is, us. Even though <strong>the</strong>y almost never interact with regular material,<br />
through weak interaction <strong>the</strong>y play a fundamental role both in shaping what<br />
<strong>the</strong> universe looks like <strong>and</strong> in <strong>the</strong> Sun’s energy production. Therefore our<br />
life would not be possible without neutrinos <strong>and</strong> without weak interaction.<br />
Something similar happens in ethnolinguistics. The interaction between<br />
two languages may not always be very salient, <strong>and</strong> it cannot necessarily be<br />
explained by <strong>the</strong> most famous types <strong>of</strong> interactions. A weak interaction in<br />
linguistics might be an interaction which is not acknowledged by <strong>the</strong><br />
speakers’ community, for instance for ideologically reasons.<br />
In <strong>the</strong> present paper I shall present three cases <strong>of</strong> weak interaction<br />
between languages, un<strong>der</strong>stood in this sense, namely Yiddish affecting
124 Tamás Bíró<br />
Hungarian, Mo<strong>der</strong>n Hebrew (Israeli Hebrew) <strong>and</strong> Esperanto. All <strong>the</strong> stories<br />
take place in <strong>the</strong> late nineteenth or early twentieth century, when a new or<br />
mo<strong>der</strong>nized language had to be created. We shall observe what kind <strong>of</strong><br />
interactions took place un<strong>der</strong> which conditions. A model for interactions<br />
combined with <strong>the</strong> better un<strong>der</strong>st<strong>and</strong>ing <strong>of</strong> <strong>the</strong> social-historical setting will<br />
enable us to do so.<br />
1. Language interactions within a given socio-historical setting<br />
1.1. Modeling interactions<br />
In physics, <strong>the</strong> interaction between two bodies depends on three factors: <strong>the</strong><br />
two “eligibilities” <strong>of</strong> <strong>the</strong> parties to interact, as well as <strong>the</strong>ir distance. For<br />
gravity <strong>and</strong> electromagnetism, <strong>the</strong> formula probably familiar from highschool<br />
physics states that <strong>the</strong> force is proportional to <strong>the</strong> product <strong>of</strong> <strong>the</strong><br />
“eligibilities” - mass or electric charge - <strong>of</strong> <strong>the</strong> two bodies, divided by <strong>the</strong><br />
square <strong>of</strong> <strong>the</strong>ir distance. In o<strong>the</strong>r words, <strong>the</strong> higher <strong>the</strong> two masses (or<br />
electric charges) <strong>and</strong> <strong>the</strong> smaller <strong>the</strong> distance, <strong>the</strong> stronger <strong>the</strong> interaction.<br />
For Newton, who formulated this formula first, gravity was a long-range<br />
interaction. Mo<strong>der</strong>n physics has completed this picture with introducing<br />
exchange particles intermediating between <strong>the</strong> interacting bodies. 3 That<br />
way, contemporary science has also incorporated <strong>the</strong> view <strong>of</strong> Newton’s<br />
opponents who argued for <strong>the</strong> only possibility <strong>of</strong> short-range interactions.<br />
To transplant this image, vaguely, into <strong>the</strong> phenomenon <strong>of</strong> language<br />
interaction, we have to identify <strong>the</strong> eligibilities <strong>of</strong> <strong>the</strong> two interacting<br />
languages, <strong>the</strong>ir distance <strong>and</strong> <strong>the</strong> exchange particles. In fact, we can do that<br />
even on two levels. <strong>On</strong> a purely linguistic level, one can easily point to<br />
words <strong>and</strong> grammatical phenomena - “exchange particles” - w<strong>and</strong>ering<br />
from language to language. But it would be har<strong>der</strong> to identify in general <strong>the</strong><br />
properties <strong>of</strong> <strong>the</strong> phenomena <strong>and</strong> <strong>of</strong> <strong>the</strong> given languages that make <strong>the</strong><br />
interaction more probable or less probable.<br />
The sociolinguistic level is more promising for such an approach. In this<br />
case, <strong>the</strong> human beings are <strong>the</strong> exchange particles: people who leave one<br />
linguistic community in or<strong>der</strong> to join a new one. By <strong>the</strong> very fact <strong>of</strong> <strong>the</strong>ir<br />
moves, <strong>the</strong>y affect <strong>the</strong>ir new language by a linguistic quantum. The closer<br />
<strong>the</strong> two language communities, <strong>the</strong> more people will act as an exchange
Weak Interactions 125<br />
particle. Here distance should be un<strong>der</strong>stood not only based on geography,<br />
but on <strong>the</strong> intensity <strong>of</strong> <strong>the</strong> social network, as well. Thus, <strong>the</strong> more people<br />
w<strong>and</strong>er to <strong>the</strong> target community, <strong>the</strong> more linguistic impulse is brought to<br />
<strong>the</strong> second language <strong>and</strong> <strong>the</strong>refore <strong>the</strong> stronger <strong>the</strong> interaction. Note that <strong>the</strong><br />
physical analogy is not complete, since <strong>the</strong> symmetry <strong>of</strong> action <strong>and</strong> reaction<br />
is not guaranteed for interacting languages.<br />
The three cases to be discussed share <strong>the</strong> feature that <strong>the</strong> role <strong>of</strong> <strong>the</strong><br />
carriers <strong>of</strong> <strong>the</strong> interaction is played by late nineteenth century Eastern<br />
European Jews. In or<strong>der</strong> to un<strong>der</strong>st<strong>and</strong> <strong>the</strong> historical background, we have<br />
to recall what is called Haskala or Jewish Enlightenment.<br />
1.2. The Haskala<br />
By <strong>the</strong> late eighteenth century, <strong>the</strong> French <strong>and</strong> German Aufklärung had<br />
raised <strong>the</strong> question whe<strong>the</strong>r to emancipate <strong>and</strong> integrate - or assimilate - <strong>the</strong><br />
Jewish population on <strong>the</strong> one side, <strong>and</strong> an increasing wish to join <strong>the</strong><br />
European culture on <strong>the</strong> o<strong>the</strong>r. Although in <strong>the</strong> second half <strong>of</strong> <strong>the</strong> siècle des<br />
lumières <strong>the</strong>re were only a few Jewish intellectuals who articulated <strong>the</strong>se<br />
ideas, most <strong>of</strong> <strong>the</strong>m belonging to <strong>the</strong> circle <strong>of</strong> <strong>the</strong> philosopher Moses<br />
Mendelssohn (1729-1786) in Berlin, <strong>the</strong> next decades witnessed <strong>the</strong><br />
acculturation <strong>of</strong> a growing segment <strong>of</strong> <strong>the</strong> Jewish population in <strong>the</strong> German<br />
territories, as well as within <strong>the</strong> Austrian Empire. The eighteenth century<br />
Berlin Haskala is called <strong>the</strong> first stage <strong>of</strong> <strong>the</strong> Jewish Enlightenment,<br />
whereas <strong>the</strong> early nineteenth century social <strong>and</strong> cultural developments<br />
represent its second stage.<br />
What <strong>the</strong> first two stages <strong>of</strong> <strong>the</strong> Haskala yielded was including a Jewish<br />
color on <strong>the</strong> contemporary Western European cultural palette. “Jewish” was<br />
un<strong>der</strong>stood exclusively as one possible faith within <strong>the</strong> list <strong>of</strong> European<br />
religions, <strong>and</strong> nothing more than a religious conviction. An enlightened Jew<br />
was supposed to fully master <strong>the</strong> educated st<strong>and</strong>ard variant <strong>of</strong> <strong>the</strong> language<br />
<strong>of</strong> <strong>the</strong> society he lived in (Hochdeutsch¸ in most <strong>of</strong> <strong>the</strong> cases), without any<br />
“Jewish-like” feature. Propagating <strong>the</strong> knowledge <strong>of</strong> Hochdeutsch <strong>and</strong><br />
rolling back Jüdischdeutsch had already been <strong>the</strong> program <strong>of</strong> Moses<br />
Mendelssohn when he began writing a mo<strong>der</strong>n targum 4 <strong>of</strong> <strong>the</strong> Bible, <strong>the</strong><br />
Biur. Fur<strong>the</strong>r, <strong>the</strong> same Jew was expected to fully master <strong>the</strong> contemporary<br />
European culture, including classical languages, sciences <strong>and</strong> arts. The only<br />
sphere in which this Jew could express his or her being Jewish was <strong>the</strong><br />
diminished <strong>and</strong> europeanized arena <strong>of</strong> religious life. Diminished, because
126 Tamás Bíró<br />
<strong>of</strong> a secularization <strong>of</strong> life style; <strong>and</strong> Europeanized, due to <strong>the</strong> inclusion <strong>of</strong><br />
philosophical ideals <strong>of</strong> <strong>the</strong> Enlightenment toge<strong>the</strong>r with aes<strong>the</strong>tic models <strong>of</strong><br />
<strong>the</strong> Romanticism. The traditional religious duty <strong>of</strong> constantly learning <strong>the</strong><br />
traditional texts with <strong>the</strong> traditional methods was sublimated into <strong>the</strong><br />
scholarly movement <strong>of</strong> <strong>the</strong> Wissenschaft des Judentums.<br />
The picture changed dramatically in <strong>the</strong> middle <strong>of</strong> <strong>the</strong> nineteenth<br />
century, when <strong>the</strong> Haskala, in its third stage, reached <strong>the</strong> Eastern European<br />
Jewry, including Jews in Pol<strong>and</strong> <strong>and</strong> Lithuania (un<strong>der</strong> Russian<br />
government), Eastern Hungary, <strong>and</strong> Rumania. Here <strong>the</strong> Jewish population<br />
was far denser, whereas <strong>the</strong> surrounding society was far behind Western<br />
Europe in <strong>the</strong> process <strong>of</strong> <strong>the</strong> social <strong>and</strong> economic development. In fact,<br />
Jews would play an important role in <strong>the</strong> mo<strong>der</strong>nization <strong>of</strong> those areas.<br />
Therefore, several people <strong>of</strong> Jewish origin could take <strong>the</strong> initiative <strong>and</strong><br />
invent absolutely new alternatives to <strong>the</strong> social constructs that people had<br />
been living with so far.<br />
<strong>On</strong>e type <strong>of</strong> those social alternatives still preserved <strong>the</strong> idea <strong>of</strong> <strong>the</strong><br />
earlier Haskala according to which Jews should become <strong>and</strong> remain an<br />
organic part <strong>of</strong> <strong>the</strong> universal human culture. These alternatives proposed<br />
thus some forms <strong>of</strong> revolutionary change to <strong>the</strong> entire humankind, as was<br />
<strong>the</strong> case in <strong>the</strong> different types <strong>of</strong> socialist movements, in which Jews<br />
unquestionably played an important role. Esperantism also belongs here,<br />
for its fa<strong>the</strong>r, Ludwig Zamenh<strong>of</strong> was a Polish-Lithuanian Jew proposing an<br />
alternative to national language as ano<strong>the</strong>r social construct.<br />
The second type <strong>of</strong> radical answer that Eastern European Jews gave to<br />
<strong>the</strong> emergence <strong>of</strong> Enlightenment in <strong>the</strong> un<strong>der</strong>developed Eastern European<br />
milieu was creating a new kind <strong>of</strong> Jewish society. Recall that <strong>the</strong>re was a<br />
dense Jewish population living within a society that itself did not represent<br />
a mo<strong>der</strong>n model to which most Jews wished to acculturate. Different<br />
streams <strong>of</strong> this type <strong>of</strong> answer emerged, although <strong>the</strong>y did not mutually<br />
exclude each o<strong>the</strong>r. Many varieties <strong>of</strong> political activism, such as early<br />
forms <strong>of</strong> Zionism, political Zionism, territorialism or cultural autonomism,<br />
embody one level <strong>of</strong> creating an autonomous Jewish society.<br />
The birth <strong>of</strong> a new Jewish secular culture, including literature,<br />
newspapers or Klezmer music is ano<strong>the</strong>r one. The question <strong>the</strong>n arose<br />
whe<strong>the</strong>r <strong>the</strong> language <strong>of</strong> this new secular culture should be Yiddish - <strong>and</strong><br />
thus a st<strong>and</strong>ardized, literary version <strong>of</strong> Yiddish was to be developed - or<br />
Hebrew - <strong>and</strong> <strong>the</strong>refore a renewal <strong>of</strong> <strong>the</strong> Hebrew language was required. In<br />
<strong>the</strong> beginning, this point was not such an enormous matter <strong>of</strong> dispute as it<br />
would later develop into, when “Hebraists”, principally connected with
Weak Interactions 127<br />
Zionism, confronted “Yiddishists”, generally claiming a cultural <strong>and</strong> / or<br />
political autonomy within Eastern Europe. It is <strong>the</strong> irony <strong>of</strong> history that <strong>the</strong><br />
far more naïve <strong>and</strong> seemingly unrealistic ideology, calling for <strong>the</strong> revival <strong>of</strong><br />
an almost unspoken language in <strong>the</strong> distant Palestine, was <strong>the</strong> one that later<br />
would become reality.<br />
1.3. Language interactions in <strong>the</strong> Haskala<br />
Let us now return to our model <strong>of</strong> language interactions. As we have seen,<br />
<strong>the</strong> intensity <strong>of</strong> <strong>the</strong> interaction depends on <strong>the</strong> number <strong>of</strong> “exchange<br />
particles” - language changing individuals - , that is a kind <strong>of</strong> “distance”<br />
measured in <strong>the</strong> social network; fur<strong>the</strong>rmore on <strong>the</strong> “eligibility” <strong>of</strong> <strong>the</strong><br />
languages to transmit <strong>and</strong> to adopt features. We shall now confront this<br />
model with <strong>the</strong> linguistic reality <strong>of</strong> <strong>the</strong> different stages <strong>of</strong> <strong>the</strong> Haskala.<br />
Concerning <strong>the</strong> first stage, when only a h<strong>and</strong>ful <strong>of</strong> followers <strong>of</strong> Moses<br />
Mendelssohn rejected <strong>the</strong> Jüdischdeutsch <strong>and</strong> started speaking<br />
Hochdeutsch, our model will correctly predict that <strong>the</strong> number <strong>of</strong> exchange<br />
particles is insufficient to affect German in a perceptible way.<br />
The number <strong>of</strong> exchange particles increases dramatically when we reach<br />
<strong>the</strong> first half <strong>of</strong> <strong>the</strong> nineteenth century. However, <strong>the</strong> people changing<br />
language more or less consciously adopted <strong>the</strong> idea <strong>of</strong> <strong>the</strong>ir original idiom<br />
being an unclean <strong>and</strong> corrupt version <strong>of</strong> <strong>the</strong> target language. Consequently,<br />
by nature <strong>the</strong>ir language change consisted <strong>of</strong> not bringing any influence on<br />
<strong>the</strong> target language with <strong>the</strong>m. By applying our vague physical model to<br />
this situation, we might say that although <strong>the</strong> two languages were indeed<br />
close - from <strong>the</strong> viewpoints <strong>of</strong> geography, linguistic similarity <strong>and</strong> social<br />
contacts - , Hochdeutsch was not “eligible” enough to be seriously affected.<br />
What happened in <strong>the</strong> third stage <strong>of</strong> <strong>the</strong> Haskala? The following three<br />
case studies represent three possibilities. The first one, <strong>the</strong> influence <strong>of</strong><br />
Yiddish on Hungarian, was actually a case where some elements <strong>of</strong> stage 2<br />
Haskala were still present. The emancipation <strong>of</strong> <strong>the</strong> Jews was closely<br />
related to <strong>the</strong>ir assimilation into <strong>the</strong> Hungarian society, culture <strong>and</strong><br />
language. As Jews wished to become an equal part <strong>of</strong> that society, let us<br />
call this case type e. Each <strong>of</strong> <strong>the</strong> many people brings only a very “light”<br />
quantum <strong>of</strong> influence, similarly to <strong>the</strong> very little mass, if any, <strong>of</strong> <strong>the</strong><br />
electron neutrinos. The type mu designates a case when Jews migrated to a<br />
newly created Jewish “l<strong>and</strong>, language <strong>and</strong> culture”, namely to Mo<strong>der</strong>n<br />
Hebrew. Here less people carry possibly more “weight”, that is why <strong>the</strong>y
128 Tamás Bíró<br />
can be paralleled by <strong>the</strong> heavier muon neutrinos. In <strong>the</strong> third case, that is<br />
<strong>the</strong> birth <strong>of</strong> Esperanto, only one person <strong>of</strong> Jewish cultural background<br />
wished to transform <strong>the</strong> entire word, with a total rejection <strong>of</strong> reference to<br />
any form <strong>of</strong> Jewishness, at least on a conscious level (type tau, referring to<br />
<strong>the</strong> probably heaviest type <strong>of</strong> neutrinos).<br />
2. Three examples <strong>of</strong> weak interaction<br />
2.1. Type e: Yiddish <strong>and</strong> Hungarian<br />
Nineteenth century Hungary was situated on <strong>the</strong> bor<strong>der</strong> <strong>of</strong> Western<br />
European Jewry, affected already by <strong>the</strong> first two stages <strong>of</strong> Haskala, <strong>and</strong><br />
Eastern European Jewry, which would be reached only by its third phase.<br />
From <strong>the</strong> second half <strong>of</strong> <strong>the</strong> previous century onward, <strong>the</strong> Jewish<br />
immigration from Bohemia <strong>and</strong> Moravia had been importing a ra<strong>the</strong>r<br />
urbanized population speaking Western Yiddish, or even Jüdischdeutsch,<br />
whereas Eastern Yiddish speaking Galician Jews inhabiting Eastern<br />
Hungary represented <strong>the</strong> westernmost branch <strong>of</strong> Eastern European Jewry.<br />
Not only were <strong>the</strong> linguistic features <strong>of</strong> <strong>the</strong> two groups strikingly different,<br />
but also <strong>the</strong>ir social, economic <strong>and</strong> cultural background.<br />
In <strong>the</strong> social <strong>and</strong> economic fields, Hungary met a first wave <strong>of</strong><br />
mo<strong>der</strong>nization in <strong>the</strong> 1830s <strong>and</strong> 1840s, which is referred to as <strong>the</strong> reform<br />
age, reaching its peak in <strong>the</strong> 1848-49 revolution. After <strong>the</strong> so-called<br />
Compromise with Austria in 1867, <strong>the</strong> consequence <strong>of</strong> which had been <strong>the</strong><br />
creation <strong>of</strong> <strong>the</strong> Austro-Hungarian Empire with a dualistic system, <strong>the</strong> most<br />
urbanized parts <strong>of</strong> <strong>the</strong> country showed an especially remarkable economic<br />
<strong>and</strong> cultural growth.<br />
Parallel to <strong>the</strong> phenomenon <strong>of</strong> general mo<strong>der</strong>nization, <strong>the</strong> Jewish<br />
population un<strong>der</strong>went a similar process to <strong>the</strong> one we have already seen<br />
apropos <strong>of</strong> <strong>the</strong> French <strong>and</strong> German Jewry that had gone through <strong>the</strong>se<br />
social changes fifty years earlier. The second quarter <strong>of</strong> <strong>the</strong> century already<br />
witnesses a few Jewish thinkers, mainly rabbis arriving from Germany or<br />
Bohemia, <strong>and</strong> bringing mo<strong>der</strong>n ideals with <strong>the</strong>m. Yet, <strong>the</strong>ir effect cannot be<br />
perceived on a larger social scale before <strong>the</strong> last third <strong>of</strong> <strong>the</strong> century.<br />
A few differences should, however, be noted between German <strong>and</strong><br />
Hungarian Haskala. First, for <strong>the</strong> larger society into which Hungarian Jews
Weak Interactions 129<br />
wished to integrate, Enlightenment was not so much <strong>the</strong> consequence <strong>of</strong> <strong>the</strong><br />
Embourgeoisement, ra<strong>the</strong>r its catalyst. Enormous heterogeneities in <strong>the</strong><br />
degree <strong>of</strong> development could be found within <strong>the</strong> country, both in social, as<br />
well as economic terms. This general picture was paralleled with a<br />
heterogeneous distribution <strong>of</strong> Eastern <strong>and</strong> Western type <strong>of</strong> Jewry. Thus,<br />
even if <strong>the</strong> most Europeanized Jews may have wished, <strong>the</strong>y could not<br />
disown <strong>the</strong>ir pre-Haskala coreligionists living close to <strong>the</strong>m.<br />
Moreover, <strong>the</strong> mo<strong>der</strong>n Hungarian society <strong>and</strong> culture had to be created<br />
in spite <strong>of</strong> <strong>the</strong> Austrian occupation. Social constructs un<strong>der</strong>went huge<br />
changes, <strong>and</strong> any group <strong>of</strong> people identifying <strong>the</strong>mselves as Hungarian -<br />
<strong>and</strong> not Austrian - could influence <strong>the</strong> new shapes <strong>of</strong> society <strong>and</strong> culture.<br />
Immigrants from all directions played a fundamental role in laying down<br />
<strong>the</strong> bases <strong>of</strong> mo<strong>der</strong>n Hungarian urban culture. These are <strong>the</strong> circumstances<br />
un<strong>der</strong> which most <strong>of</strong> <strong>the</strong> Jews chose <strong>the</strong> Hungarian, ra<strong>the</strong>r than <strong>the</strong> German<br />
or Yiddish culture <strong>and</strong> language. This decision was far from being evident.<br />
Even most <strong>of</strong> <strong>the</strong> orthodoxy adopted Hungarian, though more slowly <strong>and</strong><br />
by keeping simultaneously Yiddish.<br />
By putting toge<strong>the</strong>r <strong>the</strong> pieces, we obtain an image in which <strong>the</strong><br />
dynamically changing Hungarian culture <strong>and</strong> society is searching new,<br />
mo<strong>der</strong>n forms, <strong>and</strong> is ready to integrate foreign influences - as long as <strong>the</strong><br />
carriers identify <strong>the</strong>mselves as new Hungarians. Fur<strong>the</strong>r, a major part <strong>of</strong> <strong>the</strong><br />
Jewish population is seeking its place in this new society, wants to adopt<br />
<strong>the</strong> new culture, but is still strongly connected - <strong>of</strong>ten against its will - to<br />
<strong>the</strong> pre-Haskala Jewry living not so far from <strong>the</strong>m. Consequently, we have<br />
both a high “eligibility” for being influenced on <strong>the</strong> part <strong>of</strong> <strong>the</strong> Hungarian<br />
language, <strong>and</strong> a large number <strong>of</strong> “exchange particles” flowing from Yiddish<br />
to Hungarian. 5<br />
What is <strong>the</strong> outcome <strong>of</strong> such a situation? Let us consi<strong>der</strong> a few examples<br />
<strong>of</strong> Yiddishisms in Hungarian. I shall distinguish between three registers that<br />
Yiddishisms entered consi<strong>der</strong>ably: <strong>the</strong> Jewish sociolect <strong>of</strong> Hungarian, argot<br />
(slang), <strong>and</strong> st<strong>and</strong>ard Hungarian.<br />
The vocabulary <strong>of</strong> Hungarian speaking Jews unsurprisingly includes a<br />
large number <strong>of</strong> words specific to domains <strong>of</strong> Jewish culture <strong>and</strong> religion.<br />
In some cases only phonological assimilation takes place. The<br />
Hungarian phonological system lacks a short /a/, <strong>and</strong> <strong>the</strong> short counterpart<br />
<strong>of</strong> /��� is /�/. Therefore <strong>the</strong> Yiddish word [��������] (‘Rosh Ha-shana, name<br />
<strong>of</strong> <strong>the</strong> Jewish New Year’, from Hebrew [��� ������], i.e. [���� �������� in<br />
st<strong>and</strong>ard Hungarian Ashkenazi pronunciation) becomes optionally<br />
[��������]. Although <strong>the</strong> original Yiddish pronunciation [��������] is still
130 Tamás Bíró<br />
possible, <strong>the</strong> latter emphasizes <strong>the</strong> foreign origin <strong>of</strong> <strong>the</strong> word. An analogous<br />
example is <strong>the</strong> word barchesz ([������] or [������], ‘chala, a special bread<br />
used on Shabbat <strong>and</strong> holidays’), which is clearly from Yiddish origin, but is<br />
unknown outside Hungary; it may have belonged to <strong>the</strong> vocabulary <strong>of</strong><br />
Hungarian Yiddish.<br />
O<strong>the</strong>r words immediately un<strong>der</strong>went Hungarian morphological<br />
processes. In fact, it is a well known phenomenon in many languages <strong>of</strong> <strong>the</strong><br />
world that borrowed verbs, unlike borrowed nouns, cannot be integrated<br />
directly into <strong>the</strong> vocabulary <strong>of</strong> a given language. This is <strong>the</strong> case in words<br />
like lejnol (‘to read <strong>the</strong> Torah-scroll in <strong>the</strong> synagogue’), lejnolás (‘<strong>the</strong><br />
reading <strong>of</strong> <strong>the</strong> Torah-scroll’) as well as snó<strong>der</strong> (‘money given as donation’),<br />
snó<strong>der</strong>ol (‘to donate money, especially after <strong>the</strong> public Torah-reading’),<br />
snó<strong>der</strong>olás (‘<strong>the</strong> act <strong>of</strong> money donation’). In <strong>the</strong> first case, <strong>the</strong> Yiddish verb<br />
leyenen (‘idem’) 6 was borrowed <strong>and</strong> one <strong>of</strong> <strong>the</strong> two most frequent<br />
denominal verbal suffixes, -l, was added. 7 The word lejnolás is <strong>the</strong> nomen<br />
actionis formed with <strong>the</strong> suffix -ás. The expression tfilint légol (‘to put on<br />
<strong>the</strong> phylacteries’) originates from German <strong>and</strong> Yiddish legen, <strong>and</strong> has gone<br />
through <strong>the</strong> same processes. For snó<strong>der</strong>ol, Hungarian borrows a Yiddish<br />
noun, 8 which <strong>the</strong>n serves as <strong>the</strong> base <strong>of</strong> fur<strong>the</strong>r <strong>der</strong>ivations.<br />
The Jewish sociolect <strong>of</strong> Hungarian includes fur<strong>the</strong>r lexical items, which<br />
do not belong to <strong>the</strong> domain <strong>of</strong> religious practice or Jewish culture. <strong>On</strong>e<br />
such word is unberufn (‘without calling [<strong>the</strong> devil]’), which should be<br />
added out <strong>of</strong> superstition to any positive statement that <strong>the</strong> speaker hopes to<br />
remain true in <strong>the</strong> future. For instance: ‘My child grows in beauty,<br />
unberufn’ (Blau-Láng, 1995:66). Nowadays, many people <strong>of</strong> <strong>the</strong> generation<br />
born after World War II <strong>and</strong> raised already in an almost non-Yiddish<br />
speaking milieu judge this expression as having nothing to do with<br />
superstition, but qualifying a situation as surprisingly good, like ‘You don’t<br />
say so! It’s incredible!’ <strong>and</strong> definitely including also some irony. 9 O<strong>the</strong>rs <strong>of</strong><br />
that generation say in <strong>the</strong> same surprising-ironic context: “My gr<strong>and</strong>ma<br />
would have said: unberufn…”, even if Gr<strong>and</strong>ma had used that word in a<br />
slightly different way. This second meaning <strong>of</strong> unberufn clearly lacks any<br />
reference to superstition, since <strong>the</strong> same people would use ano<strong>the</strong>r<br />
expression (lekopogom) to say ‘touch wood! knock on wood!’.<br />
Unlike <strong>the</strong> previous interjections, <strong>the</strong> adjective betámt (‘nice, intelligent,<br />
smart, sweet, lovely’) already enters <strong>the</strong> “real” syntax <strong>of</strong> <strong>the</strong> target<br />
language, even if morphological <strong>and</strong> phonological changes have not taken<br />
place yet - which happened in <strong>the</strong> case <strong>of</strong> lejnol <strong>and</strong> snó<strong>der</strong>ol. The word<br />
betámt consists <strong>of</strong> <strong>the</strong> Hebrew root taam (‘taste’), toge<strong>the</strong>r with <strong>the</strong>
Weak Interactions 131<br />
Germanic verbal prefix be- <strong>and</strong> past participle ending –t. The resulting<br />
word denotes a person who “has some taste”: somebody who has some<br />
characteristic traits, who is interesting, who has style <strong>and</strong> some sense <strong>of</strong><br />
humour, which is kind, polite, <strong>and</strong> so on. It is typically used by “Yiddishe<br />
mammes” describing <strong>the</strong> groom <strong>the</strong>y wish <strong>the</strong>ir daughter had.<br />
So far, we have seen examples where <strong>the</strong> language changing population<br />
has kept its original expression to denote something that could be best<br />
expressed using items <strong>of</strong> <strong>the</strong>ir old vocabulary. This Jewish sociolect has<br />
become an organic part <strong>of</strong> mo<strong>der</strong>n Hungarian, acknowledged, <strong>and</strong> partially<br />
known by many non-Jewish speakers, as well. But do we also find<br />
influences <strong>of</strong> Yiddish outside <strong>of</strong> <strong>the</strong> Jewish sociolect?<br />
The register that is <strong>the</strong> most likely to be affected un<strong>der</strong> such<br />
circumstances is probably always slang: it is non-conformist by definition,<br />
<strong>and</strong>, <strong>the</strong>refore, it is <strong>the</strong> least conservative. Slang is also <strong>the</strong> field where<br />
social norms, barriers <strong>and</strong> ol<strong>der</strong> prejudices play <strong>the</strong> least role. This may be<br />
<strong>the</strong> reason why Hungarian slang created in <strong>the</strong> nineteenth century borrowed<br />
so much from <strong>the</strong> languages <strong>of</strong> two socially marginal groups: <strong>the</strong> Gipsy<br />
(Roma) languages <strong>and</strong> Yiddish. In contemporary Hungarian slang, one can<br />
find well-known words from Yiddish origin such as: kóser (‘kosher’,<br />
meaning ‘good’ in slang); tré (‘bad, crappy, grotty’, from Hebrew-Yiddish-<br />
Hungarian tréfli ‘ritually unclean, non kosher food’); majré (‘fear, dread,<br />
rabbit fever’, from Hebrew mora ‘fear’ > Ashkenazi [�����] > Yiddish<br />
moyre [�����] > Hungarian [������]), fur<strong>the</strong>r <strong>der</strong>ived to majrézik (‘to fear,<br />
to be afraid <strong>of</strong> sg.’); szajré (‘swag, loot, hot stuff’, from Hebrew sehora<br />
������� �������������� ��� �� �� ������ et al., 1967-76). An interesting<br />
construction is stikában, meaning ‘in <strong>the</strong> sly, in secret, quitely’. Its origin is<br />
<strong>the</strong> Aramaic-Hebrew noun ������] ‘remaining silent’, which receives a<br />
Hungarian inessive case ending, meaning ‘in’.<br />
Through slang, some <strong>of</strong> <strong>the</strong> Yiddish words have <strong>the</strong>n infiltrated into <strong>the</strong><br />
st<strong>and</strong>ard language <strong>and</strong> become quasi-st<strong>and</strong>ard. Thus, <strong>the</strong> word haver - from<br />
<strong>the</strong> Hebrew ������] ‘friend’ - is used nowadays as an informal synonym for<br />
a ‘good acquaintance, a friend’. Similarly, dafke means in spoken<br />
Hungarian ‘For all that! <strong>On</strong>ly out <strong>of</strong> spite!’. Fur<strong>the</strong>rmore, <strong>the</strong>re are words<br />
<strong>of</strong> Yiddish origin which did not enter Hungarian through <strong>the</strong> slang, but<br />
through cultural interaction: macesz (‘matzo, unleavened bread’, from<br />
Hebrew matzot, plural form <strong>of</strong> matza; its ending clearly shows that <strong>the</strong><br />
word arrived to Hungarian through Yiddish) or sólet (‘tsholent’, a typically<br />
Hungarian Jewish bean dish, popular among non-Jews, too). 10
132 Tamás Bíró<br />
To summarize, <strong>the</strong> high amount <strong>of</strong> “exchange particles”, that is, Jewish<br />
people gradually changing <strong>the</strong>ir language from Yiddish to Hungarian, has<br />
affected <strong>the</strong> target language in three manners. <strong>On</strong>e <strong>of</strong> <strong>the</strong>m has been <strong>the</strong><br />
creation <strong>of</strong> a special Jewish sociolect. This was not a secret language<br />
though, <strong>and</strong> non-Jews have borrowed quite a few expressions. This fact led<br />
to <strong>the</strong> second manner <strong>of</strong> influence, namely to <strong>the</strong> high amount <strong>of</strong> Yiddish<br />
words entering <strong>the</strong> slang. Some <strong>of</strong> <strong>the</strong>se words have infiltrated even into <strong>the</strong><br />
relatively more informal registers <strong>of</strong> <strong>the</strong> st<strong>and</strong>ard language. The third<br />
manner is cultural interaction: <strong>the</strong> exchange <strong>of</strong> cultural goods - for instance<br />
in <strong>the</strong> field <strong>of</strong> gastronomy - inevitably has resulted <strong>the</strong> exchange <strong>of</strong> <strong>the</strong><br />
vocabulary designating those goods.<br />
2.2. Type µ: Yiddish <strong>and</strong> Mo<strong>der</strong>n Hebrew<br />
The fruit <strong>of</strong> Western European Haskala in <strong>the</strong> field <strong>of</strong> science was <strong>the</strong> birth<br />
<strong>of</strong> Wissenschaft des Judentums. The Jewish scholars belonging to this<br />
group aimed to introduce mo<strong>der</strong>n approaches when dealing with traditional<br />
texts, Jewish history, <strong>and</strong> so forth. Their approach contrasted traditional<br />
rabbinical activity <strong>the</strong> same way as <strong>the</strong> romanticist cantorial compositions<br />
by Salomon Sulzer <strong>and</strong> Louis Lew<strong>and</strong>owski contrasted traditional synagogal<br />
music: mo<strong>der</strong>nists aimed to produce cultural goods that were esteemed by<br />
<strong>the</strong> mo<strong>der</strong>n society, both by Jews <strong>and</strong> <strong>the</strong> recipient country. A fur<strong>the</strong>r<br />
motivation <strong>of</strong> <strong>the</strong> Wissenschaft des Judentums was to expose <strong>the</strong> values <strong>of</strong><br />
post-Biblical Jewish culture, <strong>and</strong> to present <strong>the</strong>m as an organic part <strong>of</strong><br />
universal culture: by emancipating Jewish past, <strong>the</strong>y hoped to be also<br />
emancipated by contemporary society.<br />
This background illuminates why early Haskala honored so much<br />
Hebrew - <strong>the</strong> language <strong>of</strong> <strong>the</strong> contribution par excellence <strong>of</strong> <strong>the</strong> Jewish<br />
nation to universal culture, which is <strong>the</strong> Hebrew Bible, <strong>and</strong> a language that<br />
had been long studied by Christian Hebraists. And also why Yiddish, <strong>the</strong><br />
supposedly jargon <strong>of</strong> <strong>the</strong> uneducated Jews <strong>and</strong> a corrupt version <strong>of</strong> German,<br />
was so much scorned in <strong>the</strong> same time.<br />
Although <strong>the</strong> goal <strong>of</strong> <strong>the</strong> earlier phases <strong>of</strong> Haskala was to promote <strong>the</strong><br />
literary language <strong>of</strong> <strong>the</strong> recipient country among Jews, that is practically<br />
Hochdeutsch, <strong>and</strong> Hebrew was principally only <strong>the</strong> object <strong>of</strong> scholarly<br />
study, still some attempts were made to use <strong>the</strong> language in mo<strong>der</strong>n<br />
domains, at least for some restricted purposes. After a few pioneering<br />
experiments to establish Hebrew newspapers in <strong>the</strong> middle <strong>of</strong> <strong>the</strong>
Weak Interactions 133<br />
eighteenth century, <strong>the</strong> Hebrew literary quarterly Ha-Meassef appeared as<br />
early as 1784 (Sáenz-Badillos, 1993:267).<br />
However, it was not before <strong>the</strong> middle <strong>of</strong> <strong>the</strong> next century, when<br />
Haskala reached Russia, that <strong>the</strong> need <strong>of</strong> reviving <strong>the</strong> Hebrew language was<br />
really articulated. As already discussed, <strong>the</strong> major reasons for this switch<br />
were that <strong>the</strong> Jewish population did not see <strong>the</strong> un<strong>der</strong>developed<br />
surrounding society as a model to which <strong>the</strong>y wanted to assimilate; <strong>the</strong><br />
Russian society <strong>and</strong> policy did not show any real sign <strong>of</strong> wanting to<br />
emancipate <strong>and</strong> integrate Jews, ei<strong>the</strong>r; fur<strong>the</strong>rmore, <strong>the</strong> huge Jewish<br />
population reached <strong>the</strong> critical mass required to develop something in itself.<br />
The summation <strong>of</strong> <strong>the</strong>se factors led to <strong>the</strong> idea <strong>of</strong> seeing Jewry as separate a<br />
nation in its mo<strong>der</strong>n sense. A fur<strong>the</strong>r factor reinforcing Jewish national<br />
feelings both in Eastern <strong>and</strong> Western Europe was <strong>the</strong> emergence <strong>of</strong> mo<strong>der</strong>n<br />
political anti-Semitism in <strong>the</strong> 1870s in <strong>the</strong> West, accompanied by events<br />
such as <strong>the</strong> huge Russian pogroms in 1881, <strong>the</strong> blood libel <strong>of</strong> Tiszaeszlár,<br />
Hungary (1882-3) or <strong>the</strong> Dreyfus-affair in France (starting in 1894).<br />
The claims following from this idea were that <strong>the</strong> Jewish nation has <strong>the</strong><br />
right to have a country - in Palestine or elsewhere, but at least it should<br />
receive some local autonomy - , <strong>and</strong> also that <strong>the</strong> Jewish nation must have<br />
its own national language. The two major c<strong>and</strong>idates for <strong>the</strong> Jewish<br />
national language were Yiddish <strong>and</strong> Hebrew, although German was not out<br />
<strong>of</strong> <strong>the</strong> competition, ei<strong>the</strong>r (cf. e.g. Shur, 1979:VII-VIII).<br />
The first wave <strong>of</strong> attempts to revive Hebrew consisted mainly <strong>of</strong> purists,<br />
seeing Biblical Hebrew as <strong>the</strong> most precious layer <strong>of</strong> <strong>the</strong> language: some <strong>of</strong><br />
<strong>the</strong>m went so far that <strong>the</strong>y preferred to create very complicated expressions<br />
to designate mo<strong>der</strong>n concepts, ra<strong>the</strong>r than using non-Biblical vocabulary.<br />
The fruits <strong>of</strong> this early period are among o<strong>the</strong>rs <strong>the</strong> first regular Hebrew<br />
weekly, Ha-Maggid (1856), <strong>the</strong> first mo<strong>der</strong>n play by D. Zamoscz (1851),<br />
novels by A. Mapu, as well as works <strong>of</strong> S. J. Abramowitsch (Mendele<br />
Moykher Seforim), who can be consi<strong>der</strong>ed one <strong>of</strong> <strong>the</strong> foun<strong>der</strong>s <strong>of</strong> both<br />
mo<strong>der</strong>n Hebrew <strong>and</strong> mo<strong>der</strong>n Yiddish literature.<br />
The real upswing was observable in <strong>the</strong> last quarter <strong>of</strong> <strong>the</strong> century,<br />
especially after <strong>the</strong> 1881 pogroms, <strong>and</strong> when Haskala had reached <strong>the</strong><br />
broadest masses, as well. Traditionally, <strong>the</strong> publication <strong>of</strong> Eliezer Ben-<br />
Yehuda’s article in 1879 entitled ‘A burning question’ is consi<strong>der</strong>ed to be<br />
<strong>the</strong> opening <strong>of</strong> <strong>the</strong> new era (Sáenz-Badillos, 1993:269). Ben-Yehuda (1858-<br />
1922) has been portrayed as <strong>the</strong> hero <strong>of</strong> <strong>the</strong> revival: he moved to Jerusalem<br />
in 1881, where he forced himself <strong>and</strong> his family to speak Hebrew. To speak<br />
a language, that is to produce everyday, spontaneous sentences “in real-
134 Tamás Bíró<br />
time”, on a language that had been mostly used for writing <strong>and</strong> reading <strong>and</strong><br />
only in restricted domains. His son, Ithamar (1882-1943), was <strong>the</strong> first<br />
person after millennia who grew up in an exclusively Hebrew-speaking<br />
environment. Ben-Yehuda constantly introduced new words designating<br />
weekday concepts, while he was editing a newspaper <strong>and</strong> working on his<br />
monumental Thesaurus, which incorporated material from ancient <strong>and</strong><br />
medieval literature. In 1890, he founded <strong>the</strong> Va’ad ha-Lashon (‘Language<br />
Committee’), <strong>the</strong> forerunner <strong>of</strong> <strong>the</strong> Hebrew Language Academy, hereby<br />
creating a quasi-<strong>of</strong>ficial institution for language planning.<br />
However, Shur (1979) has argued against an overestimation <strong>of</strong> Ben-<br />
Yehuda’s role. Out <strong>of</strong> Fishman’s five stages <strong>of</strong> language planning (in Shur,<br />
1979) (1. code selection; 2. ideologization <strong>of</strong> <strong>the</strong> choice; 3. codification; 4.<br />
elaboration <strong>and</strong> mo<strong>der</strong>nization; 5. st<strong>and</strong>ardization, i.e. <strong>the</strong> acceptance by <strong>the</strong><br />
community), Ben-Yehuda was salient especially in codification <strong>and</strong><br />
elaboration, as well as in vitalization, which was also necessary un<strong>der</strong> <strong>the</strong><br />
given circumstances. But for socio-political reasons, he had not much<br />
influence on <strong>the</strong> initial language choice <strong>and</strong> its ideologization, as well as on<br />
<strong>the</strong> final acceptance <strong>of</strong> <strong>the</strong> codified <strong>and</strong> elaborated st<strong>and</strong>ard.<br />
It is clear that Yiddish was <strong>the</strong> mo<strong>the</strong>r tongue, or one <strong>of</strong> <strong>the</strong> main<br />
languages for a major fraction <strong>of</strong> <strong>the</strong> members <strong>of</strong> <strong>the</strong> Va’ad ha-Lashon,<br />
including Ben-Yehuda himself. Moreover, people with Yiddish as first<br />
language represented an important part <strong>of</strong> <strong>the</strong> speaker community <strong>of</strong> <strong>the</strong><br />
old-new tongue in <strong>the</strong> first half <strong>of</strong> <strong>the</strong> twentieth century. Yiddish was not<br />
scorned anymore, as it had been a century before, but it was not consi<strong>der</strong>ed<br />
as a major source for language reform, ei<strong>the</strong>r. Especially for <strong>the</strong> later<br />
generations, Yiddish would symbolize <strong>the</strong> Diaspora left behind by <strong>the</strong><br />
Zionist movement.<br />
Yiddish speaking “exchange particles” dominated <strong>the</strong> community, much<br />
more than in <strong>the</strong> Hungarian case. Yet, a very conscious ideology required<br />
changing <strong>the</strong> previous ethnic language to <strong>the</strong> old-new national language,<br />
especially after <strong>the</strong> 1913-14 “Language Quarrel”, wherein <strong>the</strong> defen<strong>der</strong>s <strong>of</strong><br />
Hebrew defeated those <strong>of</strong> German <strong>and</strong> Yiddish (Shur, 1979:VII-VIII, X).<br />
This ideology was actively present in almost each <strong>and</strong> every individual who<br />
had chosen to move to <strong>the</strong> L<strong>and</strong> <strong>of</strong> Israel in a given period - contrary to <strong>the</strong><br />
European case, where ideology <strong>of</strong> changing <strong>the</strong> language was explicit only<br />
in <strong>the</strong> cultural elite. Fur<strong>the</strong>r, <strong>the</strong> language change was not slow <strong>and</strong> gradual,<br />
but drastic in <strong>the</strong> life <strong>of</strong> <strong>the</strong> people emigrating to Palestine, combined with a<br />
simultaneous radical change in geographical location, social structure <strong>and</strong><br />
lifestyle. What phenomena would this constellation involve?
Weak Interactions 135<br />
Yiddish influence on Mo<strong>der</strong>n Hebrew vocabulary has been investigated<br />
by - among o<strong>the</strong>rs - Haim Blanc. For instance, <strong>the</strong> Mo<strong>der</strong>n Hebrew<br />
interjection davka (approx. ’necessarily, for all that’) is clearly a<br />
Hebraisation <strong>of</strong> Yiddish dafke, <strong>of</strong> Hebrew origin itself, <strong>and</strong> mentioned also<br />
in relation with Hungarian. Similarly, kumzitz ‘get-toge<strong>the</strong>r, picnic,<br />
campfire’ undoubtedly originates from <strong>the</strong> Yiddish expression ‘come [<strong>and</strong>]<br />
sit down!’, since only in Yiddish do we find [u] in <strong>the</strong> verb ‘to come’.<br />
However, <strong>the</strong> expression was probably coined in Hebrew, as st<strong>and</strong>ard<br />
Yiddish dictionaries do not mention it. <strong>On</strong>e can easily imagine <strong>the</strong> early<br />
pioneers sitting around a campfire in <strong>the</strong> first kibbutzim, chatting in a<br />
mixture <strong>of</strong> Yiddish <strong>and</strong> Hebrew, <strong>and</strong> inviting <strong>the</strong>ir comrades to join <strong>the</strong>m.<br />
Nissan Netzer (1988) analyses <strong>the</strong> use <strong>of</strong> <strong>the</strong> Mo<strong>der</strong>n Hebrew verb<br />
firgen <strong>and</strong> <strong>the</strong> corresponding de-verbal noun firgun. Officially, <strong>the</strong> word is<br />
still not consi<strong>der</strong>ed to belong to <strong>the</strong> language, for it is not attested in any<br />
dictionary <strong>of</strong> Hebrew that I know. Definitions for this word I have found on<br />
<strong>the</strong> Internet are: “<strong>the</strong> ability to allow someone else to enjoy if his or her<br />
enjoyment does not hurt one,” <strong>and</strong> “to treat favorably, with equanimity, to<br />
bear no grudge or jealousy against somebody,” <strong>and</strong> also “to be delighted at<br />
<strong>the</strong> success <strong>of</strong> <strong>the</strong> o<strong>the</strong>r”. The word can be traced back to Yiddish farginen<br />
‘not begrudge, not envy, indulge’. As Netzer has demonstrated, <strong>the</strong>re is a<br />
linguistic gap in Hebrew, for <strong>the</strong> expressions darash et tovato shel… or lo<br />
hayta eno tsara be- that should bear that meaning are cumbersome,<br />
circuitous, overly sophisticated in style <strong>and</strong> seems to cloud <strong>the</strong> true<br />
linguistic message. Therefore, <strong>the</strong>y were not accepted by <strong>the</strong> linguistic<br />
community. When a leading Hebrew linguistics pr<strong>of</strong>essor used <strong>the</strong> Yiddish<br />
equivalent in <strong>the</strong> early sixties, <strong>the</strong> situation made <strong>the</strong> listeners <strong>of</strong> an<br />
academic lecture smile, because in that time <strong>the</strong> Yiddishism was consi<strong>der</strong>ed<br />
to be a folk idiom that would finally withdraw in favor <strong>of</strong> a “real Hebrew<br />
expression”. However, firgen would have become more <strong>and</strong> more accepted<br />
in daily conversation <strong>and</strong> even in journalistic writings by <strong>the</strong> eighties. 11<br />
This example has led us to <strong>the</strong> issue <strong>of</strong> <strong>the</strong> sociolinguistic status <strong>of</strong><br />
Yiddish words in Mo<strong>der</strong>n Hebrew. Ora Schwarzwald (1995) shows that <strong>the</strong><br />
vocabulary <strong>of</strong> <strong>the</strong> most used classical texts, such as <strong>the</strong> Hebrew Bible <strong>and</strong><br />
liturgy, has become <strong>the</strong> base <strong>of</strong> Mo<strong>der</strong>n Hebrew, in all its registers.<br />
Fur<strong>the</strong>rmore, loanwords <strong>of</strong> European languages are also used both in<br />
formal <strong>and</strong> non-formal language. However, from less esteemed languages,<br />
such as Jewish languages (e.g. Yiddish <strong>and</strong> Ladino), as well as Arabic,<br />
words would infiltrate primarily into lower registers <strong>and</strong> everyday informal<br />
speech.
136 Tamás Bíró<br />
For instance, chevre ‘friends’ is used mainly when addressing<br />
informally a group <strong>of</strong> people, <strong>and</strong> it is <strong>the</strong> borrowing <strong>of</strong> <strong>the</strong> similar word in<br />
Yiddish (khevre ‘gang, bunch <strong>of</strong> friends, society’). The latter obviously<br />
comes from Hebrew chevra ‘society, company, ga<strong>the</strong>ring’, whose root is<br />
chaver ‘friend’, a well-known word for speakers <strong>of</strong> Hungarian <strong>and</strong> Dutch<br />
(gabber), too. The originally Hebrew word thus arrived back to Mo<strong>der</strong>n<br />
Hebrew, but keeping <strong>the</strong> phonological traces <strong>of</strong> its trajectory. Also note <strong>the</strong><br />
minor shifts in <strong>the</strong> semantics during <strong>the</strong> two borrowings.<br />
Ano<strong>the</strong>r example for Yiddish influence on informal speech is <strong>the</strong> use <strong>of</strong><br />
<strong>the</strong> -le diminutive suffix: abale from aba ‘dad’, Sarale ‘little Sarah’,<br />
Chanale ‘little Hanah’, <strong>and</strong> so forth. Observe that <strong>the</strong> suffix follows <strong>the</strong><br />
Hebrew word, whereas in Yiddish one would have Sorele <strong>and</strong> Chanele<br />
expect.<br />
Thus, <strong>the</strong> influence <strong>of</strong> Yiddish on Mo<strong>der</strong>n Hebrew is indeed similar to<br />
its influence on Hungarian: lower registers <strong>and</strong> informal speech constitute<br />
one <strong>of</strong> <strong>the</strong> canals through which this interaction takes place. To make <strong>the</strong><br />
similarity even more prominent, we can point to two fur<strong>the</strong>r canals, shared<br />
by <strong>the</strong> Mo<strong>der</strong>n Hebrew case <strong>and</strong> <strong>the</strong> Hungarian case. Similarly to<br />
Hungarian, <strong>the</strong> designation <strong>of</strong> goods <strong>of</strong> general culture, such as food names<br />
(beygelach ‘bagels or pretzel’) represent a domain for word borrowings.<br />
Moreover, Yiddish loan words, or Hebrew words with a Yiddish or<br />
Ashkenazi pronunciation are likely to appear in religious vocabulary (e.g.<br />
rebe ‘Chasidic charismatic lea<strong>der</strong>’); typically in <strong>the</strong> sociolect <strong>of</strong> religious<br />
groups (especially within <strong>the</strong> ultra-orthodox society), <strong>and</strong> in <strong>the</strong> language<br />
used by secular Israelis to mock <strong>the</strong> stereotypically Yiddish-speaking ultraorthodox<br />
Jews (e.g. dos ‘an ultra-orthodox person’, from Hebrew dat<br />
‘religion’; vus-vus-im ‘<strong>the</strong> Ashkenazi ultra-orthodox Jews’, who <strong>of</strong>ten say<br />
Vus? Vus? ‘What? What?’ followed by <strong>the</strong> Hebrew plural ending -im).<br />
2.3. Type τ: Yiddish <strong>and</strong> Esperanto<br />
Esperanto emerged in <strong>the</strong> very same context as Mo<strong>der</strong>n Hebrew. Its creator,<br />
Lazar Ludwik Zamenh<strong>of</strong> (1859-1917), was born one year after Eliezer Ben-<br />
Yehuda, similarly from a Jewish family living in a small Lithuanian town,<br />
whose population was composed <strong>of</strong> Russian, Polish <strong>and</strong> Lithuanian people,<br />
but was dominated by a Jewish majority. The Litvak (Lithuanian-Jewish)<br />
Haskala background <strong>of</strong> both men encouraged traditional Jewish education<br />
combined with studies in a secular Gymnasium; both <strong>of</strong> <strong>the</strong>m went on to
Weak Interactions 137<br />
study medicine. Following <strong>the</strong> 1881 wave <strong>of</strong> pogroms, in <strong>the</strong> year in which<br />
Ben-Yehuda moved to Jerusalem, Zamenh<strong>of</strong> published an article calling for<br />
mass emigration to a Jewish homel<strong>and</strong>. For a few years, he became one <strong>of</strong><br />
<strong>the</strong> first activists <strong>of</strong> <strong>the</strong> early Zionist movement Hovevei Tzion (“Lovers <strong>of</strong><br />
Zion”). Berdichevsky (1986) points out <strong>the</strong> similarities even in <strong>the</strong><br />
mentality <strong>and</strong> <strong>the</strong> physical appearance <strong>of</strong> Zamenh<strong>of</strong> <strong>and</strong> Ben-Yehuda.<br />
Never<strong>the</strong>less, two key differences should be pointed out. The first one is<br />
Zamenh<strong>of</strong>’s pragmatism. In his 1881 article, Zamenh<strong>of</strong> imagined <strong>the</strong><br />
Jewish homel<strong>and</strong> to be in <strong>the</strong> western part <strong>of</strong> <strong>the</strong> United States, a relatively<br />
unsettled area those days, which would have arisen much less sensibility<br />
from all sides. Fur<strong>the</strong>rmore, Zamenh<strong>of</strong> shared <strong>the</strong> skepticism <strong>of</strong> many <strong>of</strong><br />
his contemporaries in <strong>the</strong> feasibility to revive <strong>the</strong> Hebrew language.<br />
According to <strong>the</strong> anecdote, Theodor Herzl said once that he could not buy<br />
even a train ticket in Hebrew. Leading Jewish writers, such as Mendele<br />
Moykher Seforim, oscillated between writing in Yiddish <strong>and</strong> in Hebrew;<br />
both <strong>of</strong> <strong>the</strong>se languages called for <strong>the</strong> establishment <strong>of</strong> a mo<strong>der</strong>n, secular<br />
literary tongue. The young <strong>and</strong> pragmatic Zamenh<strong>of</strong> chose to reform<br />
Yiddish, <strong>the</strong> language with millions <strong>of</strong> native speakers; whereas <strong>the</strong> first<br />
native speaker <strong>of</strong> Mo<strong>der</strong>n Hebrew, <strong>the</strong> son <strong>of</strong> Ben-Yehuda was not born<br />
yet.<br />
In his early years, Zamenh<strong>of</strong> wrote a comprehensive Yiddish grammar<br />
(completed in 1879, partially published in 1909 in <strong>the</strong> Vilna Journal, Lebn<br />
un Vissenschaft, <strong>and</strong> fully published only in 1982). He argued for <strong>the</strong><br />
mo<strong>der</strong>nization <strong>of</strong> <strong>the</strong> language <strong>and</strong> fought for <strong>the</strong> use <strong>of</strong> <strong>the</strong> Latin alphabet,<br />
instead <strong>of</strong> <strong>the</strong> Hebrew one. How is it possible <strong>the</strong>n that a few years later<br />
Zamenh<strong>of</strong> changed his mind, <strong>and</strong> switched to Esperanto (1887)?<br />
Here comes <strong>the</strong> second key difference into <strong>the</strong> picture. Ben-Yehuda was<br />
sent by his orthodox family to a yeshiva (traditional school teaching mainly<br />
<strong>the</strong> Talmud), where one <strong>of</strong> <strong>the</strong> rabbis introduced him secretly into <strong>the</strong><br />
revolutionary ideas <strong>of</strong> <strong>the</strong> Haskala. <strong>On</strong> <strong>the</strong> contrary, Zamenh<strong>of</strong>’s fa<strong>the</strong>r <strong>and</strong><br />
gr<strong>and</strong>fa<strong>the</strong>r were enlightened high-school teachers <strong>of</strong> Western languages<br />
(French <strong>and</strong> German). For him, being Jewish probably meant a universal<br />
mission to make <strong>the</strong> world a better place for <strong>the</strong> whole humankind. This<br />
idea originates from eighteenth century German Haskala philosophers<br />
claiming that Judaism is <strong>the</strong> purest embodiment so far existing <strong>of</strong> <strong>the</strong><br />
universal moral <strong>and</strong> <strong>of</strong> <strong>the</strong> faith <strong>of</strong> <strong>the</strong> Pure Reason; even today a major part<br />
<strong>of</strong> Jews worldwide perceive Judaism this way.<br />
Zamenh<strong>of</strong> did not <strong>the</strong>refore content himself with <strong>the</strong> goal <strong>of</strong> creating a<br />
Jewish national language. For him, similarly to his semi-secularized
138 Tamás Bíró<br />
coreligionists joining <strong>the</strong> socialist movement in <strong>the</strong> same decades, unifying<br />
<strong>the</strong> human race <strong>and</strong> building a new word or<strong>der</strong> presented <strong>the</strong> solution for -<br />
among o<strong>the</strong>rs - <strong>the</strong> problems <strong>of</strong> <strong>the</strong> oppressed Eastern European Jewry.<br />
And also <strong>the</strong> o<strong>the</strong>r way around: <strong>the</strong> secular messianic idea <strong>of</strong> <strong>the</strong> unification<br />
<strong>of</strong> <strong>the</strong> dispersed <strong>and</strong> oppressed Jews into a Jewish nation was just one step<br />
behind from <strong>the</strong> secular messianic idea <strong>of</strong> <strong>the</strong> unification <strong>of</strong> <strong>the</strong> whole<br />
mankind into a supra-national unit. This explains not only <strong>the</strong> motivations<br />
<strong>of</strong> Zamenh<strong>of</strong> himself, but also why Jews played such an important role in<br />
<strong>the</strong> pre-World War II Esperanto movement in Central <strong>and</strong> Eastern Europe<br />
(Berdichevsky, 1986:60). Whereas socialists fought for a social-economic<br />
liberation <strong>of</strong> <strong>the</strong> oppressed, Zamenh<strong>of</strong> spoke about <strong>the</strong> liberation <strong>of</strong> <strong>the</strong><br />
humans from <strong>the</strong> cultural <strong>and</strong> linguistic barriers. It is not a coincidence that<br />
<strong>the</strong> twentieth century history <strong>of</strong> <strong>the</strong> Esperantist movement was so much<br />
intermingled with <strong>the</strong> one <strong>of</strong> <strong>the</strong> socialist movements.<br />
Zamenh<strong>of</strong>’s initiative was to create a language that would be equally<br />
distant from <strong>and</strong> equally close to each ethnic language, thus each human<br />
being would have equal chance using this bridge connecting cultures <strong>and</strong><br />
people. Hence Zamenh<strong>of</strong> created a vocabulary <strong>and</strong> a grammar using<br />
elements <strong>of</strong> languages he knew: Russian (<strong>the</strong> language his fa<strong>the</strong>r spoke<br />
home <strong>and</strong> <strong>the</strong> language <strong>of</strong> his highschool), German <strong>and</strong> French (<strong>the</strong><br />
languages his fa<strong>the</strong>r <strong>and</strong> gr<strong>and</strong>fa<strong>the</strong>r were teachers <strong>of</strong>), Polish (<strong>the</strong> language<br />
<strong>of</strong> his non-Jewish fellow children), Latin <strong>and</strong> Greek (from highschool), as<br />
well as English <strong>and</strong> Italian. Note that <strong>the</strong> resulting language, similarly to<br />
most artificial languages, is inherently European <strong>and</strong> Indo-European in its<br />
character, though extremely simplified.<br />
However, one should not forget that Zamenh<strong>of</strong>’s native tongue was<br />
Yiddish, this was <strong>the</strong> language he used with his school mates in <strong>the</strong> Jewish<br />
primary school (khey<strong>der</strong>, cf. Piron, 1984), <strong>and</strong> most <strong>of</strong> his life he kept<br />
contact with circles where Yiddish was alive. So one would won<strong>der</strong> why<br />
Yiddish is not mentioned overtly among <strong>the</strong> source languages <strong>of</strong> Esperanto.<br />
Seeing Zamenh<strong>of</strong>’s former devotion for <strong>the</strong> Jewish sake <strong>and</strong> <strong>the</strong> Yiddish<br />
language, as well as his later remark that Yiddish is a language similar to<br />
any o<strong>the</strong>r (in Homo Sum (1901), cf. Piron (1984:17) <strong>and</strong> Berdichevsky<br />
(1986:70)), <strong>the</strong> possibility that he despised “<strong>the</strong> corrupt version <strong>of</strong> German”<br />
or that he felt shame at his Yiddish origins, are out <strong>of</strong> question.<br />
The challenging task now is to find at least covert influences <strong>of</strong> Yiddish<br />
on Esperanto.<br />
As strange as it may sound, a consi<strong>der</strong>able literature has been devoted to<br />
etymology within Esperanto linguistics. <strong>On</strong>e <strong>of</strong> <strong>the</strong> biggest mysteries is <strong>the</strong>
Weak Interactions 139<br />
morpheme edz. As a root, it means ‘married person’ (edzo ‘husb<strong>and</strong>’;<br />
edzino ‘wife’, by adding <strong>the</strong> feminine suffix -in-). While as a suffix, it turns<br />
<strong>the</strong> word’s meaning into <strong>the</strong> wife or husb<strong>and</strong> <strong>of</strong> <strong>the</strong> stem: lavistino ’washerwoman’<br />
vs. lavistinedzo ‘washerwoman’s housb<strong>and</strong>’; doktoro ‘doctor’ vs.<br />
doktoredzino ‘doctor’s wife’. Hungarian Esperantists have tried to use this<br />
suffix to translate <strong>the</strong> Hungarian suffix -né (‘wife <strong>of</strong>…’, e.g.: Deákné ‘wife<br />
<strong>of</strong> Deák, Mrs. Deák’; cf. Goldin (1982:28)). The phonemic content <strong>of</strong> <strong>the</strong><br />
morpheme is not similar to any word with related meaning in any <strong>of</strong> <strong>the</strong><br />
languages that Zamenh<strong>of</strong> might have taken into consi<strong>der</strong>ation.<br />
Zamenh<strong>of</strong> himself wrote in a letter to Émile Boirac that <strong>the</strong> morpheme<br />
was <strong>the</strong> result <strong>of</strong> backformation, <strong>and</strong> that originally it was a bound form<br />
(Goldin, 1982:22f). Boirac suggested in 1913 <strong>the</strong> following reconstruction:<br />
if <strong>the</strong> German Kronprinz (‘heir apparent’) became kronprinco in Esperanto,<br />
while Kronprinzessin (‘wife <strong>of</strong> a crown prince’, note <strong>the</strong> double feminine<br />
ending: <strong>the</strong> French feminine suffix -esse is followed by <strong>the</strong> Germanic<br />
feminine -in) turns to kronprincedzino, <strong>the</strong>n <strong>the</strong> ending -edzin- can be<br />
identified as ‘a woman legally bound to a man’. By removing <strong>the</strong> feminine<br />
suffix -in-, we obtain <strong>the</strong> morpheme -edz-. Goldin adds to this <strong>the</strong>ory that<br />
<strong>the</strong> morphemes es <strong>and</strong> ec had already been used with o<strong>the</strong>r meanings, that<br />
is why <strong>the</strong> surprising [��] combination appeared. Summarizing, <strong>the</strong><br />
etymology <strong>of</strong> <strong>the</strong> Esperanto morpheme edz would be <strong>the</strong> French feminine<br />
ending -esse, which had been reanalyzed with a different meaning due to<br />
<strong>the</strong> additional feminine suffix in German.<br />
However, this is not <strong>the</strong> end <strong>of</strong> <strong>the</strong> story. O<strong>the</strong>r alternatives have been also<br />
proposed. Waringhien <strong>and</strong> o<strong>the</strong>rs (in Goldin, 1982) have brought forward<br />
<strong>the</strong> idea that <strong>the</strong> word serving as <strong>the</strong> base <strong>of</strong> backformation was <strong>the</strong> Yiddish<br />
word rebetsin (‘wife <strong>of</strong> a rabbi’). In fact, this word can be reanalyzed as<br />
reb+edz+in, <strong>and</strong> we obtain <strong>the</strong> edz morpheme using <strong>the</strong> same logic as<br />
above. Goldin’s counterargument that <strong>the</strong> Yiddish word is actually rebetsn<br />
with a syllabic [���] is not at all convincing: old Yiddish spelling <strong>of</strong>ten uses<br />
<strong>the</strong> letter yod to designate a schwa, or even more <strong>the</strong> syllabicity <strong>of</strong> [�� an �],<br />
similarly to <strong>the</strong> in German spelling, like in wissen. Consequently, I can<br />
indeed accept <strong>the</strong> idea that a pre-YIVO spelling rebetsin was in <strong>the</strong> mind <strong>of</strong><br />
Zamenh<strong>of</strong>.<br />
Piron (1984) adds fur<strong>the</strong>r cases <strong>of</strong> possible Yiddish influence. In words<br />
taken from German, <strong>the</strong> affricate [���] always changes to [�]: German pfeifen<br />
‘to whistle’ became Esperanto fajfi. This coincides with Yiddish fayfn.<br />
Though, one is not compelled to point to Yiddish as <strong>the</strong> origin <strong>of</strong> this word:<br />
<strong>the</strong> reason can simply be that <strong>the</strong> affricate [���] is too typical to German, not
140 Tamás Bíró<br />
occurring in any o<strong>the</strong>r languages that served “<strong>of</strong>ficially” as examples for<br />
Zamenh<strong>of</strong>. In o<strong>the</strong>r words, [���] was not seen as universal enough. But what<br />
about <strong>the</strong> consonant clusters ���], ���], ���], which are also characteristic<br />
solely to German (<strong>and</strong> to Yiddish)? May <strong>the</strong> solution be that while [���]<br />
becomes [�] in Yiddish, <strong>the</strong>se clusters are unchanged; <strong>the</strong>refore, Zamenh<strong>of</strong><br />
felt less discomfort with regard to <strong>the</strong> latter clusters than with regard to [���]<br />
which truly occurs exclusively in German? I do not believe that we can do<br />
more than speculate about <strong>the</strong> different unconscious factors acting within a<br />
person more than a hundred years ago. The only claim we can make is that<br />
some <strong>of</strong> <strong>the</strong>se factors must have been related to Yiddish, as expected from<br />
<strong>the</strong> fact that Yiddish was one <strong>of</strong> <strong>the</strong> major tongues <strong>of</strong> Zamenh<strong>of</strong>.<br />
In <strong>the</strong> field <strong>of</strong> semantics, Piron brings <strong>the</strong> differentiation in Esperanto<br />
between l<strong>and</strong>a (‘national, related to a given country’, adjective formed<br />
from l<strong>and</strong>o ‘country’) as opposed to nacia (‘national, related to a given<br />
nation’, adjective from nacio ‘nation’). This differentiation exists in<br />
Yiddish (l<strong>and</strong>ish <strong>and</strong> natsional), but not in any o<strong>the</strong>r languages that<br />
Zamenh<strong>of</strong> might have taken into consi<strong>der</strong>ation. Piron also argues against<br />
<strong>the</strong> possible claim that this is not a Yiddish influence, ra<strong>the</strong>r an inner<br />
development related to <strong>the</strong> inner logic <strong>of</strong> Esperanto.<br />
The most evident example <strong>of</strong> Piron is Esperanto superjaro ‘leap year’, a<br />
compound <strong>of</strong> super ‘on’ <strong>and</strong> jaro ‘year’. No known language uses <strong>the</strong><br />
preposition on or above to express this concept. However, Yiddish has<br />
iberyor for ‘leap year’, from Hebrew ibbur (‘making pregnant’), <strong>the</strong> term<br />
used in rabbinic literature for intercalating an extra month <strong>and</strong> making <strong>the</strong><br />
year a leap year (e.g. Tosefta Sanhedrin 2:1-7). <strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, iber also<br />
means ‘above’ in Yiddish, which explains <strong>the</strong> strange expression in<br />
Esperanto. I do not know if Zamenh<strong>of</strong> realized that <strong>the</strong> Yiddish expression<br />
iberyor is not related to German über, but this is probably not relevant.<br />
Let us summarize this section. Yiddish influence on Esperanto is a case<br />
where <strong>the</strong>re is only one exchange particle - in <strong>the</strong> first or<strong>der</strong> approximation,<br />
at least, since we have not dealt with <strong>the</strong> possible influences related to <strong>the</strong><br />
numerous later speakers <strong>of</strong> Esperanto <strong>of</strong> Yiddish background. Though, this<br />
one particle had a huge impact on <strong>the</strong> language for a very obvious reason.<br />
Even if he did not overtly acknowledge that Yiddish had played a role in<br />
creating Esperanto, it is possible to discover <strong>the</strong> - ei<strong>the</strong>r consciously hidden<br />
or unconscious - traces <strong>of</strong> Yiddish.<br />
Did Zamenh<strong>of</strong> want to deny that he had also used Yiddish, as a building<br />
block <strong>of</strong> Esperanto? Perhaps because his goal was indeed to create a<br />
universal, supra-national language, <strong>and</strong> not <strong>the</strong> language <strong>of</strong> <strong>the</strong> Jewish
Weak Interactions 141<br />
nation? Or, alternatively, was this influence unconscious? I do not dare to<br />
give an answer.<br />
3. Conclusion<br />
In linguistics, we could define weak interaction as an interaction that is not<br />
overtly acknowledged. No one would deny <strong>the</strong> influence <strong>of</strong> <strong>the</strong> Frenchspeaking<br />
ruling class on medieval English, or <strong>the</strong> impact <strong>of</strong> <strong>the</strong> Slavic<br />
neighbors on Hungarian. But sometimes, conscious factors hide <strong>the</strong> effect.<br />
Yet, weak interactions are as crucial for <strong>the</strong> development <strong>of</strong> a language, as<br />
<strong>the</strong> nuclear processes emitting neutrinos in <strong>the</strong> core <strong>of</strong> <strong>the</strong> Sun that produce<br />
<strong>the</strong> energy which is vital for us.<br />
We have seen three cases <strong>of</strong> weak interaction between languages. In<br />
fact, all three stories were about <strong>the</strong> formative phase <strong>of</strong> a new or<br />
mo<strong>der</strong>nized language, in <strong>the</strong> midst <strong>of</strong> <strong>the</strong> late nineteenth century Eastern<br />
Europe Jewry. In <strong>the</strong> cases <strong>of</strong> Yiddish influencing Hungarian <strong>and</strong> Mo<strong>der</strong>n<br />
Hebrew, <strong>the</strong> number <strong>of</strong> “exchange particles”, that is, <strong>the</strong> amount <strong>of</strong> initially<br />
Yiddish-speaking people joining <strong>the</strong> new language community, were<br />
extremely high: roughly one tenth <strong>of</strong> <strong>the</strong> Hungarian speaking population in<br />
nineteenth century Hungary, <strong>and</strong> probably above 50% <strong>of</strong> <strong>the</strong> Jews living in<br />
early twentieth century Palestine. None<strong>the</strong>less, in both cases we encounter<br />
an ideology promoting <strong>the</strong> new language <strong>and</strong> disfavoring Yiddish.<br />
Because <strong>the</strong> level <strong>of</strong> consciousness <strong>of</strong> this ideology seems to be<br />
inversely proportional to <strong>the</strong> ratio <strong>of</strong> “exchange particles” - stronger in<br />
Palestine than in Hungary - , <strong>the</strong> two factors extinguish each o<strong>the</strong>r, <strong>and</strong> we<br />
find similar phenomena. For instance, Yiddish has affected first <strong>and</strong><br />
foremost lower registers, which are less censored by society; <strong>the</strong>refrom it<br />
infiltrates into informal st<strong>and</strong>ard language. Additional trends are Yiddish<br />
words entering specific domains, such as gastronomy or Jewish religious<br />
practice. Although it is essential to note that not all concepts that are new in<br />
<strong>the</strong> target culture are expressed by <strong>the</strong>ir original Yiddish word: many new<br />
expressions in <strong>the</strong>se domains have been coined in Hungarian <strong>and</strong> Mo<strong>der</strong>n<br />
Hebrew, <strong>and</strong> accepted by <strong>the</strong> language community.<br />
The third case that we have examined is different. Zamenh<strong>of</strong> was a<br />
single person, but as <strong>the</strong> creator <strong>of</strong> Esperanto, he had an enormous<br />
influence on <strong>the</strong> new language. The influence <strong>of</strong> Yiddish was again weak in<br />
<strong>the</strong> sense that it was not overtly admitted; however, we could present<br />
examples where <strong>the</strong> native tongue <strong>of</strong> Zamenh<strong>of</strong> influenced <strong>the</strong> new
142 Tamás Bíró<br />
language. We could have cited, as <strong>the</strong> articles mentioned had done,<br />
numerous fur<strong>the</strong>r instances where <strong>the</strong> influence <strong>of</strong> Yiddish cannot be<br />
proven directly, <strong>the</strong> given phenomenon could have been taken from o<strong>the</strong>r<br />
languages, as well; however, one can hypo<strong>the</strong>size that Yiddish played -<br />
consciously or unconsciously - a reinforcing role in Zamenh<strong>of</strong>’s decisions.<br />
I do hope that I have been able to prove to <strong>the</strong> rea<strong>der</strong> that seemingly<br />
very remote fields, such as physics, social history <strong>and</strong> linguistics, can be<br />
interconnected, at least for <strong>the</strong> sake <strong>of</strong> a thought experiment. Fur<strong>the</strong>rmore,<br />
“exchange particles” in <strong>the</strong> field <strong>of</strong> science, <strong>and</strong> Tjeerd is certainly among<br />
<strong>the</strong>m, have hopefully brought at least some weak interaction among <strong>the</strong><br />
different disciplines.<br />
Notes<br />
1 According to http://cupp.oulu.fi/neutrino/nd-mass.html, <strong>the</strong> mass <strong>of</strong> <strong>the</strong><br />
electron neutrino (νe) is less than 2.2 eV, <strong>the</strong> mass <strong>of</strong> <strong>the</strong> muon neutrino (νµ)<br />
does not exceed 170 keV, while <strong>the</strong> mass <strong>of</strong> <strong>the</strong> tau neutrino (νπ) is reported to<br />
be bellow 15.5 MeV. For <strong>the</strong> sake <strong>of</strong> comparison, <strong>the</strong> mass <strong>of</strong> an electron is 511<br />
keV, while <strong>the</strong> mass <strong>of</strong> a proton is almost 940 MeV.<br />
2 Physical phenomena are thought to be reducible to four fundamental forces.<br />
These are gravity, electromagnetism, weak interaction <strong>and</strong> strong interaction.<br />
The last two play a role in sub-atomic physics.<br />
3 The photons (particles <strong>of</strong> <strong>the</strong> light) are <strong>the</strong> exchange particles for <strong>the</strong><br />
electromagnetic interaction; <strong>the</strong> hypo<strong>the</strong>tical gravitons should transmit gravitation;<br />
in <strong>the</strong> case <strong>of</strong> <strong>the</strong> weak interaction, <strong>the</strong> W + , W - <strong>and</strong> Z vector bosons play<br />
that role; whereas <strong>the</strong> strong interaction is mediated by pions.<br />
4 Targumim (plural <strong>of</strong> targum) are <strong>the</strong> Jewish Aramaic versions <strong>of</strong> <strong>the</strong> Hebrew<br />
Bible from <strong>the</strong> late antiquity, including also many commentaries beside <strong>the</strong><br />
pure translation. The same way as late antiquity Jews created <strong>the</strong> commented<br />
translation <strong>of</strong> <strong>the</strong> Holy Scriptures to <strong>the</strong>ir native tongue <strong>and</strong> using <strong>the</strong>ir way <strong>of</strong><br />
thinking, Moses Mendelssohn expected his version <strong>of</strong> <strong>the</strong> Bible to fit <strong>the</strong><br />
mo<strong>der</strong>n way <strong>of</strong> thinking <strong>and</strong> <strong>the</strong> “correct language” <strong>of</strong> its future rea<strong>der</strong>s.<br />
Obviously, <strong>the</strong> Biur should first have to fulfil its previous task, namely to teach<br />
<strong>the</strong> mo<strong>der</strong>n way <strong>of</strong> thinking <strong>and</strong> <strong>the</strong> “correct tongue” to <strong>the</strong> first generation <strong>of</strong><br />
its rea<strong>der</strong>s. Interestingly enough, script was not such a major issue for Mendelssohn<br />
as “language purity”, thus he wrote Hochdeutsch in Hebrew characters; in<br />
or<strong>der</strong> to better disseminate his work among <strong>the</strong> Jewish population.<br />
5 I assume that <strong>the</strong> formative phase <strong>of</strong> mo<strong>der</strong>n Dutch society <strong>and</strong> culture in <strong>the</strong><br />
17 th <strong>and</strong> 18 th century is comparable to that <strong>of</strong> 19 th century Hungary; even more
Weak Interactions 143<br />
is so <strong>the</strong> role <strong>of</strong> Jewry in both countries, as a group which was simultaneously<br />
integrating into <strong>the</strong> new society <strong>and</strong> also forming it. In both cases, <strong>the</strong> presence<br />
<strong>of</strong> <strong>the</strong> continuous spectrum from <strong>the</strong> pre-Haskala Yid to <strong>the</strong> self-mo<strong>der</strong>nizing<br />
Israelite led to a gradual, though determined giving up <strong>of</strong> <strong>the</strong> Yiddish language.<br />
This socio-historical parallelism could partially explain why phenomena <strong>of</strong><br />
Yiddish influence on Dutch are <strong>of</strong>ten similar to that on Hungarian.<br />
Concerning Dutch-Jewish linguistic interactions, rea<strong>der</strong>s interested in Jewish<br />
aspects <strong>of</strong> Papiamentu, a creole language spoken in <strong>the</strong> Ne<strong>the</strong>rl<strong>and</strong>s Antilles,<br />
are referred to Richard E. Wood’s article in Jewish Language Review 3<br />
(1983):15-18.<br />
6 The etymology <strong>of</strong> <strong>the</strong> Yiddish word itself is also interesting. The origin is <strong>the</strong><br />
late Latin or Old French root [���] ‘to read’ (cf. to Latin lego, legere, mo<strong>der</strong>n<br />
French je lis, lire), which was borrowed by <strong>the</strong> Jews living in early medieval<br />
Western Europe. The latter would <strong>the</strong>n change <strong>the</strong>ir language to Old High<br />
German, <strong>the</strong> ancestor <strong>of</strong> Yiddish. At some point, <strong>the</strong> meaning <strong>of</strong> <strong>the</strong> Old French<br />
word was restricted to <strong>the</strong> public reading <strong>of</strong> <strong>the</strong> Torah-scroll in <strong>the</strong> synagogue.<br />
7 Compare to sí ‘ski’ > síel ‘to ski’, ��� ‘fire’ > tüzel ‘to fire’; also: printel ‘to<br />
print with a computer printer’. It is extremely surprising that <strong>the</strong> word lejnol<br />
does not follow vowel harmony, one would expect * lejnel. Even though <strong>the</strong> [�]<br />
sound can be transparent for vowel harmony, this fact is not enough to explain<br />
<strong>the</strong> word lejnol. Probably <strong>the</strong> dialectal Yiddish laynen was originally borrowed,<br />
<strong>and</strong> this form served as <strong>the</strong> base for word formation, before <strong>the</strong> <strong>of</strong>ficial Yiddish<br />
form leynen influenced <strong>the</strong> Hungarian word. Some people still say lájnol.<br />
8 When being called to <strong>the</strong> Torah during <strong>the</strong> public reading, one recites a<br />
blessing, <strong>the</strong> text <strong>of</strong> which says: “He Who blessed our forefa<strong>the</strong>rs Abraham,<br />
Isaac <strong>and</strong> Jacob, may He bless [<strong>the</strong> name <strong>of</strong> <strong>the</strong> person] because he has come up<br />
to <strong>the</strong> Torah / who has promised to contribute to charity on behalf <strong>of</strong>… etc.”<br />
The part <strong>of</strong> <strong>the</strong> text ‘who has promised’ sounds in <strong>the</strong> Ashkenazi pronunciation<br />
[�� �������]. This is most probably <strong>the</strong> source <strong>of</strong> <strong>the</strong> word snó<strong>der</strong>, after vowel<br />
in <strong>the</strong> unstressed last syllable has become a schwa, a process that is crucial for<br />
un<strong>der</strong>st<strong>and</strong>ing <strong>the</strong> Yiddishization <strong>of</strong> Hebrew words. The exciting part <strong>of</strong> <strong>the</strong><br />
story is that <strong>the</strong> proclitic [��] (‘that’) was kept toge<strong>the</strong>r with <strong>the</strong> following finite<br />
verbal form ([�����] ‘he promised’), <strong>and</strong> <strong>the</strong>y were reanalysed as one word.<br />
9 When I asked people about <strong>the</strong> meaning <strong>of</strong> unberufn on <strong>the</strong> mailing list 2nd-<br />
Generation-Jews-Hungary@yahoogroups.com, somebody reported that her<br />
non-Jewish gr<strong>and</strong>mo<strong>the</strong>r also used to say unberufn with a similar meaning.<br />
10 O<strong>the</strong>r Hungarian words <strong>of</strong> Hebrew origin do not come from Yiddish, as shown<br />
by <strong>the</strong>ir non-Askenazi pronunciation: Tóra ([�����] ‘Torah’, as opposed to its<br />
Yiddish counterpart Toyre) or rabbi (<strong>and</strong> not rov or rebe). Words like behemót<br />
(‘big hulking fellow’), originally from Biblical Hebrew behema (‘cattle’, plural:
144 Tamás Bíró<br />
behemot; appearing also as a proper name both in Jewish <strong>and</strong> in Christian<br />
mythology) should be ra<strong>the</strong>r traced back to Christian Biblical tradition.<br />
11 Note, that <strong>the</strong> word has kept its original word initial [�], without transforming it<br />
into [�], which would have been predicted by Hebrew phonology. Although<br />
this is a remarkable fact for Netzer, it turns out that almost no word borrowed<br />
by Mo<strong>der</strong>n Hebrew would change its initial [�] to [�]. Even not verbs that have<br />
had to un<strong>der</strong>go morpho-phonological processes (e.g. fibrek from English to<br />
fabricate). The only exception I have found in dictionaries is <strong>the</strong> colloquial<br />
form pilos<strong>of</strong>iya for filos<strong>of</strong>iya ‘philosophy’, as well as <strong>the</strong> verb formed from it,<br />
pilsef ‘to philosophise’. Fur<strong>the</strong>rmore, it can be argued that pilos<strong>of</strong>iya is not even<br />
a mo<strong>der</strong>n borrowing. The only reason why one would still expect firgen<br />
������� ��� ����������� �� ������ ��������� �� ���� ��� ������� �������� �<br />
��<br />
�� ��� ����� ������� �� � ����� ���� �� ��� ��������� ���������à ���<br />
���<br />
�� �������� ������ ����� ��������� ��� ������� ��]. <strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, one<br />
�<br />
may claim that /�/ <strong>and</strong> /�/ should be consi<strong>der</strong>ed as distinct phonemes in Mo<strong>der</strong>n<br />
Hebrew, even if no proposed minimal pair that I know <strong>of</strong> is really convincing.<br />
References<br />
������ ������ �� ��� ������ ������ ����� ������ A Magyar Nyelv Történeti-<br />
Etimológiai Szótára [The Historical-Etymological Dictionary <strong>of</strong><br />
The Hungarian Language]. Akadémiai Kiadó, Budapest.<br />
Berdichevsky, Norman (1986). Zamenh<strong>of</strong> <strong>and</strong> Esperanto. Ariel, A Review <strong>of</strong><br />
Arts <strong>and</strong> Letters in Israel, 64: 58-71.<br />
Blau Henrik, Károly Láng (1995). Szájról-szájra, Magyar-jiddis<br />
�������������� ����� ����� ������� ��������� ���� 2 .<br />
Goldin, Bernard (1982). The Supposed Yiddish Origin <strong>of</strong> <strong>the</strong> Esperanto<br />
Morpheme. edz, Jewish Language Review, 2: 21-33.<br />
Graaf, Tjeerd, de (1969). Aspects <strong>of</strong> neutrino astrophysics. Wolters-Noordh<strong>of</strong>f<br />
nv, Groningen.<br />
Netzer, Nissan (1988). “Fargen” - Employing a Yiddish Root to Bridge a<br />
Linguistic Gap in <strong>the</strong> Hebrew Language (in Hebrew, with<br />
English abstract). Hebrew Computational Linguistics, 26: 49-58.<br />
Piron, Claude (1984). Contribution à l’étude des apports du yidiche à<br />
l’ésperanto. Jewish Language Review, 4: 15-29.<br />
Sáenz-Badillos, Angel (1993). A History <strong>of</strong> <strong>the</strong> Hebrew Language. University<br />
Press, Cambridge.<br />
Schwarzwald, Ora (Rodrigue) (1995). The Components <strong>of</strong> <strong>the</strong> Mo<strong>der</strong>n Hebrew<br />
Lexicon: The Influence <strong>of</strong> Hebrew Classical Sources, Jewish
Weak Interactions 145<br />
Languages <strong>and</strong> O<strong>the</strong>r Foreign Languages on Mo<strong>der</strong>n Hebrew (in<br />
Hebrew, with English abstract). Hebrew Linguistics, 39: 79-90.<br />
Shur, Shimon (1979). Language Innovation <strong>and</strong> Socio-political Setting: The<br />
Case <strong>of</strong> Mo<strong>der</strong>n Hebrew. Hebrew Computational Linguistics, 15:<br />
IV-XIII.
Prosodic Acquisition: a Comparison <strong>of</strong> Two<br />
Theories<br />
Angela Grimm<br />
1. Introduction<br />
During language development, children’s word productions are target <strong>of</strong> a<br />
variety <strong>of</strong> prosodic processes as e.g. syllable deletions, syllable additions<br />
<strong>and</strong> stress shift. Using current phonological <strong>the</strong>ory, investigators have<br />
explained <strong>the</strong> production pattern in a number <strong>of</strong> different ways.<br />
In this paper, I review two approaches to <strong>the</strong> development <strong>of</strong> word<br />
stress: Fikkert’s (1994) <strong>the</strong>ory <strong>of</strong> trochaic template mapping <strong>and</strong> Demuth &<br />
Fee’s (1995) prosodic hierarchy account. Both <strong>the</strong>ories assume that<br />
children build up <strong>the</strong> prosodic representation <strong>of</strong> words step-by-step, starting<br />
with <strong>the</strong> smallest unit <strong>and</strong> ending with an adult-like representation. I argue<br />
that both <strong>the</strong>ories are problematic because <strong>the</strong>y overgenerate certain<br />
structures (e.g. level stress), but that <strong>the</strong> model <strong>of</strong> Demuth & Fee can better<br />
account for <strong>the</strong> data presented so far.<br />
This paper is organized as follows: since it is crucial in both <strong>the</strong>ories,<br />
paragraph 2 briefly introduces <strong>the</strong> basic assumptions <strong>of</strong> <strong>the</strong> prosodic<br />
hierarchy. In paragraph 3, I will give a survey <strong>of</strong> Fikkert’s (1994) model <strong>of</strong><br />
stress development <strong>and</strong> Demuth <strong>and</strong> Fee’s (1995) model based on <strong>the</strong><br />
prosodic hierarchy. In paragraph 4, I discuss <strong>the</strong> problems arising with <strong>the</strong><br />
models <strong>and</strong> paragraph 5 concludes.<br />
2. The prosodic hierarchy <strong>of</strong> words<br />
The prosodic hierarchy up to <strong>the</strong> word level consists <strong>of</strong> four constituents.<br />
The lowest element <strong>of</strong> <strong>the</strong> prosodic hierarchy is <strong>the</strong> mora (µ). Since <strong>the</strong>re<br />
are <strong>of</strong>ten no segmental slots in moraic models, <strong>the</strong> mora has a double<br />
function as <strong>the</strong> unit <strong>of</strong> syllable weight <strong>and</strong> as <strong>the</strong> unique sub-syllabic<br />
constituent. The moraic level is dominated by <strong>the</strong> syllabic level (σ), <strong>and</strong><br />
syllables are parsed into feet (F) at <strong>the</strong> foot level above. The highest unit is
148 Angela Grimm<br />
<strong>the</strong> prosodic word (Wd) which directly dominates <strong>the</strong> foot level (see Figure<br />
1):<br />
Prosodic word (Wd)<br />
Foot (F)<br />
Syllable (σ)<br />
Mora (µ)<br />
Figure 4. The prosodic hierarchy (Selkirk, 1980)<br />
Syllables differ with respect to <strong>the</strong> number <strong>of</strong> moras <strong>the</strong>y contain. Light<br />
syllables contain one mora, while heavy syllables contain at least two. The<br />
tendency <strong>of</strong> languages to assign stress to heavy syllables is expressed by<br />
<strong>the</strong> Weight-to-Stress-Principle (WSP). In a parametric approach to word<br />
stress (cf. Hayes, 1995), languages ei<strong>the</strong>r respect this principle (quantitysensitive<br />
languages) or do not (quantity-insensitive languages).<br />
The next constituent <strong>of</strong> <strong>the</strong> prosodic organization above <strong>the</strong> syllable<br />
level is <strong>the</strong> foot. Ideally, <strong>the</strong> foot is binary branching which implies that it<br />
should consist <strong>of</strong> two moras or <strong>of</strong> two syllables. Thus, a binary foot can be<br />
monosyllabic if it contains two moras (e.g. ����� ‘duck’) or it can be<br />
disyllabic if it consists <strong>of</strong> two syllables or two moras (e.g. ������ ‘papa’).<br />
The head constituent <strong>of</strong> <strong>the</strong> foot receives stress.<br />
The prosodic word is <strong>the</strong> domain <strong>of</strong> stress application. It can also<br />
coincide with a single foot. Because <strong>the</strong> foot size is <strong>the</strong> smallest shape a<br />
prosodic word can have, it is called Minimal Word. Many languages have<br />
restrictions such that content words must not be smaller than <strong>the</strong> minimal<br />
word. There is ample evidence that <strong>the</strong> minimal word restriction also<br />
governs <strong>the</strong> shape <strong>of</strong> <strong>the</strong> early words in language acquisition (Demuth &<br />
Fee, 1995; Demuth, 1996; Fikkert, 1994; Ota, 2001).<br />
A very important principle <strong>of</strong> <strong>the</strong> prosodic hierarchy is <strong>the</strong> Strict Layer<br />
Hypo<strong>the</strong>sis (Selkirk, 1984) which dem<strong>and</strong>s that layers must not be skipped,<br />
i.e. that a given prosodic constituent(n-1) is contained in <strong>the</strong> constituent(n)<br />
immediately above. Fur<strong>the</strong>rmore, it requires that constituents have one <strong>and</strong><br />
only one head, which implies that <strong>the</strong>re is always a difference in<br />
prominence among <strong>the</strong> elements forming a given prosodic unit.
Prosodic Acquisition: a Comparison <strong>of</strong> Two Theories 149<br />
3. The acquisition <strong>of</strong> word stress: two current models<br />
3.1. Fikkert (1994)<br />
Fikkert’s study <strong>of</strong> Dutch children is <strong>the</strong> most detailed research on stress<br />
acquisition to date. Fikkert mainly focused on disyllabic words <strong>and</strong> argued<br />
for <strong>the</strong> foot as <strong>the</strong> basic unit <strong>of</strong> development.<br />
Although Fikkert’s model is based on Dutch, she claims that <strong>the</strong><br />
trochaic template is universal in child language since it is <strong>the</strong> only quantityinsensitive<br />
foot in <strong>the</strong> typology <strong>of</strong> Hayes (1991). Thus, children should not<br />
show sensitivity to syllable weight at <strong>the</strong> earliest stages <strong>of</strong> prosodic<br />
acquisition. The postulation <strong>of</strong> a universal foot template implies that <strong>the</strong><br />
child always makes reference to <strong>the</strong> foot level in <strong>the</strong> word productions.<br />
Consequently, it is a foot, not a syllable that is being truncated in forms like<br />
below:<br />
����� ��������<br />
������� �����������<br />
Example 1.<br />
child form adult target gloss<br />
‘ballon’<br />
‘holiday’<br />
Fikkert assumes that <strong>the</strong> output a child produces is directed by <strong>the</strong> mapping<br />
<strong>of</strong> a melody template onto a trochaic template via prosodic circumscription.<br />
Based on phenomena such as truncation, stress shift <strong>and</strong> epen<strong>the</strong>sis, four<br />
different stages <strong>of</strong> prosodic development are postulated.<br />
Stage 1<br />
According to Fikkert, <strong>the</strong> child circumscribes <strong>the</strong> stressed syllable <strong>of</strong> <strong>the</strong><br />
adult form toge<strong>the</strong>r with its segmental material <strong>and</strong> maps it onto a trochaic<br />
template. The presumed representation <strong>of</strong> <strong>the</strong> child is given in Figure 2 (‘S’<br />
denotes <strong>the</strong> prominent position <strong>and</strong> ‘W’ <strong>the</strong> non-prominent position within<br />
<strong>the</strong> foot):
150 Angela Grimm<br />
Wd<br />
F<br />
σS σW<br />
�����<br />
Figure 5. The prosodic representation at stage 1<br />
Prosodic circumscription forces <strong>the</strong> child to divide <strong>the</strong> input into two parts,<br />
<strong>the</strong> kernel (i.e. <strong>the</strong> stressed syllable) <strong>and</strong> <strong>the</strong> residue. In <strong>the</strong> mapping<br />
process, <strong>the</strong> kernel (�����) is mapped onto <strong>the</strong> strong position in <strong>the</strong><br />
prosodic template. The residue (/��/) becomes truncated because <strong>the</strong>re are<br />
no empty positions in <strong>the</strong> template. The mapping onto <strong>the</strong> trochaic template<br />
accounts for <strong>the</strong> fact that, if <strong>the</strong> result <strong>of</strong> prosodic circumscription is a<br />
monosyllabic foot, sometimes a syllable is added to receive a disyllabic<br />
output, for example ������� instead <strong>of</strong> ��������.<br />
Stage 2<br />
At stage 2, <strong>the</strong> child circumscribes a trochaic foot. Thus, if <strong>the</strong> prosodic<br />
circumscription already results in a trochee as in ����������/ ‘holiday’, <strong>the</strong><br />
trochee remains unchanged in <strong>the</strong> output <strong>and</strong> appears as ��������. Words<br />
consisting <strong>of</strong> more than a single foot are circumscribed differently. Fikkert<br />
argues that <strong>the</strong> child selects <strong>the</strong> next stressed syllable to <strong>the</strong> left in addition<br />
to <strong>the</strong> stressed final syllable. For instance, Dutch /������������ ‘crocodile’<br />
should be realized as ���������� because <strong>the</strong> ultimate, main stressed<br />
syllable <strong>and</strong> <strong>the</strong> antepenultimate, secondary stressed syllable are kept. The<br />
disyllabic representation is <strong>the</strong>n mapped onto <strong>the</strong> trochaic template<br />
resulting in a trochaic pattern. Since <strong>the</strong> production template still consists<br />
<strong>of</strong> one single trochaic foot, stress shifts to <strong>the</strong> initial syllable. The<br />
representation <strong>of</strong> <strong>the</strong> child is depicted in Figure 3:
Prosodic Acquisition: a Comparison <strong>of</strong> Two Theories 151<br />
Wd<br />
F<br />
σS σW<br />
����� ����<br />
������ �����<br />
Figure 6. The prosodic representation at stage 2<br />
Stage 3<br />
At stage 3, <strong>the</strong> productions are extended to two feet. According to Fikkert,<br />
<strong>the</strong> children have noticed that <strong>the</strong> target words can consist <strong>of</strong> more than a<br />
single foot. She claims that her subjects realized two syllables <strong>of</strong> <strong>the</strong> target<br />
word with equal prominence (level stress). However, her argument for <strong>the</strong><br />
level stress stage is ra<strong>the</strong>r weak: she stipulates that <strong>the</strong> children have to<br />
produce two equally stressed feet because <strong>the</strong>y are unable to realize stress<br />
at word level.<br />
The prosodic representation at stage 3 is depicted in Figure 4 below:<br />
Wd<br />
F F<br />
σS σW σS σW<br />
������ �����<br />
Figure 7. The prosodic representation at stage 3<br />
Since <strong>the</strong> trochaic foot still governs <strong>the</strong> productions, weak positions in <strong>the</strong><br />
template can be filled with extra syllables.
152 Angela Grimm<br />
Stage 4<br />
The representations are now adult-like. The word level stress has been<br />
acquired <strong>and</strong> <strong>the</strong> child is able to operate at <strong>the</strong> level <strong>of</strong> <strong>the</strong> prosodic word.<br />
3.2. Demuth & Fee (1995)<br />
Demuth & Fee propose a more abstract approach which, although primarily<br />
based on data <strong>of</strong> English acquiring children, aims to capture <strong>the</strong> prosodic<br />
development universally. The basic assumption in Demuth & Fee’s model<br />
is that prosodic development goes along <strong>the</strong> prosodic hierarchy (see Figure<br />
1). In contrast to Fikkert, Demuth & Fee avoid <strong>the</strong> notion <strong>of</strong> prosodic<br />
circumscription <strong>and</strong> trochaic template mapping. According to <strong>the</strong>m,<br />
sensitivity to <strong>the</strong> moraic structure <strong>of</strong> <strong>the</strong> mo<strong>the</strong>r tongue is already <strong>the</strong>re<br />
from <strong>the</strong> onset <strong>of</strong> word production on. They distinguish between <strong>the</strong><br />
following stages:<br />
Stage 1<br />
The first stage is characterized by sub-minimal (monomoraic) words. The<br />
productions consist <strong>of</strong> a single CV-syllable <strong>and</strong> <strong>the</strong>re are no vowel length<br />
distinctions yet. Thus, <strong>the</strong> phonological representation <strong>of</strong> <strong>the</strong> words also is<br />
CV.<br />
Stage 2<br />
At stage 2, children realize words <strong>of</strong> foot-size (Minimal Words). Stage 2 is<br />
characterized by three successional sub-stages: at <strong>the</strong> beginning, <strong>the</strong> foot is<br />
disyllabic as for example in ������ ‘papa’. Second, as soon as <strong>the</strong> child is<br />
able to produce coda consonants <strong>the</strong> foot can also have a monosyllabic<br />
form, e.g. ����� ‘duck’. Third, <strong>the</strong> vowel length distinction becomes<br />
phonemic. The child is now aware <strong>of</strong> <strong>the</strong> fact that <strong>the</strong> stressed syllable <strong>of</strong><br />
Dutch ���������� ‘banana’ has to be realized with a long vowel ������, while<br />
in �������� ‘giraffe’ <strong>the</strong> second vowel remains short (examples from Robin,<br />
see Fikkert, 1994). Demuth & Fee assume a direct relationship between<br />
distinctive vowel length <strong>and</strong> <strong>the</strong> appearance <strong>of</strong> coda consonants. Thus, a<br />
CVV structure counts as sub-minimal, <strong>and</strong> a CVVCVV structure as<br />
minimal as long as <strong>the</strong> child does not produce coda consonants.
Stage 3<br />
Prosodic Acquisition: a Comparison <strong>of</strong> Two Theories 153<br />
Beyond <strong>the</strong> minimal word stage, syllable structure can be more complex<br />
<strong>and</strong> words can have a larger size than a single foot. This is also <strong>the</strong> stage<br />
where <strong>the</strong> largest progress in <strong>the</strong> development <strong>of</strong> <strong>the</strong> word stress is<br />
predicted. The child seems to become aware that feet have to be stressed<br />
<strong>and</strong> that <strong>the</strong>re are language-specific stress rules. Demuth & Fee do not<br />
assume a trochaic template. However, <strong>the</strong>y adopt Fikkert’s assumption <strong>of</strong><br />
an obligatory intermediate stage <strong>of</strong> level stress where two feet are produced<br />
with primary stress.<br />
At <strong>the</strong> end <strong>of</strong> stage 3, children acquire stress at <strong>the</strong> word level <strong>and</strong> <strong>the</strong>y<br />
realize one primary stress per word.<br />
Stage 4<br />
At <strong>the</strong> final stage, extrametrical (i.e. unfooted) syllables are permitted.<br />
Children at this stage operate at <strong>the</strong> level <strong>of</strong> <strong>the</strong> prosodic word.<br />
4. Discussion <strong>of</strong> <strong>the</strong> models<br />
Although both models can explain a number <strong>of</strong> frequently observed<br />
patterns like syllable deletions <strong>and</strong> word size restrictions, <strong>the</strong>re are a<br />
number <strong>of</strong> empirical <strong>and</strong> <strong>the</strong>oretical problems related with <strong>the</strong> models.<br />
First, Fikkert <strong>and</strong> Demuth & Fee assume that <strong>the</strong> prosodic development<br />
proceeds bottom-up, i.e. from a lower level <strong>of</strong> representation (<strong>the</strong> foot or<br />
<strong>the</strong> mora) to <strong>the</strong> top <strong>of</strong> <strong>the</strong> prosodic hierarchy (<strong>the</strong> prosodic word). Children<br />
invariably have to pass trough one stage before <strong>the</strong>y can go to <strong>the</strong> next. For<br />
example, multisyllabic words like ‘elephant’ or ‘crocodile’ have to show a<br />
level stress pattern before <strong>the</strong>y can be produced adult-like.<br />
Fikkert explicitly points to that fact. Missing evidence in her data is<br />
explained by <strong>the</strong> recording modalities or is due to <strong>the</strong> fact that a given stage<br />
took a very short time. Demuth & Fee, in contrast, are not explicit with<br />
respect to <strong>the</strong> or<strong>der</strong>ing <strong>of</strong> <strong>the</strong> stages. However, <strong>the</strong>y claim that prosodic<br />
development proceeds along <strong>the</strong> prosodic hierarchy. Since in <strong>the</strong> prosodic<br />
hierarchy one constituent strictly dominates <strong>the</strong> constituent below, stages<br />
cannot vary with respect to <strong>the</strong>ir temporal or<strong>der</strong>. According to <strong>the</strong> models,<br />
<strong>the</strong> following realizations for /������������ ‘crocodile’ <strong>of</strong> Jule, a girl<br />
acquiring German, should be chronologically impossible (data from my<br />
own corpus):
154 Angela Grimm<br />
Example 2.<br />
child form age description<br />
�����<br />
��������<br />
����������<br />
(1;08,12)<br />
(1;08,29)<br />
<strong>the</strong> main stressed syllable is realised<br />
a foot with final stress is realized<br />
(1;10,14) level stress emerges<br />
As <strong>the</strong> examples illustrate, level stress can occur after a finally stressed<br />
variant <strong>of</strong> <strong>the</strong> target word was produced, contrary to <strong>the</strong> predictions <strong>of</strong> <strong>the</strong><br />
models. Such an acquisition or<strong>der</strong> provides empirical evidence against level<br />
stress as an obligatory component <strong>of</strong> prosodic development. Additional<br />
empirical support comes from <strong>the</strong> data <strong>of</strong> English acquiring children<br />
examined by Kehoe & Stoel-Gammon (1997) who also could not find a<br />
systematic emergence <strong>of</strong> level stress.<br />
Level stress as assumed in <strong>the</strong> models above is problematic also from a<br />
grammatical point <strong>of</strong> view: <strong>the</strong> representation intended to create level stress<br />
(see Figure 4 above) essentially violates <strong>the</strong> strict layer hypo<strong>the</strong>sis because<br />
<strong>the</strong> two feet are not correctly bounded into <strong>the</strong> prosodic word. The problem<br />
is that <strong>the</strong> strict layer hypo<strong>the</strong>sis never can be kept by such a representation<br />
because <strong>the</strong>re is no gradation in prominence at <strong>the</strong> word level. According to<br />
prosodic <strong>the</strong>ory, two equally stressed feet must not occur within a single<br />
prosodic word:<br />
*Wd<br />
FS FS<br />
σS σW σS σW<br />
������ �����<br />
Figure 8. The ill-formed representation <strong>of</strong> <strong>the</strong> prosodic hierarchy as implied by<br />
Fikkert (1994) <strong>and</strong> Demuth & Fee (1995)<br />
Both models remain vague with respect to <strong>the</strong> source <strong>of</strong> level stress: it is<br />
unclear how <strong>the</strong> stages <strong>of</strong> level stress fit to <strong>the</strong> assumption that prosodic<br />
development is directed by universal prosodic principles. Since <strong>the</strong>y do not<br />
discuss <strong>the</strong> possibility <strong>of</strong> a child-specific representation, <strong>the</strong> representation<br />
according to <strong>the</strong> prosodic hierarchy should look like illustrated in Figure 6:
Prosodic Acquisition: a Comparison <strong>of</strong> Two Theories 155<br />
Wd Wd<br />
FS<br />
FS<br />
σS σW σS σW<br />
������ �����<br />
Figure 9. A prosodic representation that incorporates <strong>the</strong> requirements <strong>of</strong> <strong>the</strong><br />
prosodic hierarchy <strong>and</strong> that allows for level stress<br />
The representation in Figure 6 admits <strong>the</strong> co-occurrence <strong>of</strong> two equally<br />
stressed feet because every foot projects its own prosodic word. The<br />
drawback is that this assumption is ad hoc. There is no motivation for<br />
separating a single prosodic word like Dutch /������������ ‘crocodile’ into<br />
two prosodic words. In addition, it is an open question which factors could<br />
trigger <strong>the</strong> merging <strong>of</strong> <strong>the</strong> two prosodic words into a single one later.<br />
Ano<strong>the</strong>r problem is that <strong>the</strong> models described above are primarily based<br />
on truncation patterns in multisyllabic words. This is critical from a<br />
methodological point <strong>of</strong> view because it is presupposed that <strong>the</strong> truncation<br />
<strong>of</strong> syllables is exclusively triggered by prosodic size restrictions. Recent<br />
evidence, however, suggests that segmental properties <strong>of</strong> syllables can also<br />
affect <strong>the</strong> truncation rate. For example, syllables with sonorant onsets seem<br />
to be more prone to truncation than syllables with obstruent onsets (Kehoe<br />
& Stoel-Gammon 1997).<br />
A comparison <strong>of</strong> both models suggests that <strong>the</strong> predictions <strong>of</strong> <strong>the</strong><br />
template mapping model <strong>of</strong> Fikkert (1994) are sometimes too strong. Thus,<br />
<strong>the</strong> prosodic hierarchy model <strong>of</strong> Demuth & Fee (1995) seems to be superior<br />
because <strong>of</strong> its greater flexibility. First <strong>of</strong> all, it prevents Fikkert’s circular<br />
process <strong>of</strong> assigning a trochaic structure via prosodic circumscription that<br />
actually should be created by <strong>the</strong> foot template. Fur<strong>the</strong>rmore, <strong>the</strong> prosodic<br />
hierarchy model allows for more variability in <strong>the</strong> productions <strong>of</strong> children.<br />
For example, it allows for <strong>the</strong> co-occurrence <strong>of</strong> monosyllabic <strong>and</strong> disyllabic<br />
feet in contrast to Fikkert’s model that only proposes disyllabic trochees for<br />
a very long period <strong>of</strong> time. As <strong>the</strong> data <strong>of</strong> children acquiring English<br />
suggest, <strong>the</strong>re are doubts on Fikkert’s view that <strong>the</strong> disyllabic trochee is <strong>the</strong><br />
unique representation at <strong>the</strong> early stages (Kehoe & Stoel-Gammon, 1997;
156 Angela Grimm<br />
Salidis & Johnson, 1997). Moreover, Fikkert predicts a systematic stress<br />
shift to <strong>the</strong> left in disyllabic iambs, a pattern that still needs empirical<br />
evaluation. It is also possible that stress shift is ra<strong>the</strong>r <strong>the</strong> result <strong>of</strong> a<br />
complex interplay <strong>of</strong> factors like edge preferences, weight sensitivity <strong>and</strong><br />
segmental factors than <strong>of</strong> a simple template mapping mechanism. If this is<br />
true, stress shift can be bidirectional to <strong>the</strong> left or to <strong>the</strong> right, depending on<br />
<strong>the</strong> relative importance <strong>of</strong> <strong>the</strong> factors involved.<br />
Fikkert’s model is more detailed than <strong>the</strong> model <strong>of</strong> Demuth & Fee. It is<br />
at best elaborated for stage 1 <strong>and</strong> 2. With respect to <strong>the</strong> later stages she<br />
remains somewhat inconsistent. For example, she strongly argues for <strong>the</strong><br />
foot as <strong>the</strong> relevant prosodic unit, but already at stage 2 <strong>the</strong> syllable, not <strong>the</strong><br />
foot, becomes <strong>the</strong> target <strong>of</strong> circumscription:<br />
“[...] <strong>the</strong> child realises both syllables <strong>of</strong> <strong>the</strong> target word. However, stress<br />
falls on <strong>the</strong> first syllable. The segmental material <strong>of</strong> both syllables <strong>of</strong> <strong>the</strong><br />
adult word is taken out <strong>and</strong> mapped onto <strong>the</strong> child’s trochaic template [...]”<br />
(p. 210).<br />
Fikkert also consi<strong>der</strong>s <strong>the</strong> possibility <strong>of</strong> circumscribing a foot. She<br />
concludes that <strong>the</strong> children circumscribe syllables because <strong>the</strong> surviving<br />
syllables do not constitute a foot in <strong>the</strong> adult word. But <strong>the</strong> examples she<br />
presents (p. 211) do form two feet within a weight-sensitive model, with<br />
each foot containing at least two moras (Example 3):<br />
��������� �������������<br />
���������� �������������<br />
���������� �����������������<br />
��������� �������������<br />
Example 3.<br />
child form adult target gloss<br />
‘elephant’<br />
‘pelican’<br />
‘locomotive’<br />
‘farm’<br />
Fikkert cannot account for this fact because she exclusively assumes<br />
weight-insensitive trochees at stage 2. Demuth & Fee’s model, in contrast,<br />
would allow for <strong>the</strong> retention <strong>of</strong> <strong>the</strong> foot as <strong>the</strong> relevant unit since it<br />
assumes sensitivity to syllable weight with <strong>the</strong> emergence <strong>of</strong> <strong>the</strong> foot<br />
structure.<br />
Demuth & Fee, in contrast, have problems to explain <strong>the</strong> stress shift to<br />
<strong>the</strong> first syllable in <strong>the</strong> examples above for two reasons: first, recent<br />
evidence suggests that <strong>the</strong> relationship between distinctive vowel length<br />
<strong>and</strong> <strong>the</strong> emergence <strong>of</strong> coda consonants is not as categorial as <strong>the</strong>y claim. In<br />
an examination <strong>of</strong> Fikkert’s data, Salidis & Johnson (1997) found that,
Prosodic Acquisition: a Comparison <strong>of</strong> Two Theories 157<br />
contrary to <strong>the</strong>ir English acquiring child, <strong>the</strong> vowel length was not<br />
controlled by <strong>the</strong> Dutch children even if <strong>the</strong>y correctly produced coda<br />
consonants. If in turn, children cannot control vowel length appropriately,<br />
<strong>the</strong>y cannot assign two moras to a long vowel. The authors relate <strong>the</strong><br />
divergence between <strong>the</strong> languages to <strong>the</strong> impact <strong>of</strong> vowel quantity on <strong>the</strong><br />
stress pattern: in English, <strong>the</strong> long vowels in (C)VV(C) syllables count as<br />
heavy <strong>and</strong> thus attract stress, contrary to Dutch which ra<strong>the</strong>r relies on <strong>the</strong><br />
open-closed distinction. In Dutch, a (C)VC syllable counts as heavy, while<br />
a (C)VV does not. Thus, a learner <strong>of</strong> Dutch presumably does not rely on<br />
vowel length as an indicator for stress, while it is crucial for a learner <strong>of</strong><br />
English to identify <strong>the</strong> relationship between vowel quantity <strong>and</strong> stress.<br />
Second, given that <strong>the</strong> absence <strong>of</strong> <strong>the</strong> vowel length distinction is an artifact<br />
<strong>of</strong> <strong>the</strong> investigation <strong>and</strong> children have mastered <strong>the</strong> vowel length<br />
distinctions if <strong>the</strong>y produce bimoraic feet. Then nei<strong>the</strong>r universal nor<br />
language-specific constraints could account for <strong>the</strong> fact that <strong>the</strong> superheavy<br />
finals lose <strong>the</strong>ir primary stress in favor <strong>of</strong> <strong>the</strong> less heavy ultimates because<br />
Dutch follows <strong>the</strong> universal generalization that a (C)VCC (e.g. /����/) or a<br />
(C)VVC (e.g. /����/) syllable is heavier than a (C)VV syllable (/��/,/���/).<br />
The observation that Dutch children need more time to acquire vowel<br />
length distinctions indicates that language-specific properties may influence<br />
<strong>the</strong> prosodic representation in a more detailed way than assumed so far.<br />
Thus, fur<strong>the</strong>r empirical work is needed to shed light on <strong>the</strong> interplay <strong>of</strong><br />
universal principles <strong>and</strong> language-specific conditions in prosodic<br />
development.<br />
5. Conclusion<br />
In <strong>the</strong> present paper, two models <strong>of</strong> prosodic development are introduced<br />
<strong>and</strong> examined. As <strong>the</strong>y evidence, <strong>the</strong> acquisition <strong>of</strong> word prosody largely<br />
conforms to <strong>the</strong> prosodic hierarchy in such a way that universal prosodic<br />
constituents as <strong>the</strong> foot or <strong>the</strong> mora govern children’s word productions.<br />
This is essential in both models. However, it has turned out in <strong>the</strong><br />
discussion that <strong>the</strong>re are empirical, <strong>the</strong>oretical <strong>and</strong> methodical<br />
shortcomings. Common problems <strong>of</strong> both accounts are <strong>the</strong> absence <strong>of</strong><br />
empirical <strong>and</strong> <strong>the</strong>oretical motivation <strong>of</strong> level stress <strong>and</strong> <strong>the</strong> reliance on<br />
truncations as <strong>the</strong> primary diagnostics <strong>of</strong> prosodic development.<br />
In sum, <strong>the</strong> evidence so far ra<strong>the</strong>r supports <strong>the</strong> prosodic hierarchy model<br />
<strong>of</strong> Demuth & Fee (1995) because it is more flexible than Fikkert’s template<br />
mapping model.
158 Angela Grimm<br />
References<br />
Demuth, K. (1996). The prosodic structure <strong>of</strong> early words. In: J. Morgan & K.<br />
Demuth (eds.) From signal to syntax: Bootstrapping from speech<br />
to grammar in early acquisition. Lawrence Erlbaum Associates,<br />
Hillsdale, N.J., 171-184.<br />
Demuth, C. & Fee, J. (1995). Minimal words in early phonological<br />
development. Ms., Brown University <strong>and</strong> Dalhousie University.<br />
Fikkert, P.M. (1994). <strong>On</strong> <strong>the</strong> acquisition <strong>of</strong> prosodic structure. Holl<strong>and</strong> Institute<br />
<strong>of</strong> Generative Linguistics, Dordrecht,<br />
Hayes, B. (1991). Metrical stress <strong>the</strong>ory: principles <strong>and</strong> case studies. Ms,<br />
UCLA.<br />
Hayes, B. (1995). Metrical stress <strong>the</strong>ory. Chicago University Press, Chicago.<br />
Johnson, J. & Salidis, J.S. (1997). The production <strong>of</strong> minimal words: A<br />
longitudinal case study <strong>of</strong> phonological development. Language<br />
Acquisition, 6 (1): 1-36.<br />
Kehoe, M. & Stoel-Gammon, C. (1997). The acquisition <strong>of</strong> prosodic structure:<br />
An investigation <strong>of</strong> current accounts <strong>of</strong> children’s prosodic<br />
development. Language, 73 (1): 113-144.<br />
Ota, M. (2001). Phonological Theory <strong>and</strong> <strong>the</strong> Development <strong>of</strong> Prosodic<br />
Structure: Evidence from Child Japanese. Available at<br />
http://www.ling.ed.ac.uk/~mits/downloadables.shtml<br />
Selkirk, E. (1980). The role <strong>of</strong> prosodic categories in English word stress.<br />
Linguistic Inquiry, 11: 563-605.<br />
Selkirk, E. (1984). <strong>Phonology</strong> <strong>and</strong> Syntax: The relation between Sound <strong>and</strong><br />
Structure. MIT Press, Cambridge, MA.
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in<br />
Nivkh<br />
Hidetoshi Shiraishi<br />
1. Introduction<br />
1.1. Background<br />
Morphologically complex words <strong>of</strong>ten exhibit phonological similarities<br />
with <strong>the</strong>ir morphologically related base forms which <strong>the</strong>y are <strong>der</strong>ived from.<br />
In a number <strong>of</strong> cases, <strong>the</strong>se similarities yield a marked phonological pattern<br />
given <strong>the</strong> general rules or phonotactics <strong>of</strong> <strong>the</strong> language (Kenstowicz, 1996;<br />
Burzio, 1997, 2002 etc.). In Optimality-Theory (OT), similarity between<br />
existing words is captured by Output-to-Output (OO) correspondence<br />
constraints (Burzio, 1996, 2002; Kenstowicz, 1996, 1997; Benua, 1997ab;<br />
Ito <strong>and</strong> Mester, 1997; Steriade, 2000 etc.). The marked phonological<br />
pattern arises when similarity between words takes priority over <strong>the</strong><br />
canonical phonology <strong>of</strong> <strong>the</strong> language. OT expresses this situation by<br />
ranking OO-correspondence constraints above phonological markedness<br />
constraints. OO-correspondence constraints evaluate <strong>the</strong> output c<strong>and</strong>idates<br />
<strong>and</strong> select <strong>the</strong> one which is most similar to <strong>the</strong> base.<br />
Since <strong>the</strong> base plays a crucial role in computing <strong>the</strong> phonology <strong>of</strong> its<br />
<strong>der</strong>ivatives, it is important to identify <strong>the</strong> correct surface form as <strong>the</strong> base.<br />
Many authors have observed that OO-constraints have access to <strong>the</strong> base<br />
only if <strong>the</strong> latter occurs as an independent word (Kenstowicz, 1996; Benua,<br />
1997a; Ito <strong>and</strong> Mester, 1997). 1 Consi<strong>der</strong> <strong>the</strong> s-voicing observed in <strong>the</strong><br />
nor<strong>the</strong>rn dialects <strong>of</strong> Italian. In <strong>the</strong>se dialects, s <strong>and</strong> z are in complementary<br />
distribution. Z appears intervocalically, when <strong>the</strong> flanking vowels belong to<br />
<strong>the</strong> same phonological word (examples from Kenstowicz, 1996: 373-374).
160 Hidetoshi Shiraishi<br />
1.1<br />
a. a[z]ola ‘button hole’<br />
a[z]ilo ‘nursery school’<br />
ca[z]-a ‘house’<br />
ca[z]-ina ‘house - diminutive’<br />
b. lo [s]apevo ‘I knew it’<br />
telefonati [s]i ‘having called each o<strong>the</strong>r’<br />
The distribution <strong>of</strong> s-voicing in lexical items containing a prefix is more<br />
complicated. When <strong>the</strong> target precedes <strong>the</strong> boundary, s-voicing applies<br />
(1.2a). But when <strong>the</strong> target follows <strong>the</strong> boundary, s-voicing may or may not<br />
apply, even if <strong>the</strong> structural description <strong>of</strong> s-voicing is met (1.2b, c).<br />
1.2<br />
a. di[z]-onesto ‘dishonest’<br />
di[z]-ugale ‘unequal’<br />
b. re-[z]istenza ‘resistance’<br />
pre-[z]entire ‘to have a presentiment’<br />
c. a-[s]ociale ‘asocial’<br />
bi-[s]essuale ‘bisexual’<br />
pre-[s]entire ‘to hear in advance’<br />
The unexpected blocking <strong>of</strong> s-voicing in 1.2c is in sharp contrast with <strong>the</strong><br />
items in 1.2b where z surfaces intervocalically, following <strong>the</strong> phonological<br />
norm <strong>of</strong> <strong>the</strong> language. Nespor <strong>and</strong> Vogel (1986) pointed out that <strong>the</strong> crucial<br />
difference between <strong>the</strong> items in 1.2b <strong>and</strong> 1.2c lies in <strong>the</strong> lexical status <strong>of</strong> <strong>the</strong><br />
stem to which <strong>the</strong> prefix is attached; in 1.2c <strong>the</strong> stem occurs as an<br />
independent word (sociale, sessuale, etc.) whereas in 1.2b it does not<br />
(*sistenza, etc.). Following this view, Kenstowicz (1996) claimed that <strong>the</strong>re<br />
is a lexico-morphological pressure from <strong>the</strong> independently occurring stem<br />
to surface its <strong>der</strong>ivative as similar as possible. The presence <strong>of</strong> such an<br />
independently occurring immediate constituent is thus crucial in computing<br />
<strong>the</strong> phonology <strong>of</strong> a morphologically complex item. Kenstowicz dubbed this<br />
generalization Base-Identity; <strong>the</strong> base forces its <strong>der</strong>ivative to be formally as<br />
similar as possible in or<strong>der</strong> to “improve <strong>the</strong> transparency <strong>of</strong> morphological<br />
relationships between words <strong>and</strong> enhance lexical access” (Kenstowicz,<br />
1996: 372).
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 161<br />
1.3 Base-Identity: Given an input structure [X Y] output c<strong>and</strong>idates are<br />
evaluated for how well <strong>the</strong>y match [X] <strong>and</strong> [Y] if <strong>the</strong> latter occur as<br />
independent words. (Kenstowicz, 1996: 372)<br />
The languages in East Asia provide an interesting test for this<br />
generalization. Languages as Korean or Japanese show a systematic<br />
difference in <strong>the</strong> composition <strong>of</strong> verbs <strong>and</strong> nouns; while verbal stems<br />
always surface with a morphological extension, nominal stems may surface<br />
without such an extension. This means that complex words formed from a<br />
nominal stem always have an independently occurring base to which <strong>the</strong>y<br />
phonologically should conform, whereas verbal <strong>der</strong>ivatives lack such a<br />
base <strong>and</strong> hence should not show such conformity. This prediction is borne<br />
out in Korean in which <strong>der</strong>ivatives <strong>of</strong> nominal <strong>and</strong> verbal stems are subject<br />
to different phonology (Kenstowicz, 1996. See section 2.3 below.). In this<br />
paper, I discuss ano<strong>the</strong>r language <strong>of</strong> East Asia, Nivkh, which also has an<br />
asymmetric composition <strong>of</strong> nouns <strong>and</strong> verbs like Korean <strong>and</strong> Japanese. I<br />
will focus on two phonological phenomena, Consonant Alternation <strong>and</strong><br />
Final Fricative Devoicing <strong>and</strong> show that both phenomena exhibit<br />
asymmetries between nominal <strong>and</strong> verbal phonology. I will discuss each<br />
case in detail <strong>and</strong> argue that Base-Identity is <strong>the</strong> driving force <strong>of</strong> <strong>the</strong>se<br />
asymmetries.<br />
The article is organized as follows. I will start with a descriptive sketch<br />
<strong>of</strong> Consonant Alternation (section 2.1) <strong>and</strong> <strong>the</strong>n illustrate <strong>the</strong> exceptional<br />
behavior <strong>of</strong> nominal stems as a case <strong>of</strong> noun-verb asymmetry (section 2.2).<br />
While most previous works, including my own, somehow stipulated <strong>the</strong><br />
asymmetric behavior <strong>of</strong> nominal <strong>and</strong> verbal stems, I will argue that Base-<br />
Identity provides a superior analysis which is free from such a stipulation.<br />
Section 3 discusses <strong>the</strong> second phenomenon, Final Fricative Devoicing. I<br />
will illustrate <strong>the</strong> asymmetric behavior <strong>of</strong> fricative-final nominal <strong>and</strong> verbal<br />
stems when followed by a suffix. The pattern <strong>of</strong> asymmetry is as in CA:<br />
while verbal phonology is subject to canonical phonology, nominal<br />
phonology is not. Section 4 concludes.<br />
1.2. About Nivkh<br />
Nivkh (also called Gilyak) is an isolated language spoken by <strong>the</strong> people <strong>of</strong><br />
Nivkh, who live on <strong>the</strong> isl<strong>and</strong> <strong>of</strong> Sakhalin <strong>and</strong> in <strong>the</strong> lower reaches <strong>of</strong> <strong>the</strong><br />
Amur River in <strong>the</strong> Russian Far East. The language has four dialects <strong>and</strong> <strong>the</strong>
162 Hidetoshi Shiraishi<br />
major discrepancy is between <strong>the</strong> Amur dialect, spoken in <strong>the</strong> Amur area on<br />
<strong>the</strong> continent <strong>and</strong> <strong>the</strong> west coast <strong>of</strong> north Sakhalin, <strong>and</strong> <strong>the</strong> Sakhalin dialect<br />
spoken in <strong>the</strong> east coast <strong>of</strong> Sakhalin. Nivkh is listed in <strong>the</strong> UNESCO Red<br />
Book on endangered languages as being seriously endangered. According<br />
to <strong>the</strong> census <strong>of</strong> 1989, <strong>the</strong> percentage <strong>of</strong> speakers is 23, 3% <strong>of</strong> <strong>the</strong> total<br />
population <strong>of</strong> 4,681. 2 This article concerns <strong>the</strong> phonology <strong>of</strong> <strong>the</strong> Amur<br />
dialect spoken by <strong>the</strong> continental Nivkh. All <strong>the</strong> examples are from <strong>the</strong><br />
following sources, unless o<strong>the</strong>rwise mentioned: Krejnovich (1937), <strong>and</strong><br />
Saveleva <strong>and</strong> Taksami (1970).<br />
2. Consonant Alternation<br />
2.1. A descriptive sketch<br />
I will first outline <strong>the</strong> segmental inventory <strong>of</strong> Nivkh.<br />
2.1 Consonantal inventory <strong>of</strong> Nivkh<br />
(I) aspirated plosives p� t� c� k� q�<br />
(II) non-aspirated plosives p t c k q<br />
(III) voiceless fricatives f r� s x �<br />
(IV) voiced fricatives v r 3 z � �<br />
nasals m n � �<br />
lateral l<br />
glides j h<br />
2.2 Vowels<br />
i � u<br />
e o<br />
a<br />
Consonant Alternation (henceforth CA) is a phonological process which<br />
changes <strong>the</strong> feature [continuant] in obstruents when <strong>the</strong>y are placed in<br />
certain phonological <strong>and</strong> morphosyntactic contexts. Descriptively, CA<br />
consists <strong>of</strong> two processes: spirantization, in which a plosive changes to a<br />
fricative, <strong>and</strong> hardening, in which a fricative changes to a plosive.<br />
Laryngeal features are also relevant since aspirated plosives only alternate
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 163<br />
with voiceless fricatives <strong>and</strong> non-aspirated plosives with voiced fricatives,<br />
i.e. <strong>the</strong> alternation is strictly between <strong>the</strong> obstruents <strong>of</strong> row (I) <strong>and</strong> (III), or<br />
4, 5<br />
(II) <strong>and</strong> (IV).<br />
2.3 Spirantization: (I) > (III), (II) > (IV)<br />
a. (I) > (III) mac�a [r�]om (< t�om) ‘fat <strong>of</strong> a seal’<br />
seal fat<br />
c�ol�i [�]os (< q�os ) ‘neck <strong>of</strong> a<br />
reindeer neck reindeer’<br />
b. (II) > (IV) p�eq [v]��x (< p��x ) ‘chicken soup’<br />
chicken soup<br />
mac�a [z]us (< cus) ‘meat <strong>of</strong> a seal’<br />
seal meat<br />
2.4 Hardening: (III) > (I), (IV) > (II)<br />
a. (III) > (I) c�x�f [q�]a- (< �a-) ‘to shoot a bear’<br />
bear shoot<br />
cus [t�]a- (< r�a-) ‘to bake meat’<br />
meat bake<br />
b. (IV) > (II) tux [k]e- (< �e-) ‘to take an axe’<br />
axe take<br />
p�n�nx [t]�u- (< r�u-) ‘to teach one's<br />
one's sister teach sister’<br />
The phonological contexts <strong>of</strong> spirantization <strong>and</strong> hardening are in<br />
complementary distribution. Spirantization takes place when <strong>the</strong> target<br />
(plosive) follows a vowel, a glide, or a plosive (2.5). There is no<br />
spirantization when <strong>the</strong> target follows a fricative or a nasal (2.6).<br />
2.5 Spirantization Preceding segment<br />
Vowel mac�a [r�]om ‘fat <strong>of</strong> a seal’<br />
Glide k��nraj [r�]om ‘fat <strong>of</strong> a duck’<br />
k��nraj [v]��x ‘duck soup’<br />
Plosive �t [r�]om ‘fat <strong>of</strong> a species <strong>of</strong><br />
duck’<br />
amsp [v]��x ‘soup <strong>of</strong> a species<br />
<strong>of</strong> seal’
164 Hidetoshi Shiraishi<br />
2.6 No spirantization<br />
Fricative c�x�f t�om ‘bear fat’<br />
c�x�f p��x ‘bear soup’<br />
Nasal k�e� t�i ‘sun ray’<br />
rum d�f ‘Rum(person)’s house’<br />
<strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, hardening occurs when <strong>the</strong> target (fricative) follows<br />
ei<strong>the</strong>r a fricative or a nasal (2.7). When a segment o<strong>the</strong>r than fricative<br />
precedes <strong>the</strong> target, hardening does not occur (2.8).<br />
2.7 Hardening Preceding segment<br />
Fricative cx�f [q�]a- (< �a-) ‘to shoot a bear’<br />
lovr� [c]osq-(< zosq-) ‘to break a spoon’<br />
Nasal qan [d]�u- 6 (
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 165<br />
whereas hardening activates when a fricative is in <strong>the</strong> input. In <strong>the</strong> past,<br />
many approaches have overlooked this generalization <strong>and</strong> described <strong>the</strong><br />
rules as if <strong>the</strong>y had independent structural goals. This is not <strong>the</strong> case.<br />
Let us now move to <strong>the</strong> morphosyntactic conditioning. CA targets a<br />
segment at <strong>the</strong> left edge <strong>of</strong> a <strong>der</strong>ived morphosyntactic unit in <strong>the</strong> presence<br />
<strong>of</strong> a preceding segment. CA applies cyclically to every left edge <strong>of</strong> a<br />
morpho-syntactic unit until <strong>the</strong> maximal projection (NP, VP) is reached.<br />
2.10 Means <strong>of</strong> <strong>der</strong>ivation<br />
Prefixation p�-[r�]u (< t�u) ‘one’s own sledge’<br />
REF-sledge<br />
Postposition t��x-tox ‘towards <strong>the</strong> top’<br />
top-ALL<br />
tu-rox ‘towards a lake’<br />
qan-dox ‘towards a dog’<br />
Reduplication t�k[r�]�k- ‘to be silent’<br />
(Sakhalin dialect, Hattori, 1962: 107)<br />
NP formation mac�a [r�]om ‘fat <strong>of</strong> a seal’<br />
VP formation cx�f [q�]a- (< �a-) ‘to shoot a bear’<br />
<strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, CA never targets segments in a non-<strong>der</strong>ived<br />
environment, nor does it apply across XP boundary, as shown in 2.11 <strong>and</strong><br />
2.12, respectively.<br />
2.11 CA does not apply in non-<strong>der</strong>ived environment<br />
utku *ut[�]u ‘man’<br />
n��s *n��[c�] ‘teeth’<br />
e�l�<br />
2.12 No CA across XP boundary (subject-predicate)<br />
r�o- ‘The child holds (something)’<br />
= [NPe�l�] [VPr�o-] (‘child’ is subject)<br />
Example 2.13 below differs minimally from example 2.12 above with<br />
respect to <strong>the</strong> application <strong>of</strong> CA. In <strong>the</strong> former, CA applies since <strong>the</strong> noun<br />
is <strong>the</strong> object <strong>of</strong> <strong>the</strong> following predicate. Thus <strong>the</strong>se two words form a VP,<br />
differing minimally from example 2.12.
166 Hidetoshi Shiraishi<br />
2.13<br />
e�l� [t�]o- ‘(Someone) holds <strong>the</strong> child’<br />
= [VP[NPe�l�][V t�o-]] (‘child’ is object)<br />
2.2. The spirantization – hardening asymmetry<br />
There is one environment in which <strong>the</strong> regular pattern <strong>of</strong> CA as depicted<br />
above fails to apply. Nouns beginning with a fricative never un<strong>der</strong>go<br />
hardening. In such a case, <strong>the</strong> structural goal <strong>of</strong> CA (2.9) is not achieved. In<br />
this context <strong>the</strong> o<strong>the</strong>rwise illicit fricative-fricative or nasal-fricative<br />
sequence appears.<br />
2.14<br />
a. t�ulv vo *t�ulv [b]o ‘winter village’<br />
winter village<br />
b. c���r vox *c���r [b]ox 'a hill covered with grass'<br />
grass hill<br />
c. t�f r�� *t�f [t�]� ‘entrance door’<br />
house door<br />
d. t�e� vaqi *t�e� [b]aqi ‘coal box’<br />
coal box<br />
Previous works have ei<strong>the</strong>r described this context as an exception to CA, or<br />
did not discuss it. In most cases, <strong>the</strong>se works simply stipulated that a) nouns<br />
do not un<strong>der</strong>go hardening, or alternatively b) only transitive verbs un<strong>der</strong>go<br />
hardening. <strong>On</strong>ce stated as a condition this way, <strong>the</strong> application <strong>of</strong> hardening<br />
to nouns can indeed be avoided. However, adding such a condition (in<br />
ei<strong>the</strong>r form) to a phonological rule pairs prosodic phonology with specific<br />
category labels (transitive verb, noun), which is unlikely to occur in natural<br />
languages (Nespor <strong>and</strong> Vogel, 1986; Selkirk, 1986 etc.). 7 But most<br />
critically, it is explanatorily unsatisfying; why should hardening be<br />
restricted to transitive verbs (or alternatively, why should nouns be an<br />
exception to hardening)? No literature provides a satisfactory answer to this<br />
question.<br />
The tacit assumption prevailing in <strong>the</strong> previous works is that <strong>the</strong> input to<br />
CA is <strong>the</strong> citation form, i.e. <strong>the</strong> form that appears in isolation. Following<br />
this assumption, <strong>the</strong> transitive verbs ought to un<strong>der</strong>go hardening since <strong>the</strong>y<br />
initiate with a fricative in <strong>the</strong> citation form. However, <strong>the</strong>re is no a priori
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 167<br />
reason that <strong>the</strong> citation form should be <strong>the</strong> un<strong>der</strong>lying form. In Shiraishi<br />
(2000), I defended <strong>the</strong> position that <strong>the</strong> citation form <strong>of</strong> <strong>the</strong>se transitive<br />
verbs cannot be <strong>the</strong> un<strong>der</strong>lying form, if we want to advocate a<br />
phonologically plausible analysis for <strong>the</strong> observed spirantization-hardening<br />
asymmetry. The lack <strong>of</strong> hardening in nouns could be interpreted as<br />
evidence that CA consists solely <strong>of</strong> spirantization, without hardening. I<br />
argued that transitive verbs <strong>of</strong> Nivkh initiate with a plosive at <strong>the</strong><br />
un<strong>der</strong>lying level, instead <strong>of</strong> a fricative that appears in <strong>the</strong> citation form.<br />
Initiating with a plosive, transitive verbs now un<strong>der</strong>go spirantization in <strong>the</strong><br />
8, 9<br />
same way as nouns do.<br />
2.15<br />
Previous analyses Shiraishi (2000)<br />
VP 'shoot a NP 'bird soup' VP 'shoot NP 'bird<br />
bear'<br />
a bear' soup'<br />
Un<strong>der</strong>lying<br />
form<br />
cx�f �a- p�eq p��x cx�f q�a- p�eq p��x<br />
Spirantization not<br />
applicable<br />
p�eq [v]��x blocked p�eq [v]��x<br />
Hardening cx�f [q�]a- not applicable<br />
Surface form cx�f q�a- p�eq v��x cx�f q�a- p�eq v��x<br />
The analysis in Shiraishi (2000) leaves hardening out <strong>of</strong> <strong>the</strong> list <strong>of</strong><br />
phonological processes; nouns do not un<strong>der</strong>go hardening since <strong>the</strong>re is no<br />
hardening in <strong>the</strong> phonology <strong>of</strong> <strong>the</strong> language.<br />
2.16<br />
Previous<br />
analyses<br />
Shiraishi (2000)<br />
Un<strong>der</strong>lying<br />
form<br />
t�ulv vo t�ulv vo<br />
Spirantization not applicable not applicable<br />
Hardening t�ulv [b]o<br />
Surface form �t�ulv bo t�ulv vo �: incorrect output<br />
This analysis is free from category-specific specification in <strong>the</strong> structural<br />
description <strong>of</strong> <strong>the</strong> rule, which was inevitable in <strong>the</strong> previous analyses.<br />
Although this analysis explains nicely why fricative-initial nouns never<br />
un<strong>der</strong>go hardening in Nivkh, it is not without problems. First, it
168 Hidetoshi Shiraishi<br />
manipulates <strong>the</strong> un<strong>der</strong>lying form <strong>of</strong> a specific lexical category (transitive<br />
verb) in or<strong>der</strong> to explain phonologically exceptional behavior. Although<br />
such a 'prespecification' at <strong>the</strong> un<strong>der</strong>lying level is not an uncommon way to<br />
approach phonological exceptions (cf. Inkelas, Orgun <strong>and</strong> Zoll, 1997<br />
amongst o<strong>the</strong>rs), such an approach does not explain why only this<br />
particular class <strong>of</strong> words needs to un<strong>der</strong>go such manipulation. Since<br />
prespecification puts unpredictable information into <strong>the</strong> lexicon, it is a<br />
strong descriptive device which leaves little space for phonological<br />
generalizations. Contrary to what seems to be <strong>the</strong> case at first glance, <strong>the</strong><br />
analytical gain <strong>of</strong> Shiraishi (2000) from previous analyses is not so<br />
obvious. <strong>On</strong>e may ask correctly what <strong>the</strong> difference between <strong>the</strong> two<br />
analyses is, which claim that a) nouns are exceptions to hardening<br />
(previous analyses) or b) transitive verbs un<strong>der</strong>go spirantization because<br />
<strong>the</strong>y initiate with plosives un<strong>der</strong>lyingly (Shiraishi 2000). In o<strong>the</strong>r words, it<br />
remains an arbitrary choice that only transitive verbs, <strong>and</strong> not o<strong>the</strong>r<br />
categories, un<strong>der</strong>go prespecification.<br />
Secondly, <strong>the</strong> relationship between <strong>the</strong> un<strong>der</strong>lying form <strong>and</strong> <strong>the</strong> citation<br />
form is obscured in transitive verbs. By positing a form o<strong>the</strong>r than <strong>the</strong><br />
citation form as <strong>the</strong> un<strong>der</strong>lying form, <strong>the</strong> citation form would always be<br />
<strong>der</strong>ived from <strong>the</strong> un<strong>der</strong>lying form by some morphological operation. That<br />
is, Shiraishi (2000) created asymmetry between <strong>the</strong> morpholexical make-up<br />
between nominal <strong>and</strong> verbal stems.<br />
2.17<br />
Nominal stem Verbal stem<br />
Un<strong>der</strong>lying form p��x q�a-<br />
Surface form p��x �a-<br />
In fact, this asymmetry describes <strong>the</strong> historical path <strong>of</strong> <strong>der</strong>ivation <strong>of</strong><br />
transitive verbs (Jakobson, 1957; Austerlitz, 1977). <strong>On</strong> synchronic grounds,<br />
however, it is highly doubtful whe<strong>the</strong>r such a morphological operation can<br />
be justified.<br />
In <strong>the</strong> next section I propose an alternative approach to <strong>the</strong><br />
spirantization-hardening (or noun-transitive verb) asymmetry, which makes<br />
use nei<strong>the</strong>r <strong>of</strong> prespecification nor <strong>of</strong> information about category labels.<br />
Instead, I will argue that correspondence relation between output forms<br />
plays a decisive role in distinguishing <strong>the</strong> phonological behavior <strong>of</strong> <strong>the</strong> two<br />
groups. <strong>On</strong>ce stated this way, nothing ought to be stipulated in or<strong>der</strong> to
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 169<br />
<strong>der</strong>ive <strong>the</strong> surface form; this follows naturally from <strong>the</strong> phonological<br />
principles <strong>of</strong> <strong>the</strong> language.<br />
2.3. Noun-verb asymmetry as Base-Identity<br />
In Nivkh, verbal <strong>and</strong> nominal stems differ from each o<strong>the</strong>r in one crucial<br />
morphological aspect; verbal stems should always end in a morphological<br />
extension but nominal stems do not. Or put differently, verbal stems never<br />
surface in isolation, whereas nominal stems do. This means that bare verbal<br />
stems cannot function as citation forms. Usually, <strong>the</strong> form with an<br />
infinitival suffix (-d�, -t�) provides <strong>the</strong> citation form.<br />
2.18<br />
2.19<br />
Stem /�a/ ‘to shoot~’ /r�o/ ‘to take’<br />
Infinitive �a-d� r�o-d�<br />
(citation form)<br />
‘when~’ �a-�an r�o-�an<br />
/vo/ ‘village’ /�ota/ ‘town’<br />
Citation form vo �ota<br />
Allative vo-rox �ota-rox<br />
As mentioned in section 1, independent forms <strong>of</strong>ten exercise special<br />
influence on <strong>the</strong> realization <strong>of</strong> morphologically related forms in <strong>der</strong>ived<br />
contexts. For instance, in certain varieties <strong>of</strong> English <strong>the</strong> existence <strong>of</strong> <strong>the</strong><br />
form condense guarantees that <strong>the</strong> vowel <strong>of</strong> <strong>the</strong> second syllable in <strong>the</strong><br />
morphologically related word condensation does not reduce to a schwa.<br />
2.20<br />
co�nd[�]nsa�tion co�mp[�]nsa�tion<br />
cond[��]nse co�mp[�]nsa�te<br />
<strong>Phonology</strong> would expect <strong>the</strong> unstressed vowel <strong>of</strong> condensation to surface<br />
with a schwa, as is <strong>the</strong> case with <strong>the</strong> structurally similar compensation. The<br />
usual explanation for this asymmetry is that <strong>the</strong> vowel reduction in<br />
condensation is blocked by virtue <strong>of</strong> <strong>the</strong> existence <strong>of</strong> <strong>the</strong> morphologically<br />
related form condense, which appears with a full vowel [�] (Chomsky <strong>and</strong>
170 Hidetoshi Shiraishi<br />
Halle, 1968: 110-116). <strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, compensation lacks such a<br />
morphologically related form with a full vowel. Hence unstressed vowel<br />
reduces to a schwa, following <strong>the</strong> phonological norm <strong>of</strong> <strong>the</strong> language.<br />
Ano<strong>the</strong>r example comes from Korean. In Korean a stem-final consonant<br />
cluster surfaces only when it is followed by a vowel-initial suffix. In<br />
combination with a consonant-initial suffix, <strong>the</strong> cluster is simplified to a<br />
single consonant (Kenstowicz, 1996: 375).<br />
2.21<br />
Stem /kaps/ ‘price’ /talk/ ‘chicken’<br />
Citation form kap tak<br />
Nominative kaps-i talk-i<br />
Comitative kap-k'wa tak-k'wa<br />
In <strong>the</strong> speech <strong>of</strong> younger generation <strong>of</strong> Seoul, however, simplification overapplies<br />
to contexts where vowel-initial suffix follows <strong>the</strong> stem.<br />
2.22<br />
Nominative kap-i tak-i<br />
Interestingly, this overgeneralization does not apply to verbal stems. Here<br />
<strong>the</strong> consonant cluster surfaces.<br />
2.23<br />
�ps-�ss-� palk-�ss-�<br />
Stem /�ps/ ‘not have’ /palk/<br />
‘be bright’<br />
Past-informal (*�p-�ss-�)<br />
(*pak-�ss-�)<br />
Non-past-formal �p-t'a pak-t'a<br />
Kenstowicz analyzed <strong>the</strong> absence <strong>of</strong> <strong>the</strong> cluster simplification in verbal<br />
stems to be due to a lack <strong>of</strong> corresponding citation forms. As in Nivkh,<br />
verbal stems in Korean never appear in isolation; <strong>the</strong>y should always<br />
appear with an inflectional ending. In contrast, nominal stems are free to<br />
appear without any inflectional ending, so <strong>the</strong>y exercise strong influence on<br />
<strong>the</strong> realization <strong>of</strong> <strong>the</strong>ir <strong>der</strong>ivatives. Verbal stems, on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, surface<br />
with consonant clusters since <strong>the</strong>re are no isolated counterparts which<br />
forces conformity to it. This is an instance <strong>of</strong> Base-Identity, which requires<br />
forms in <strong>der</strong>ived contexts to be formally similar to <strong>the</strong> base. This is <strong>the</strong>
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 171<br />
generalization captured in <strong>the</strong> Base-Identity constraint <strong>of</strong> Kenstowicz (1.3),<br />
repeated below.<br />
2.24 (=1.3) Base-Identity: Given an input structure [X Y] output<br />
c<strong>and</strong>idates are evaluated for how well <strong>the</strong>y match [X] <strong>and</strong> [Y]<br />
if <strong>the</strong> latter occur as independent words. (Kenstowicz, 1996:<br />
372)<br />
We can account for <strong>the</strong> noun-verb asymmetry in Korean using Base-<br />
Identity as a high-ranked constraint. By ranking Base-Identity above a<br />
faithfulness constraint which prohibits deletion <strong>of</strong> a segment in <strong>the</strong> input<br />
(MAX), nominal stems surface with a single consonant in concordance<br />
with <strong>the</strong> base.<br />
2.25<br />
constraints →<br />
/kaps+i/ base: kap<br />
c<strong>and</strong>idates ↓<br />
Base-Identity *CLUSTER MAX<br />
kapsi *!<br />
�kapi *<br />
Base-Identity is vacuously satisfied in verbal stems. Since <strong>the</strong>re is no base<br />
to which verbal stems should conform, verbal stems exhibit canonical<br />
phonology. Consonant clusters surface only if a vowel-initial suffix<br />
follows, elsewhere <strong>the</strong>y are simplified. A phonological markedness<br />
constraint *CLUSTER penalizes every output c<strong>and</strong>idate containing a triconsonantal<br />
cluster.<br />
2.26<br />
constraints →<br />
/�ps+�ss+�/ base: ø<br />
c<strong>and</strong>idates ↓<br />
��ps-�ss-�<br />
Base-Identity *CLUSTER MAX<br />
�p-�ss-� *!
172 Hidetoshi Shiraishi<br />
2.27<br />
constraints →<br />
/�ps-t'a/ base: ø<br />
c<strong>and</strong>idates ↓<br />
Base-Identity *CLUSTER MAX<br />
�ps-t'a *!<br />
��p-t'a *<br />
The noun-verb asymmetry <strong>of</strong> hardening in Nivkh is strikingly similar to <strong>the</strong><br />
case <strong>of</strong> Korean. As in Korean, verbal stems <strong>of</strong> Nivkh are not allowed to<br />
surface in isolation; <strong>the</strong>y always require a morpho-syntactic extension<br />
(2.18). This is in contrast to nominal stems, which may surface in isolation<br />
(2.19). The difference is reflected directly in <strong>the</strong>ir phonological behavior;<br />
verbal stems un<strong>der</strong>go hardening, nominal stems do not. In <strong>the</strong> next section I<br />
will show how this analysis formally works.<br />
2.4. Base-Identity blocks hardening<br />
I assume that <strong>the</strong> phonological markedness constraint that induces<br />
hardening to be <strong>the</strong> Obligatory Contour Principle (OCP) [fric]. 10 OCP [fric]<br />
prohibits adjacent fricatives. Base-Identity, as defined in <strong>the</strong> previous<br />
section, prefers output c<strong>and</strong>idates which are similar to <strong>the</strong> base. With <strong>the</strong><br />
ranking Base-Identity >> OCP [fric], we obtain <strong>the</strong> desired output;<br />
hardening does not apply to nominal stems.<br />
2.28<br />
constraints →<br />
/tulv vo/ base: vo<br />
c<strong>and</strong>idates ↓<br />
Base-Identity OCP [fric] IDENT [cont]<br />
�tulv vo *<br />
tulv bo *! *<br />
Base-Identity is satisfied vacuously in verbal stems since <strong>the</strong>y lack a base.<br />
Being free from Base-Identity, an initial fricative now hardens to a plosive<br />
in or<strong>der</strong> to circumvent an OCP violation.
2.29<br />
constraints →<br />
/c�x�f �a-/ base: ø<br />
c<strong>and</strong>idates ↓<br />
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 173<br />
Base-Identity OCP IDENT [cont]<br />
c�x�f �a- *!<br />
�c�x�f [q�]a- *<br />
Since Base-Identity refers to <strong>the</strong> base <strong>and</strong> not to <strong>the</strong> input, this ranking<br />
always <strong>der</strong>ives <strong>the</strong> correct output no matter <strong>of</strong> <strong>the</strong> input value. This is<br />
illustrated in <strong>the</strong> tableau below in which <strong>the</strong> verbal stem initiates with a<br />
plosive in <strong>the</strong> input (cf. Shiraishi, 2000).<br />
2.30<br />
constraints →<br />
/c�x�f q�a-/ base: ø<br />
c<strong>and</strong>idates ↓<br />
Base-Identity OCP IDENT [cont]<br />
�c�x�f q�ac�x�f<br />
[�]a- *! *<br />
The present analysis correctly <strong>der</strong>ives <strong>the</strong> observed output no matter <strong>of</strong> <strong>the</strong><br />
input. There is thus no prespecification, in which input strings are fixed to<br />
take a particular form. Nor does it make use <strong>of</strong> information <strong>of</strong> category<br />
labels, a condition that was inevitable in previous descriptions in or<strong>der</strong> to<br />
let hardening apply appropriately. The current analysis makes a totally<br />
different claim. There is no exception to <strong>the</strong> hardening rule (nominal<br />
stems), nor should <strong>the</strong> specific un<strong>der</strong>goer (verbal stems) be prespecified at<br />
<strong>the</strong> un<strong>der</strong>lying level. Ra<strong>the</strong>r, <strong>the</strong> asymmetry <strong>of</strong> nominal <strong>and</strong> verbal stems<br />
follows from <strong>the</strong> existence <strong>of</strong> a base, which is an independent fact <strong>of</strong> <strong>the</strong><br />
language. By making use <strong>of</strong> such morpho-lexical information, <strong>the</strong> current<br />
analysis accounts for <strong>the</strong> noun-verb asymmetry without appealing to<br />
language-specific stipulations.<br />
3. Final Fricative Devoicing<br />
Base-Identity plays a crucial role in ano<strong>the</strong>r phonological phenomenon <strong>of</strong><br />
Nivkh. In this section, I will discuss such a case.
174 Hidetoshi Shiraishi<br />
3.1. Distribution <strong>of</strong> laryngeal features<br />
Like Danish, a full contrast <strong>of</strong> laryngeal features <strong>of</strong> Nivkh obstruents is<br />
realized only at <strong>the</strong> stem-initial position, which is <strong>the</strong> most prominent<br />
position as in many o<strong>the</strong>r languages (cf. Beckman, 1996). In o<strong>the</strong>r<br />
positions, laryngeal features do not exercise a phonemic contrast <strong>and</strong> <strong>the</strong><br />
feature value at <strong>the</strong> surface level is predictable from <strong>the</strong> context (Jakobson,<br />
1957: 83). In principle, non-prominent (stem-medial <strong>and</strong> final) positions<br />
only allow non-aspirated plosives <strong>and</strong> voiced fricatives. Aspirated plosives<br />
<strong>and</strong> voiceless fricatives, on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, are excluded from <strong>the</strong>se<br />
positions. Following Jakobson (1957), I will call <strong>the</strong>m <strong>the</strong> lenis <strong>and</strong> fortis<br />
series, respectively.<br />
3.1<br />
3.2<br />
Lenis obstruents non-aspirated plosives : p t c k q<br />
voiced fricatives : v r z � �<br />
Fortis obstruents aspirated plosives : p� t� c� k� q�<br />
voiceless fricatives : f r s x �<br />
pal ‘forest’ �t�k ‘fa<strong>the</strong>r’<br />
p�al ‘floor’ �k�n ‘mo<strong>the</strong>r’<br />
ra-d� ‘to drink’ ova ‘flour’<br />
r�a-d� ‘to bake’ muvi ‘porridge’<br />
eri ‘river’<br />
There are two exceptional contexts in which a voiceless fricative appears in<br />
a non-prominent position: i) when preceding a plosive, <strong>and</strong>/or ii) before an<br />
I[ntonational] P[hrase] boundary (Jakobson, 1957: 83).<br />
3.3<br />
a. esqa-d� ‘to hate’<br />
taft� ‘salt’<br />
kins ‘evil spirit’<br />
kins k�u-d� ‘to kill an evil spirit’<br />
c�x�f ‘bear’<br />
c�x�f k�u-d� ‘to kill a bear’<br />
als ‘berry’<br />
als p�e- ‘to pick berries’
. nivx ‘human’<br />
erx ‘to him/her’<br />
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 175<br />
The examples in 3.3b indicate that it is only <strong>the</strong> absolute final position that<br />
matters; <strong>the</strong> fricative second from <strong>the</strong> right appears as voiced. In Nivkh,<br />
<strong>the</strong>re are no words ending in consecutive voiceless fricatives, indicating<br />
that voicelessness is required only for <strong>the</strong> very last fricative in an IP. I<br />
assume this to be due to a restriction which I will call Final Fricative<br />
Devoicing (FFD). FFD targets every final fricative within an IP.<br />
Stem-final voiceless fricatives appear as voiced, however, as soon as <strong>the</strong><br />
above-mentioned conditions are removed. Thus if a stem-final fricative is<br />
embedded in an IP, i.e. not final in <strong>the</strong> domain, <strong>and</strong> if it is not adjacent to a<br />
plosive it becomes voiced (3.4a). This is in concordance with <strong>the</strong><br />
phonotactics <strong>of</strong> stem-medial fricatives which are always voiced (3.4b)<br />
unless adjacent to a plosive. This distribution is not surprising since stemmedial<br />
fricatives are expected not to coincide with an IP-boundary.<br />
3.4<br />
a. [kinz it-]I ‘go insane’<br />
[c�x�v l�j-]I<br />
'to kill a bear'<br />
[alz �a-]I<br />
‘to pick berry’<br />
b. ezmu- ‘to like~’<br />
urla ‘good’<br />
pa�la ‘red’<br />
Outside <strong>of</strong> <strong>the</strong>se two contexts, only lenis obstruents appear in nonprominent<br />
positions. Apparently, lenis obstruents have more distributional<br />
freedom than fortis obstruents, indicating <strong>the</strong>ir unmarked status in <strong>the</strong><br />
phonology <strong>of</strong> Nivkh. Since non-prominent positions are predictably<br />
occupied by lenis obstruents, I assume that obstruents in <strong>the</strong>se positions are<br />
unspecified for laryngeal features in <strong>the</strong> un<strong>der</strong>lying form. Unless contextsensitive<br />
requirements contravene, obstruents without laryngeal<br />
specifications surface as lenis, <strong>the</strong> unmarked obstruent <strong>of</strong> <strong>the</strong> language.
176 Hidetoshi Shiraishi<br />
3.2. Base-Identity in suffixation<br />
Having discussed <strong>the</strong> unmarked nature <strong>of</strong> <strong>the</strong> lenis obstruents, we are now<br />
ready to look at <strong>the</strong> way FFD interacts with Base-Identity. Such a case<br />
arises when a suffix attaches to a fricative-final stem.<br />
Like stem-medial <strong>and</strong> final positions, <strong>the</strong> initial obstruent <strong>of</strong> a suffix<br />
does not exhibit a laryngeal contrast, indicating that it is a non-prominent<br />
position. Except for a few exceptional cases, only lenis obstruents are<br />
allowed. 11<br />
3.5<br />
-tox/rox/dox allative (case suffix)<br />
-�u/gu/ku plural<br />
-t�/d� infinitive<br />
-gu/ku causative<br />
When affixed to a stem, <strong>the</strong> redundant [+voice] specification <strong>of</strong> <strong>the</strong> stemfinal<br />
segment spreads to <strong>the</strong> initial obstruent <strong>of</strong> <strong>the</strong> suffix.<br />
3.6<br />
‘to drink-INF’<br />
ra-d�<br />
‘big-INF’<br />
pil-d�<br />
‘to walk-INF’<br />
amam-d�<br />
‘to harness-INF’<br />
ifk-t�<br />
‘to bind-INF’<br />
jup-t�<br />
‘to help-CAU-INF’<br />
ro-gu-d�<br />
‘to do-CAU-INF’<br />
l�t-ku-d�<br />
c�am-gu ‘shaman-PL’<br />
c�am-dox ‘shaman-ALL’<br />
There is an interesting discrepancy between fricative-final nominal <strong>and</strong><br />
verbal stems in this context; following a verbal stem, <strong>the</strong> initial segment <strong>of</strong><br />
a suffix is always voiced (3.7a), while following a nominal stem, it is<br />
always voiceless (3.7b).<br />
fuv-d�<br />
i�-d�<br />
t�v�-d�<br />
3.7<br />
a. ‘to blow/to saw-INF’<br />
‘to kill-INF’<br />
‘to go inside <strong>the</strong> house-INF’
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 177<br />
jar-d�<br />
roz-gu-d�<br />
t�mz-gu-d�<br />
b. kins-ku<br />
‘to feed-INF’<br />
‘to divide-CAU-INF’<br />
‘drop-CAU-INF’<br />
‘evil spirit-PL’<br />
c�x�f-ku ‘bear-PL’<br />
or�r�-ku ‘Uilta-PL’<br />
t�f-tox ‘house-ALL’<br />
ti�r�-tox ‘wood-ALL’<br />
The reason <strong>of</strong> this discrepancy is not immediately clear. In particular, <strong>the</strong><br />
final voiceless fricative <strong>of</strong> nominal stems is a mystery. Being affixed by a<br />
suffix, it is no longer in <strong>the</strong> context <strong>of</strong> FFD, so nothing prevents it from<br />
appearing in <strong>the</strong> unmarked voiced fricative. In fact, this is <strong>the</strong> case with<br />
verbal stems; final fricatives <strong>of</strong> verbal stems are systematically voiced<br />
(3.7a). The o<strong>the</strong>r context-sensitive requirement, namely, <strong>the</strong> precedence to<br />
a plosive cannot be <strong>the</strong> reason ei<strong>the</strong>r since <strong>the</strong>se suffixes have a voiced<br />
variant, which surfaces when following a (redundant) [+voice] segment<br />
(3.6, 3.7a). The <strong>der</strong>ivatives <strong>of</strong> verbal stems in 3.7a show that <strong>the</strong> initial<br />
plosive <strong>of</strong> <strong>the</strong>se suffixes can accommodate a (preceding) voiced fricative,<br />
unlike plosives in a stem. But in fact, this option is not adopted in nominal<br />
stems. In short, <strong>the</strong>se context-sensitive requirements cannot explain <strong>the</strong><br />
different behavior <strong>of</strong> final-fricatives in nominal <strong>and</strong> verbal stems.<br />
Un<strong>der</strong> Base-Identity, however, such a discrepancy is explicable. Recall<br />
that nominal <strong>and</strong> verbal stems have different morpho-lexical compositions.<br />
Nominal stems can surface without any morphological ending, making <strong>the</strong><br />
last fricative target <strong>of</strong> FFD. In contrast, final fricative <strong>of</strong> a verbal stem is<br />
always followed by a morphological extension, making it irrelevant to<br />
FFD. Since Base-Identity claims that <strong>der</strong>ivatives should phonologically<br />
conform to <strong>the</strong> base, nominal <strong>der</strong>ivatives conform to <strong>the</strong>ir base, which ends<br />
in a voiceless fricative (due to FFD). This is not <strong>the</strong> case, however, for<br />
verbal stems since <strong>the</strong>y have no base <strong>and</strong> <strong>the</strong>refore do not un<strong>der</strong>lie such<br />
pressure. As a consequence, verbal stems un<strong>der</strong>go canonical phonology <strong>and</strong><br />
fricatives in non-prominent positions do appear as lenis, <strong>the</strong> unmarked<br />
obstruents <strong>of</strong> <strong>the</strong> language.<br />
Finally, it is important to note that reference to laryngeal specifications<br />
using Input-to-Output correspondence constraints is not a viable option in<br />
this context. Recall that <strong>the</strong>re is no laryngeal contrast in stem-final position<br />
in Nivkh. A phonological <strong>the</strong>ory which minimizes <strong>the</strong> specification <strong>of</strong><br />
predictable features in un<strong>der</strong>lying representations, which is <strong>the</strong> one adopted
178 Hidetoshi Shiraishi<br />
here, makes it impossible for Input-to-Output constraints to refer to <strong>the</strong><br />
voiceless status <strong>of</strong> stem-final fricatives. 12 Thus <strong>the</strong>ir voicelessness should<br />
come from somewhere else. According to <strong>the</strong> current analysis it originates<br />
from <strong>the</strong> base, <strong>the</strong> independently occurring isolated form.<br />
4. Conclusion<br />
In this paper I have discussed phonological asymmetries between nominal<br />
<strong>and</strong> verbal stems <strong>of</strong> Nivkh, as observed in two phonological phenomena<br />
CA <strong>and</strong> FFD. Though <strong>the</strong> asymmetries <strong>the</strong>mselves look very different on<br />
<strong>the</strong> surface, this article has made explicit that <strong>the</strong>y are subject to a common<br />
generalization, Base-Identity. Given <strong>the</strong> asymmetric composition <strong>of</strong> nouns<br />
<strong>and</strong> verbs, Base-Identity makes two predictions: i) nominal <strong>and</strong> verbal<br />
<strong>der</strong>ivatives exhibit different phonological patterns, <strong>and</strong> ii) it is <strong>the</strong> nominal<br />
stem which exhibits <strong>the</strong> non-canonical phonology given <strong>the</strong> strong pressure<br />
from <strong>the</strong> base. Both predictions were borne out in <strong>the</strong> phonological<br />
phenomena discussed above. The base plays a decisive role in computing<br />
<strong>the</strong> phonology <strong>of</strong> nominal <strong>and</strong> verbal <strong>der</strong>ivatives in both CA <strong>and</strong> FFD. As<br />
for CA, <strong>the</strong> current analysis correctly predicts that nominal <strong>der</strong>ivatives<br />
accommodate <strong>the</strong> o<strong>the</strong>rwise illicit segmental sequence (fricative-fricative,<br />
nasal-fricative), while verbal <strong>der</strong>ivatives do not. This analysis is superior to<br />
previous accounts since it makes no direct use <strong>of</strong> <strong>the</strong> notion <strong>of</strong> exception,<br />
which was inevitable in previous works. Ra<strong>the</strong>r, <strong>the</strong> suggested analysis<br />
relates <strong>the</strong> asymmetry in phonology to <strong>the</strong> compositional asymmetry<br />
between nouns <strong>and</strong> verbs.<br />
As for FFD, nominal <strong>der</strong>ivatives showed conformity to <strong>the</strong>ir base, in<br />
ending in a voiceless fricative. Verbal stems on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, do not show<br />
such conformity since <strong>the</strong>y lack a base. Unlike nominal <strong>der</strong>ivatives, <strong>the</strong><br />
stem-final fricatives <strong>of</strong> verbal <strong>der</strong>ivatives appear as lenis, following <strong>the</strong><br />
canonical phonology <strong>of</strong> Nivkh. Base-Identity provides us with <strong>the</strong><br />
mechanism un<strong>der</strong>lying <strong>the</strong> noun-verb asymmetry, <strong>and</strong> it correctly predicts<br />
<strong>the</strong>ir phonological behavior with respect to <strong>the</strong> canonical phonology <strong>of</strong> <strong>the</strong><br />
language.
Acknowledgements<br />
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 179<br />
I would like to thank Dicky Gilbers, Angela Grimm, Maartje Schreu<strong>der</strong>,<br />
Jeroen van de Weijer <strong>and</strong> <strong>the</strong> audiences <strong>of</strong> ULCL <strong>Phonology</strong> meeting at<br />
Leiden (27-05-2003) <strong>and</strong> TABU dag (20-06-2003, Groningen) for<br />
comments on parts <strong>of</strong> this article. I bear all responsibility for errors.<br />
Notes<br />
1<br />
“…identity effects will come into play only to <strong>the</strong> extent that <strong>the</strong> immediate<br />
constituents composing <strong>the</strong> complex structure constitute independently<br />
occurring outputs…(Kenstowicz 1996: 373)”, “The base <strong>of</strong> an OOcorrespondence<br />
relation is a licit output word, which is both morphologically<br />
<strong>and</strong> phonologically well-formed (Benua 1997a: 29)”, “The bound form <strong>of</strong> a<br />
stem is segmentally identical with its corresponding free form (Ito <strong>and</strong> Mester<br />
1997: 431)”.<br />
2<br />
See www.let.rug.nl/~toshi/ for more information.<br />
3<br />
The rhotic r <strong>of</strong> Nivkh is classified here <strong>and</strong> elsewhere in <strong>the</strong> literature (e.g.<br />
Trubetzkoj 1939) as a voiced fricative since it patterns as such in <strong>the</strong> CA<br />
system. Its voiceless r� counterpart is an apical trill containing portions without<br />
vocal cord vibration (Ladefoged <strong>and</strong> Maddieson 1996: 236).<br />
4<br />
Regarding this nature <strong>of</strong> CA, one may postulate a single laryngeal feature<br />
(ra<strong>the</strong>r than two) for both plosives <strong>and</strong> fricatives, e.g. [+spread glottis] for both<br />
aspirated plosives <strong>and</strong> voiceless fricatives. Such an analysis is proposed by<br />
Jakobson (1957) <strong>and</strong> Blevins (1993). See also section 3 below.<br />
5<br />
Segments that un<strong>der</strong>went CA are put in square brackets. Abbreviations are:<br />
ALL= allative, asp= aspiration, I=Intonational phrase, INF=infinitive, NP =<br />
noun phrase, PL= plural, VP = verb phrase, XP = maximal projection.<br />
6<br />
The alternation (r >) t > d is due to post-nasal voicing.<br />
7<br />
CA exhibits aspects <strong>of</strong> prosodic phonology (I am using this term to contrast<br />
with lexical phonology); it is sensitive to pause insertions <strong>and</strong> to speech rate. I<br />
would classify it as a P-structure rule in <strong>the</strong> terminology <strong>of</strong> Selkirk (1986). Pstructure<br />
rules exhibit phonological properties <strong>of</strong> prosodic phonology, yet <strong>the</strong>y<br />
are sensitive to syntactic bracketing (Selkirk 1986).<br />
8<br />
This line <strong>of</strong> analysis has antecedents. Amongst <strong>the</strong>m are: Kenstowicz <strong>and</strong><br />
Kisserberth (1979), Rushchakov (1981), Kaisse (1985), <strong>and</strong> Blevins (1993).<br />
Interestingly, Lev Shternberg, <strong>the</strong> pioneer <strong>of</strong> Nivkh study, assumed plosiveinitial<br />
forms to be <strong>the</strong> input to transitive structures, as well (Shternberg 1908).
180 Hidetoshi Shiraishi<br />
9<br />
Spirantization <strong>and</strong> hardening are not or<strong>der</strong>ed relative to each o<strong>the</strong>r in <strong>the</strong><br />
tableau below.<br />
10<br />
Post-nasal context requires different markedness constraint but I omit it from<br />
<strong>the</strong> discussion below. See Shiraishi (2000) for details.<br />
11<br />
Following a velar or a uvular plosive, <strong>the</strong> initial velar <strong>of</strong> a suffix appears as [x],<br />
spirantizing <strong>the</strong> former at <strong>the</strong> same time: �t�x-xu
Base-Identity <strong>and</strong> <strong>the</strong> Noun-Verb Asymmetry in Nivkh 181<br />
Ito, J. <strong>and</strong> A. Mester (1997). Correspondence <strong>and</strong> Compositionality: The Gagyo<br />
Variation in Japanese <strong>Phonology</strong>. In: I.Roca (ed.), 419-462.<br />
Jakobson, R. (1957). Notes on Gilyak. Roman Jakobson. Selected Writings II.<br />
Word <strong>and</strong> language. Mouton, The Hague <strong>and</strong> Paris, 72-102.<br />
Kaisse, E. (1985). Connected Speech. Academic Press, Orl<strong>and</strong>o.<br />
Kenstowicz, M. (1996). Base-Identity <strong>and</strong> Uniform Exponence: Alternatives to<br />
Cyclicity. In: J. Dur<strong>and</strong> <strong>and</strong> B. Laks (eds.), 365-395.<br />
Kenstowicz, M. (1997). Uniform exponence: Exemplification <strong>and</strong> extension.<br />
In: V. Miglio <strong>and</strong> B. Moren (eds.), 139-155.<br />
Kenstowicz, M. <strong>and</strong> C. Kisserberth. (1979). Generative <strong>Phonology</strong>: description<br />
<strong>and</strong> <strong>the</strong>ory. Academic Press, New York.<br />
Krejnovich, E. (1937). Fonetika nivxskogo (giljackogo) jazyka [<strong>Phonetics</strong> <strong>of</strong> <strong>the</strong><br />
Nivkh (Gilyak) language]. Uchpedgiz, Moskva - Leningrad.<br />
Ladefoged, P. <strong>and</strong> I. Maddieson. (1996). The Sounds <strong>of</strong> <strong>the</strong> World’s Languages.<br />
Blackwell, Oxford.<br />
Miglio, V. <strong>and</strong> B. Moren (eds.) (1997). University <strong>of</strong> Maryl<strong>and</strong> Working<br />
Papers in Linguistics, vol.5.<br />
Nepor, M. <strong>and</strong> I. Vogel (1986). Prosodic <strong>Phonology</strong>. Foris, Dordrecht.<br />
Roca, I. (ed.) (1997). Derivations <strong>and</strong> Constraints in <strong>Phonology</strong>. Clarendon<br />
Press, Oxford.<br />
Rushchakov, V. (1981). Akusticheskie xarakteristiki soglasnyx nivxskogo<br />
jazyka (avtoreferat). Ph.D.dissertation, Akademija Nauk CCCP,<br />
Leningradskoe otdelenie instituta jazykoznanija.<br />
Savel’eva,V. <strong>and</strong> C.Taksami. (1970). Nivxsko-russkij slovar. [Nivkh-Russian<br />
dictionary] Sovetskaja Enciklopedija, Moskva.<br />
Selkirk, E. (1986). <strong>On</strong> <strong>der</strong>ived domains in sentence phonology . <strong>Phonology</strong><br />
Yearbook, 3: 371-405.<br />
Shiraishi, H. (2000). Nivkh consonant alternation does not involve hardening.<br />
Journal <strong>of</strong> Chiba University Eurasian Society. No.3. 89-119<br />
(Also available at www.let.rug.nl/˜toshi/list_<strong>of</strong>_publication.htm).<br />
Abridged version has appeared in <strong>the</strong> Proceedings <strong>of</strong> <strong>the</strong> 120 th<br />
meeting <strong>of</strong> <strong>the</strong> Japanese Society <strong>of</strong> Linguists, 42-47.<br />
Shternberg, L. (1908). Materialy po izucheniju gilijackogo jazyka i fol’klora.<br />
In: Obrachy narodnoj slovesnosti. Vol. 1, Part I. Imper.<br />
Akademii Nauk, St.Petersburg.<br />
Steriade, D. (2000). Paradigm Uniformity <strong>and</strong> <strong>the</strong> <strong>Phonetics</strong>-<strong>Phonology</strong><br />
boundary. In: M. Broe <strong>and</strong> J. Pierrehumbert (eds.). Papers in<br />
Laboratory <strong>Phonology</strong> 5. Cambridge University Press,<br />
Cambridge, 313-334<br />
Trubetzkoj, N. (1939). Grundzuge <strong>der</strong> Phonologie. Travaux du Cercle<br />
Linguistique de Prague, Prague.
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns<br />
Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
1. Introduction 1<br />
The topic <strong>of</strong> this paper is how rhythmic variability in speech can be<br />
accounted for both phonologically <strong>and</strong> phonetically. The question is<br />
whe<strong>the</strong>r a higher speech rate leads to adjustment <strong>of</strong> <strong>the</strong> phonological<br />
structure, or just to 'phonetic compression', i.e. shortening <strong>and</strong> merging <strong>of</strong><br />
vowels <strong>and</strong> consonants, with preservation <strong>of</strong> <strong>the</strong> phonological structure. We<br />
claim that <strong>the</strong> melodic content <strong>of</strong> a phonological domain is indeed adjusted<br />
optionally when <strong>the</strong> speech rate increases. In o<strong>the</strong>r words, every speech rate<br />
has its own preferred register, in terms <strong>of</strong> Optimality Theory (Prince <strong>and</strong><br />
Smolensky, 1993) its own ranking <strong>of</strong> constraints.<br />
We will investigate prosodic variability as part <strong>of</strong> our main research<br />
project, which involves a comparison <strong>of</strong> <strong>the</strong> analyses <strong>of</strong> music <strong>and</strong><br />
language. Our ultimate aim is to provide evidence for <strong>the</strong> assumption that<br />
every temporal behavior is structured similarly (cf. Liberman, 1975).<br />
Gilbers <strong>and</strong> Schreu<strong>der</strong> (to appear) show that Optimality Theory owes a lot<br />
to <strong>the</strong> constraint-based music <strong>the</strong>ory <strong>of</strong> Lerdahl <strong>and</strong> Jackend<strong>of</strong>f (1983).<br />
Based on <strong>the</strong> great similarities between language <strong>and</strong> music we claim that<br />
musical knowledge can help in solving linguistic issues.<br />
In this paper, we will show that clashes are avoided in allegro tempo. In<br />
both language <strong>and</strong> music distances between beats are enlarged, i.e. <strong>the</strong>re<br />
appears to be more melodic content between beats. To illustrate this, we ran<br />
a pilot experiment in which we elicited fast speech. As expected, speech<br />
rate plays a role in rhythmic variability.<br />
The paper is organized as follows. In section 2 <strong>the</strong> data <strong>of</strong> <strong>the</strong><br />
experiment is introduced. Section 3 is addressed to <strong>the</strong> phonological<br />
framework <strong>of</strong> Optimality Theory <strong>and</strong> <strong>the</strong> different rankings <strong>of</strong> <strong>and</strong>ante <strong>and</strong><br />
allegro speech. The method <strong>of</strong> <strong>the</strong> experiment is discussed in section 4 <strong>and</strong><br />
<strong>the</strong> auditive <strong>and</strong> acoustic analyses plus <strong>the</strong> results follow in section 5. The<br />
perspectives <strong>of</strong> our analysis will be discussed in <strong>the</strong> final section.
184 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
2. Data<br />
We will discuss three types <strong>of</strong> rhythmic variability in Dutch. The first we<br />
will call “stress shifts to <strong>the</strong> right”; <strong>the</strong> second “stress shifts to <strong>the</strong> left” <strong>and</strong><br />
<strong>the</strong> third “beat reduction”. In <strong>the</strong> first type as exemplified in stúdietòelage<br />
(s w s w w) ‘study grant’, we assume that this compound can be realized as<br />
stúdietoelàge (s w w s w) in allegro speech. Perfèctioníst (w s w s) is an<br />
example <strong>of</strong> “stress shift to <strong>the</strong> left” <strong>and</strong> we expect a realization pèrfectioníst<br />
(s w w s) in allegro speech. The last type does not concern a stress shift, but<br />
a stress reduction. In zùidàfrikáans (s s w s) ‘South African’ compounding<br />
<strong>of</strong> zuid <strong>and</strong> afrikaans results in a stress clash. In fast speech this clash is<br />
avoided by means <strong>of</strong> reducing <strong>the</strong> second beat: zùidafrikáans (s w w s).<br />
Table 1 shows a selection <strong>of</strong> our data.<br />
Table 1. Data<br />
Type 1: stress shift to <strong>the</strong> right (<strong>and</strong>ante: s w s w w; allegro: s w w s w)<br />
stu die toe la ge ‘study grant’<br />
weg werp aan ste ker ‘disposable lighter’<br />
ka mer voor zit ter ‘chairman <strong>of</strong> <strong>the</strong> House <strong>of</strong> Parliament’<br />
Type 2: stress shift to <strong>the</strong> left (<strong>and</strong>ante: w s w s; allegro: s w w s)<br />
per fec tio nist ‘perfectionist’<br />
a me ri kaan ‘American’<br />
vi ri li teit ‘virility’<br />
Type 3: beat reduction (<strong>and</strong>ante: s s w s; allegro: s w w s)<br />
zuid a fri kaans ‘South African’<br />
schier mon nik oog ‘name <strong>of</strong> an isl<strong>and</strong>’<br />
gre go ri aans ‘Gregorian’<br />
The different rhythmic patterns are accounted for phonologically within <strong>the</strong><br />
framework <strong>of</strong> OT.<br />
3. Framework <strong>and</strong> phonological analysis<br />
The mechanism <strong>of</strong> constraint interaction, <strong>the</strong> essential characteristic <strong>of</strong> OT,<br />
is also used in <strong>the</strong> generative <strong>the</strong>ory <strong>of</strong> tonal music (Lerdahl <strong>and</strong><br />
Jackend<strong>of</strong>f, 1983). In both frameworks, constraint satisfaction determines<br />
grammaticality <strong>and</strong> in both frameworks <strong>the</strong> constraints are potentially
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 185<br />
conflicting <strong>and</strong> s<strong>of</strong>t, which means violable. Violation, however, is only<br />
allowed if it leads to satisfaction <strong>of</strong> a more important, higher ranked<br />
constraint. The great similarities between <strong>the</strong>se <strong>the</strong>oretical frameworks<br />
make comparison <strong>and</strong> interdisciplinary research possible.<br />
For example, restructuring rhythm patterns as a consequence <strong>of</strong> a higher<br />
playing rate is a very common phenomenon in music. In Figure 1 we give<br />
an example <strong>of</strong> re-/misinterpretation <strong>of</strong> rhythm in accelerated or sloppy<br />
playing.<br />
Dotted notes rhythm � triplet rhythm<br />
Figure 1. Rhythmic restructuring in music<br />
In Figure 1, <strong>the</strong> “dotted notes rhythm” (left <strong>of</strong> <strong>the</strong> arrow) is played as a<br />
triplet rhythm (right <strong>of</strong> <strong>the</strong> arrow). In <strong>the</strong> dotted notes rhythm <strong>the</strong> second<br />
note has a duration which is three times as long as <strong>the</strong> third, <strong>and</strong> in <strong>the</strong><br />
triplet rhythm <strong>the</strong> second note is twice as long as <strong>the</strong> third. In fast playing it<br />
is easier to have equal durations between note onsets. Clashes are thus<br />
avoided <strong>and</strong> one tries to distribute <strong>the</strong> notes, <strong>the</strong> melodic content, over <strong>the</strong><br />
measures as evenly as possible, even if this implies a restructuring <strong>of</strong> <strong>the</strong><br />
rhythmic pattern. To ensure that <strong>the</strong> beats do not come too close to each<br />
o<strong>the</strong>r in fast playing, <strong>the</strong> distances are enlarged, thus avoiding a staccatolike<br />
rhythm. In short, in fast tempos <strong>the</strong> musical equivalents <strong>of</strong> <strong>the</strong><br />
Obligatory Contour Principle (OCP), a prohibition on adjacency <strong>of</strong><br />
identical elements in language (McCarthy, 1986), become more important.<br />
We claim that - just as in music - <strong>the</strong> allegro patterns in all <strong>the</strong> different<br />
types <strong>of</strong> data in Table 1 are caused by clash avoidance. There is a<br />
preference for beats that are more evenly distributed over <strong>the</strong> phrase. The<br />
different structures can be described phonologically as a conflict between<br />
markedness constraints, such as FOOT REPULSION (���) (Kager, 1994), <strong>and</strong><br />
OUTPUT - OUTPUT CORRESPONDENCE constraints (cf. Burzio, 1998) within<br />
<strong>the</strong> framework <strong>of</strong> OT. FOOT REPULSION prohibits adjacent feet <strong>and</strong><br />
consequently prefers a structure in which feet are separated from each o<strong>the</strong>r<br />
by an unparsed syllable. This constraint is in conflict with PARSE-σ, which
186 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
dem<strong>and</strong>s that every syllable is part <strong>of</strong> a foot. OUTPUT - OUTPUT<br />
CORRESPONDENCE compares <strong>the</strong> structure <strong>of</strong> a phonological word with <strong>the</strong><br />
structure <strong>of</strong> its individual parts. For example, in a word such as fototoestel<br />
'photo camera', OUTPUT - OUTPUT CORRESPONDENCE dem<strong>and</strong>s that <strong>the</strong><br />
rhythmic structure <strong>of</strong> its part tóestel 'camera' with a stressed first syllable is<br />
reflected in <strong>the</strong> rhythmic structure <strong>of</strong> <strong>the</strong> output. In o<strong>the</strong>r words, OUTPUT -<br />
OUTPUT CORRESPONDENCE prefers fótotòestel, with secondary stress on toe,<br />
to fótotoestèl, with secondary stress on stel.<br />
Whereas <strong>the</strong> normal patterns in <strong>and</strong>ante speech satisfy OUTPUT -<br />
OUTPUT CORRESPONDENCE, <strong>the</strong> preference for triplet patterns in fast speech<br />
is accounted for by means <strong>of</strong> dominance <strong>of</strong> <strong>the</strong> markedness constraint,<br />
FOOT REPULSION, as illustrated in Table 2. 2<br />
Table 2. Rhythmic restructuring in language<br />
a. ranking in <strong>and</strong>ante speech:<br />
constraints →<br />
fototoestel<br />
c<strong>and</strong>idates ↓<br />
OUTPUT - OUTPUT<br />
CORRESPONDENCE<br />
��� PARSE-σ<br />
� (fóto)(tòestel) *<br />
(fóto)toe(stèl) *! *<br />
b. ranking in allegro speech:<br />
constraints →<br />
fototoestel<br />
c<strong>and</strong>idates ↓<br />
(fóto)(tòestel) *!<br />
OUTPUT - OUTPUT<br />
���<br />
CORRESPONDENCE<br />
� (fóto)toe(stèl) * *<br />
PARSE-σ<br />
Dutch is described as a trochaic language (Neijt <strong>and</strong> Zonneveld, 1982).<br />
Table 2a shows a preference for an alternating rhythm. The dactyl pattern<br />
as preferred in Table 2b, however, is a very common rhythmic pattern <strong>of</strong><br />
prosodic words in languages such as Estonian <strong>and</strong> Cayuvava: every strong<br />
syllable alternates with two weak syllables (cf. Kager, 1994). We assume<br />
that <strong>the</strong> rhythm grammar, i.e. constraint ranking, <strong>of</strong> Dutch allegro speech
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 187<br />
resembles <strong>the</strong> grammar <strong>of</strong> <strong>the</strong>se languages. In <strong>the</strong> next section we will<br />
explore whe<strong>the</strong>r we can find empirical evidence for our hypo<strong>the</strong>sis.<br />
4. Method<br />
To find out whe<strong>the</strong>r people indeed prefer triplet patterns in allegro speech,<br />
we ran a pilot experiment in which we tried to elicit fast speech. Six<br />
subjects participated in a multiple-choice quiz in which <strong>the</strong>y competed each<br />
o<strong>the</strong>r in answering twenty simple questions as quickly as possible. In this<br />
way, we expected <strong>the</strong>m to speak fast without concentrating too much on<br />
<strong>the</strong>ir own speech. In Table 3 one <strong>of</strong> <strong>the</strong> quiz items is depicted.<br />
Table 3. Quiz item<br />
Q4 President Bush is een typische ‘President Bush is a typical ’<br />
A1 intellectueel ‘intellectual’<br />
A2 amerikaan ‘American’<br />
A3 taalkundige ‘linguist’<br />
We categorized <strong>the</strong> obtained data as allegro speech. As a second task <strong>the</strong><br />
subjects were asked to read out <strong>the</strong> answers at a normal speaking rate<br />
embedded in <strong>the</strong> sentence ik spreek nu het woord … uit 'now I pronounce<br />
<strong>the</strong> word … '. This normal speaking rate generally means that <strong>the</strong> subjects<br />
will produce <strong>the</strong> words at a rate <strong>of</strong> approximately 180 words per minute,<br />
which we categorize as <strong>and</strong>ante speech. All data were recorded on minidisk<br />
in a soundpro<strong>of</strong> studio <strong>and</strong> normalized in CoolEdit in or<strong>der</strong> to improve <strong>the</strong><br />
signal-noise (S/N) ratio. Normalizing to 100% yields an S/N ratio<br />
approaching 0 dB.<br />
Six trained listeners judged <strong>the</strong> data auditively <strong>and</strong> indicated where <strong>the</strong>y<br />
perceived secondary stress. After this auditive analysis <strong>the</strong> data were<br />
phonetically analyzed in PRAAT (Boersma <strong>and</strong> Weenink, 1992). We<br />
compared <strong>the</strong> <strong>and</strong>ante <strong>and</strong> allegro data by measuring duration, pitch,<br />
intensity, spectral balance <strong>and</strong> rhythmic timing (Sluijter, 1995; Couper-<br />
Kuhlen, 1993; Cummins & Port, 1998; Quené & Port, 2002; a.o.). Sluijter<br />
claims that, respectively, duration <strong>and</strong> spectral balance are <strong>the</strong> main<br />
correlates <strong>of</strong> primary stress. In our experiment, we are concerned with<br />
secondary stress.
188 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
For <strong>the</strong> duration measurements, <strong>the</strong> rhymes <strong>of</strong> <strong>the</strong> relevant syllables<br />
were observed. For example, in <strong>the</strong> allegro style answer A2 amerikaan in<br />
Table 3, we measured <strong>the</strong> first two rhymes <strong>and</strong> compared <strong>the</strong> values in<br />
Msec. with <strong>the</strong> values for <strong>the</strong> same rhymes at <strong>the</strong> <strong>and</strong>ante rate. In or<strong>der</strong> to<br />
make this comparison valid, we equalized <strong>the</strong> total durations <strong>of</strong> both<br />
realizations by multiplying <strong>the</strong> duration <strong>of</strong> <strong>the</strong> allegro with a so-called<br />
'acceleration factor', i.e. <strong>the</strong> duration <strong>of</strong> <strong>the</strong> <strong>and</strong>ante version divided by <strong>the</strong><br />
duration <strong>of</strong> <strong>the</strong> allegro version. According to Eefting <strong>and</strong> Rietveld (1989)<br />
<strong>and</strong> Rietveld <strong>and</strong> Van Heuven (1997), <strong>the</strong> just noticeable difference for<br />
duration is 4,5%. If <strong>the</strong> difference in duration between <strong>the</strong> <strong>and</strong>ante <strong>and</strong> <strong>the</strong><br />
allegro realization did not exceed this threshold, we consi<strong>der</strong>ed <strong>the</strong><br />
realizations as examples <strong>of</strong> <strong>the</strong> same speech rate <strong>and</strong> neglected <strong>the</strong>m for<br />
fur<strong>the</strong>r analysis.<br />
For <strong>the</strong> pitch measurements, we took <strong>the</strong> value in Hz in <strong>the</strong> middle <strong>of</strong><br />
<strong>the</strong> vowel. The just noticeable difference for pitch is 2,5% ('t Hart et al,<br />
1990). For <strong>the</strong> intensity measurements, we registered <strong>the</strong> mean value in dB<br />
<strong>of</strong> <strong>the</strong> whole syllable.<br />
The next parameter we consi<strong>der</strong>ed concerns spectral balance. Sluijter<br />
(1995) claims that <strong>the</strong> spectral balance <strong>of</strong> <strong>the</strong> vowel <strong>of</strong> a stressed syllable is<br />
characterized by more perceived loudness in <strong>the</strong> higher frequency region,<br />
because <strong>of</strong> <strong>the</strong> changes in <strong>the</strong> source spectrum due to a more pulse-like<br />
shape <strong>of</strong> <strong>the</strong> glottal waveform. The vocal effort, which is used for stress,<br />
generates a strongly asymmetrical glottal pulse. As a result <strong>of</strong> <strong>the</strong> shortened<br />
closing phase, <strong>the</strong>re is an increase <strong>of</strong> intensity around <strong>the</strong> four formants in<br />
<strong>the</strong> frequency region above 500 Hz. Following Sluijter (1995) we compared<br />
<strong>the</strong> differences in intensity <strong>of</strong> <strong>the</strong> higher <strong>and</strong> lower frequencies <strong>of</strong> <strong>the</strong><br />
relevant syllables in both tempos.<br />
Finally, we consi<strong>der</strong>ed rhythmic timing. The idea is that <strong>the</strong> beats in<br />
speech are separated from each o<strong>the</strong>r at an approximately equal distance<br />
independent <strong>of</strong> <strong>the</strong> speech rate. In o<strong>the</strong>r words, a speaker more or less<br />
follows an imaginary metronome. If he/she speaks faster, more melodic<br />
content will be placed between beats, which results in a shift <strong>of</strong> secondary<br />
stress. This hypo<strong>the</strong>sis will be confirmed if <strong>the</strong> distance between <strong>the</strong><br />
stressed syllables in <strong>the</strong> <strong>and</strong>ante realization <strong>of</strong> an item, e.g. stu <strong>and</strong> toe in<br />
studietoelage, approximates <strong>the</strong> distance between <strong>the</strong> stressed syllables in<br />
<strong>the</strong> allegro realization <strong>of</strong> <strong>the</strong> same item, e.g. stu <strong>and</strong> la. If <strong>the</strong> quotient <strong>of</strong> <strong>the</strong><br />
<strong>and</strong>ante beat interval duration divided by <strong>the</strong> allegro beat interval duration<br />
approximates 1, we expect perceived restructuring.
5. Results<br />
5.1. Auditive analysis<br />
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 189<br />
Before we can present an auditive analysis <strong>of</strong> <strong>the</strong> data, we have to find out<br />
whe<strong>the</strong>r or not <strong>the</strong> quiz design was successful. The results show that <strong>the</strong><br />
quiz indeed triggers faster speech by all subjects. Figure 2 shows <strong>the</strong>ir<br />
acceleration factors. Subjects 1, 2 <strong>and</strong> 4 turned out to be <strong>the</strong> best<br />
accelerating speakers, whereas subjects 3, 5 <strong>and</strong> 6 showed less difference in<br />
duration between <strong>and</strong>ante <strong>and</strong> allegro realizations. The mean acceleration<br />
factor for <strong>the</strong> three fast speakers is 1.31, whereas <strong>the</strong> mean acceleration<br />
factor for <strong>the</strong> three slow speakers is 1.13.<br />
factor<br />
1.4<br />
1.35<br />
1.3<br />
1.25<br />
1.2<br />
1.15<br />
1.1<br />
1.05<br />
1<br />
acceleration factors<br />
p1 p2 p3 p4 p5 p6<br />
subjects<br />
Figure 2. Acceleration factors <strong>of</strong> all subjects<br />
Figure 3 shows <strong>the</strong> mean durations <strong>of</strong> <strong>the</strong> items at both speech rates. It<br />
shows that <strong>the</strong> best accelerating speakers are also <strong>the</strong> fastest speakers. We<br />
expect to find more restructured patterns for <strong>the</strong>se speakers, mainly subjects<br />
1 <strong>and</strong> 4, in comparison to <strong>the</strong> slower speakers, such as subjects 3 <strong>and</strong> 6.
190 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
seconds<br />
1.15<br />
1.05<br />
0.95<br />
0.85<br />
0.75<br />
0.65<br />
mean word durations<br />
p1 p2 p3 p4 p5 p6<br />
subjects<br />
Figure 3. Mean word durations<br />
<strong>and</strong>ante<br />
Figure 4 shows that most subjects prefer patterns in which from a<br />
phonological point <strong>of</strong> view markedness constraints dominate <strong>the</strong><br />
correspondence constraints at both rates for right <strong>and</strong> left shift data, but not<br />
for beat reduction data. There are slightly more restructured patterns in<br />
allegro tempo, although <strong>the</strong> differences are quite small.<br />
Number<br />
40<br />
35<br />
30<br />
25<br />
20<br />
15<br />
10<br />
5<br />
0<br />
Right Shifts<br />
(N=42)<br />
Left Shifts<br />
(N=36)<br />
Types<br />
All Subjects<br />
Beat<br />
Reductions<br />
(N=42)<br />
Figure 4. All subjects: Number <strong>of</strong> restructured items per type<br />
When we take <strong>the</strong> results <strong>of</strong> two fast subjects apart, subjects 1 <strong>and</strong> 4, we<br />
observe a stronger preference for restructuring in allegro speech <strong>and</strong> no<br />
restructuring in <strong>and</strong>ante speech, as shown in Figure 5. In o<strong>the</strong>r words, <strong>the</strong><br />
fast subjects display both a greater difference in word durations in <strong>and</strong>ante<br />
<strong>and</strong> allegro speech, <strong>and</strong> more variability in <strong>the</strong>ir speech patterns due to<br />
tempo than <strong>the</strong> slow subjects do.<br />
allegro<br />
Not Shifted Andante<br />
Shifted Andante<br />
Shifted Allegro<br />
Not Shifted Allegro
Number<br />
16<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
0<br />
Right Shifts<br />
(N=14)<br />
Fast Subjects<br />
Left Shifts<br />
(N=12)<br />
Types<br />
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 191<br />
Beat<br />
Reductions<br />
(N=14)<br />
Not Shifted Andante<br />
Shifted Andante<br />
Shifted Allegro<br />
Not Shifted Allegro<br />
Figure 5. Fast subjects: Number <strong>of</strong> restructured items per type<br />
Obviously, <strong>the</strong> preference for restructuring <strong>the</strong> rhythmic pattern in allegro<br />
speech is not an absolute preference. Sometimes restructuring does not take<br />
place in allegro speech, but on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong> restructured patterns also<br />
show up in <strong>and</strong>ante speech. 3 Some items were realized with <strong>the</strong> same<br />
rhythmic pattern irrespective <strong>of</strong> <strong>the</strong> tempo. Therefore, we also looked at <strong>the</strong><br />
word pairs with a different rhythmic pattern in both tempos for each<br />
subject. We observe that <strong>the</strong> relatively fast speakers p1, p2 <strong>and</strong> p4, show<br />
<strong>the</strong> expected pattern according to our hypo<strong>the</strong>sis, which means that <strong>the</strong>y<br />
show a restructured pattern in allegro tempo, as shown in Figure 6 for <strong>the</strong><br />
right shifts.<br />
number<br />
4<br />
3<br />
2<br />
1<br />
0<br />
Right Shifts Word Pairs<br />
p1 p2 p3 p4 p5 p6<br />
subjects<br />
Expected combinations<br />
Counterexamples<br />
Figure 6. Right Shifts: Expected combinations<br />
Two <strong>of</strong> <strong>the</strong> relatively slow speakers, p3 <strong>and</strong> p6, show one counterexample<br />
each, where <strong>the</strong> subject prefers <strong>the</strong> restructured patterns in <strong>and</strong>ante tempo.<br />
The o<strong>the</strong>r slow speaker, P5, displays no different patterns in <strong>and</strong>ante <strong>and</strong>
192 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
allegro at all. Clearly, we have two different groups <strong>of</strong> speakers <strong>and</strong> this<br />
observation streng<strong>the</strong>ns our claim that restructuring relates to speech rate.<br />
Some items, such as hobbywerkruimte (Type 1) 'hobby room', never<br />
show a stress shift <strong>and</strong> o<strong>the</strong>r items, such as viriliteit (Type 2) ‘virility’,<br />
prefer <strong>the</strong> shifted pattern in both tempos for all subjects. Possibly, <strong>the</strong><br />
syllable structure plays an important role; open syllables seem to lose stress<br />
more easily than closed ones.<br />
5.2. Acoustic analysis<br />
In <strong>the</strong> current state <strong>of</strong> phonological research, embodied in e.g. laboratory<br />
phonology, much value is set on acoustic evidence for phonological<br />
analyses. Studies such as Sluijter (1995) <strong>and</strong> Sluijter <strong>and</strong> Van Heuven<br />
(1996) provide acoustic correlates for primary stress. In our study we are<br />
concerned with beat reduction <strong>and</strong> secondary stress shifts <strong>and</strong> we won<strong>der</strong><br />
whe<strong>the</strong>r or not <strong>the</strong> same acoustic correlates hold for secondary stress.<br />
Shattuck Hufnagel et al (1994) <strong>and</strong> Cooper <strong>and</strong> Eady (1986) do not find<br />
acoustic correlates <strong>of</strong> rhythmic stress at all. They claim that it is not entirely<br />
clear which acoustic correlates are appropriate to measure, since <strong>the</strong>se<br />
correlates are dependent on <strong>the</strong> relative strength <strong>of</strong> <strong>the</strong> syllables <strong>of</strong> an<br />
utterance. The absolute values <strong>of</strong> a single syllable can hardly be compared<br />
without reference to <strong>the</strong>ir context <strong>and</strong> <strong>the</strong> intonation pattern <strong>of</strong> <strong>the</strong> complete<br />
phrase. Huss (1978) claims that some cases <strong>of</strong> perceived rhythmic stress<br />
shift may be perceptual ra<strong>the</strong>r than acoustic in nature. Grabe <strong>and</strong> Warren<br />
(1995) also suggest that stress shifts can only be perceived in rhythmic<br />
contexts. In isolation, <strong>the</strong> prominence patterns are unlikely to be judged<br />
reliably. In <strong>the</strong> remain<strong>der</strong> <strong>of</strong> this paper we try to find out if we can support<br />
one <strong>of</strong> <strong>the</strong>se lines <strong>of</strong> reasoning. In o<strong>the</strong>r words, are we able to support our<br />
perceived rhythmic variability with a phonetic analysis? Therefore, we<br />
measured <strong>the</strong> duration, pitch, intensity, spectral balance <strong>and</strong> rhythmic<br />
timing <strong>of</strong> <strong>the</strong> relevant syllables as realized by subject P1.<br />
Because Dutch is a quantity-sensitive language, <strong>the</strong> duration <strong>of</strong> <strong>the</strong><br />
relevant syllable rhymes was consi<strong>der</strong>ed. <strong>On</strong>sets do not contribute to <strong>the</strong><br />
weight <strong>of</strong> a syllable. In Figure 7, <strong>the</strong> duration analysis is shown for Type 2<br />
data (left shifts). The four columns indicate, respectively, <strong>the</strong> duration <strong>of</strong><br />
<strong>the</strong> rhyme <strong>of</strong> <strong>the</strong> first <strong>and</strong> second syllable in <strong>and</strong>ante speech, <strong>and</strong> <strong>the</strong><br />
duration <strong>of</strong> <strong>the</strong> first <strong>and</strong> second one in allegro speech. According to Sluijter<br />
(1995), duration is <strong>the</strong> main correlate <strong>of</strong> primary stress. As a starting point,
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 193<br />
we adopt her claim for our analysis <strong>of</strong> secondary stress. Our measurements<br />
would confirm our hypo<strong>the</strong>sis <strong>and</strong> our auditive analysis, if <strong>the</strong> second<br />
column were higher than <strong>the</strong> first one <strong>and</strong> if <strong>the</strong> fourth column were lower<br />
than <strong>the</strong> third one. In that case, <strong>the</strong> subject would realize a word such as<br />
perfectionist as perfèctioníst in <strong>and</strong>ante tempo <strong>and</strong> as pèrfectioníst in<br />
allegro tempo.<br />
In <strong>the</strong> <strong>and</strong>ante tempo, three out <strong>of</strong> six items show <strong>the</strong> dominant<br />
correspondence pattern <strong>and</strong> in <strong>the</strong> allegro tempo, four out <strong>of</strong> six items show<br />
<strong>the</strong> dominant markedness pattern. That is hardly a preference <strong>and</strong> it does<br />
not confirm our auditive analysis <strong>of</strong> <strong>the</strong> same data. Fur<strong>the</strong>rmore, if we<br />
consi<strong>der</strong> <strong>the</strong> word pairs with different patterns, <strong>the</strong>re is only one pair that<br />
has <strong>the</strong> ideal ratio: <strong>the</strong> patterns <strong>of</strong> amerikaan.<br />
seconds<br />
0.16<br />
0.14<br />
0.12<br />
0.1<br />
0.08<br />
0.06<br />
0.04<br />
0.02<br />
0<br />
perfectionist<br />
amerikaan<br />
duration rhymes<br />
piraterij<br />
verbaliseren<br />
item<br />
banaliteit<br />
Figure 7. Duration (Left Shifts by Subject P1)<br />
If duration does not enable us to confirm our auditive findings, maybe pitch<br />
is <strong>the</strong> main stress correlate for this speaker. However, pitch measurements<br />
reveal <strong>the</strong> same fuzzy result as <strong>the</strong> duration measurements. Again, only one<br />
pattern confirms <strong>the</strong> auditive analysis. This time it is not <strong>the</strong> item<br />
amerikaan, but <strong>the</strong> item perfectionist. Moreover, <strong>the</strong> differences in pitch in<br />
this item do not exceed <strong>the</strong> threshold <strong>of</strong> <strong>the</strong> 2.5%, which is <strong>the</strong> just<br />
noticeable difference for pitch. We also analyzed <strong>the</strong> mean intensity value<br />
<strong>of</strong> <strong>the</strong> relevant vowels without recognizable patterns between allegro <strong>and</strong><br />
<strong>and</strong>ante style. These results support <strong>the</strong> analyses <strong>of</strong> Sluijter (1995) <strong>and</strong><br />
Sluijter <strong>and</strong> Van Heuven (1996), who also claim that <strong>the</strong> intensity<br />
parameter does not contribute much to <strong>the</strong> perception <strong>of</strong> stress.<br />
Next, we consi<strong>der</strong>ed <strong>the</strong> spectral balance. In or<strong>der</strong> to rule out <strong>the</strong><br />
influence <strong>of</strong> <strong>the</strong> o<strong>the</strong>r parameters, we monotonized <strong>the</strong> data for volume <strong>and</strong><br />
viriliteit<br />
<strong>and</strong>ante r1<br />
<strong>and</strong>ante r2<br />
allegro r1<br />
allegro r2
194 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
pitch. Then we selected <strong>the</strong> relevant vowels <strong>and</strong> analyzed <strong>the</strong>m as a<br />
cochleagram in PRAAT. The cochleagram simulates <strong>the</strong> way <strong>the</strong> tympanic<br />
membrane functions, in o<strong>the</strong>r words <strong>the</strong> way in which we perceive sounds.<br />
In Figure 8 we show two cochleagrams <strong>of</strong> <strong>the</strong> vowel [a] in <strong>the</strong> fourth<br />
syllable <strong>of</strong>, respectively, stúdietòelage 'study grant' (Type 1) in <strong>and</strong>ante<br />
tempo <strong>and</strong> stúdietoelàge in allegro tempo. This item was taken from a prestudy.<br />
The allegro data show <strong>the</strong> expected increased perceived loudness in<br />
<strong>the</strong> higher frequencies, indicated by means <strong>of</strong> shades <strong>of</strong> gray; <strong>the</strong> darker<br />
gray <strong>the</strong> more perceived loudness.<br />
25<br />
20<br />
15<br />
10<br />
5<br />
0<br />
0 0.169371<br />
Time (s)<br />
Figure 8. Cochleagrams <strong>of</strong> [�] in studietoelage<br />
0<br />
0 0.143209<br />
Time (s)<br />
The right cochleagram (stressed [a]) in Figure 8 shows increased perceived<br />
loudness in <strong>the</strong> regions <strong>of</strong> approximately 5 to 22 Bark in <strong>the</strong> allegro version<br />
<strong>of</strong> [a] in comparison with <strong>the</strong> left cochleagram (unstressed [a]). This<br />
confirms <strong>the</strong> results <strong>of</strong> <strong>the</strong> study <strong>of</strong> primary stress in Sluijter (1995). If we<br />
convert this perceptive, almost logarithmic, Bark scale into its linear<br />
counterpart, <strong>the</strong> Hertz scale, this area correlates with <strong>the</strong> frequency region<br />
<strong>of</strong> 3 to 10 kHz.<br />
In or<strong>der</strong> to measure perceived secondary stress, we will measure <strong>the</strong><br />
relative loudness in <strong>the</strong> different frequency regions in Phon. 4 According to<br />
Sluijter (1995) stressed vowels have increased loudness above 500 Hz<br />
compared to <strong>the</strong> same vowel in an unstressed position. This can be shown if<br />
we take a point in time from both cochleagrams in Figure 8 in which <strong>the</strong> F1<br />
reaches its highest value (following Sluijter, 1995). In Figure 9 <strong>the</strong> values<br />
in Phon are depicted for <strong>the</strong>se points <strong>and</strong> plotted against <strong>the</strong> Bark values in<br />
25 steps.<br />
25<br />
20<br />
15<br />
10<br />
5
phon<br />
60<br />
50<br />
40<br />
30<br />
20<br />
1<br />
stressed <strong>and</strong> unstressed [a]<br />
5<br />
9<br />
13<br />
bark<br />
Figure 9. Loudness in Phon<br />
17<br />
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 195<br />
21<br />
25<br />
<strong>and</strong>ante<br />
a lle g ro<br />
The white line in Figure 9 indicates <strong>the</strong> pattern <strong>of</strong> <strong>the</strong> allegro stressed [a] in<br />
studietoelage <strong>and</strong> <strong>the</strong> black line indicates <strong>the</strong> pattern <strong>of</strong> <strong>the</strong> <strong>and</strong>ante<br />
unstressed [a]. We see increased loudness in <strong>the</strong> region <strong>of</strong> 13 to 21 Bark,<br />
which correlates with <strong>the</strong> most sensitive region <strong>of</strong> our ear. The mean Phon<br />
value in Figure 9 between 5 <strong>and</strong> 21 Bark is 43.6 Phon for <strong>the</strong> <strong>and</strong>ante<br />
unstressed [a] <strong>and</strong> 47.4 Phon for <strong>the</strong> allegro stressed [a]; a mean difference<br />
<strong>of</strong> 3.8 Phon.<br />
Now, let us see whe<strong>the</strong>r or not we can find similar results for our subject<br />
P1. Figure 10 shows that <strong>the</strong> spectral balance confirms <strong>the</strong> leftward stress<br />
shift we perceived in <strong>the</strong> allegro realization <strong>of</strong> amerikaan. The first syllable<br />
vowel in allegro tempo is characterized by more loudness in <strong>the</strong> higher<br />
frequency regions than its <strong>and</strong>ante counterpart. In <strong>the</strong> second syllable vowel<br />
it is just <strong>the</strong> o<strong>the</strong>r way around.<br />
Phon<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
Stressed <strong>and</strong> unstressed [a] in<br />
[a]merikaan<br />
1<br />
4<br />
7<br />
10<br />
13<br />
16<br />
19<br />
22<br />
25<br />
Bark<br />
<strong>and</strong>ante<br />
allegro<br />
Stressed <strong>and</strong> unstressed [e] in<br />
a[me]rikaan<br />
Figure 10. Spectral balance comparison <strong>of</strong> <strong>the</strong> first two vowels <strong>of</strong> amerikaan<br />
Unfortunately, not all spectral balance data confirm our auditive analysis.<br />
For example, we claimed that <strong>the</strong> pitch analysis <strong>of</strong> <strong>the</strong> stress shift in<br />
Phon<br />
60<br />
40<br />
20<br />
0<br />
1<br />
4<br />
7<br />
10<br />
13<br />
16<br />
19<br />
22<br />
25<br />
Bark<br />
<strong>and</strong>ante<br />
allegro
196 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
perfectionist did confirm our auditive analysis. Therefore, we expected<br />
more loudness in <strong>the</strong> allegro realization <strong>of</strong> <strong>the</strong> first vowel <strong>and</strong> less loudness<br />
in <strong>the</strong> allegro realization <strong>of</strong> <strong>the</strong> second vowel, but it appeared that <strong>the</strong>re is<br />
relatively more loudness in <strong>the</strong> <strong>and</strong>ante realization <strong>of</strong> per. This result<br />
contradicts our auditive <strong>and</strong> our pitch analysis.<br />
We have to conclude that <strong>the</strong> different phonetic analyses contradict each<br />
o<strong>the</strong>r. Sometimes <strong>the</strong> perceived stress shift is characterized by a longer<br />
duration <strong>of</strong> <strong>the</strong> stressed syllable; sometimes a relatively higher pitch<br />
characterizes it. The results <strong>of</strong> our spectral balance analysis show that <strong>the</strong><br />
differences in loudness pattern with differences in duration. In our<br />
perceived stress shift in allegro perfectionist, pitch turned out to be <strong>the</strong><br />
decisive correlate, whereas duration <strong>and</strong> spectral balance measurements<br />
indicated no shift at all. <strong>On</strong> <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>, <strong>the</strong> perceived shift in allegro<br />
amerikaan was confirmed by <strong>the</strong> duration <strong>and</strong> spectral balance analyses<br />
toge<strong>the</strong>r, whereas pitch measurements indicated <strong>the</strong> opposite pattern. For<br />
most perceived stress shifts, however, <strong>the</strong> acoustic correlates did not give<br />
any clue.<br />
Finally, we will consi<strong>der</strong> whe<strong>the</strong>r <strong>the</strong> perception <strong>of</strong> restructuring<br />
depends on rhythmic timing. Just as in music, speech can be divided into a<br />
melodic string <strong>and</strong> a rhythmic string as partly independent entities. With<br />
respect to speech, <strong>the</strong> melodic string seems to be more flexible than <strong>the</strong><br />
rhythmic one. Imagine that <strong>the</strong> rhythm constitutes a kind <strong>of</strong> metronome<br />
pulse to which <strong>the</strong> melodic content has to be aligned. The listener expects<br />
prominent syllables to occur with beats. This behavior is formulated as <strong>the</strong><br />
Equal Spacing Constraint: prominent vowel onsets are attracted to<br />
periodically spaced temporal locations (Couper-Kuhlen, 1993; Cummins &<br />
Port, 1998; Quené & Port, 2002; a.o.). Dependent on speech rate <strong>the</strong><br />
number <strong>of</strong> intervening syllables between beats may differ. Suppose <strong>the</strong> beat<br />
interval is constant at 300 msec., <strong>the</strong>re will be more linguistic material in<br />
between in allegro speech, e.g. <strong>the</strong> two syllables die <strong>and</strong> toe in<br />
stúdietoelàge, than in <strong>and</strong>ante speech, e.g. only one syllable die in<br />
stúdietòelage.<br />
If indeed <strong>the</strong> perception <strong>of</strong> secondary stress shifts depends on rhythmic<br />
timing, i.e. <strong>the</strong> beat interval between prominent syllables in <strong>and</strong>ante <strong>and</strong><br />
allegro speech is approximately equal, than we expect that <strong>the</strong> duration<br />
quotient <strong>of</strong> <strong>the</strong> interval between, for example, stu <strong>and</strong> toe in <strong>the</strong> <strong>and</strong>ante<br />
realization <strong>of</strong> studietoelage <strong>and</strong> stu <strong>and</strong> la in <strong>the</strong> allegro realization<br />
approximates 1.
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 197<br />
In our pre-study, <strong>the</strong> interval between <strong>the</strong> vowel onsets <strong>of</strong> <strong>the</strong> first <strong>and</strong><br />
third syllable in studietoelage (<strong>and</strong>ante) is 0.358 sec, whereas <strong>the</strong> interval<br />
between <strong>the</strong> first <strong>and</strong> <strong>the</strong> fourth syllable in <strong>the</strong> allegro realization <strong>of</strong> <strong>the</strong><br />
same word is 0.328 sec. This means that <strong>the</strong> duration quotient is 1.091,<br />
which indeed approximates 1. In o<strong>the</strong>r words, this example supports <strong>the</strong><br />
idea <strong>of</strong> <strong>the</strong> Equal Spacing Constraint.<br />
Does <strong>the</strong> same result hold for our present data? We measured <strong>the</strong> beat<br />
intervals between all possible stress placement sites for all six subjects.<br />
Figure 11 depicts <strong>the</strong> duration quotients for subject 1. Figure 12 shows <strong>the</strong><br />
beat intervals <strong>of</strong> <strong>the</strong> same data. It depicts as well <strong>the</strong> duration interval<br />
between <strong>the</strong> first <strong>and</strong> <strong>the</strong> third, as <strong>the</strong> first <strong>and</strong> fourth syllable for both<br />
speech rates. We expect restructuring for those data in which <strong>the</strong> line <strong>of</strong> <strong>the</strong><br />
first to third syllable interval (<strong>and</strong>ante (black line)) coincides with <strong>the</strong> line<br />
<strong>of</strong> <strong>the</strong> first to fourth syllable interval (allegro (white line)).<br />
quotient<br />
1.1<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
studietoelage<br />
Right Shifts beat intervals <strong>and</strong>ante:allegro<br />
wegwerpaanst...<br />
trimesterindeling<br />
kamervoorzitter<br />
hobbywerkruimte<br />
gemeenteinschr...<br />
Figure 11. Quotient beat intervals <strong>of</strong> Subject P1<br />
beat intervals in sec<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
studietoelage<br />
wegwerpaansteker<br />
trimesterindeling<br />
Figure 12. Beat intervals <strong>of</strong> Subject P1<br />
winkelopheffing<br />
Right Shifts Beat Intervals P1<br />
kamervoorzitter<br />
hobbywerkruimte<br />
gemeenteinschrijving<br />
winkelopheffing<br />
quotient<br />
<strong>and</strong>ante s1-3<br />
<strong>and</strong>ante s1-4<br />
allegro s1-4<br />
allegro s1-3
198 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
The Figures 11 <strong>and</strong> 12 indicate that <strong>the</strong> relevant beat intervals <strong>of</strong> <strong>the</strong> items<br />
1, 4 <strong>and</strong> 7, studietoelage 'study grant', kamervoorzitter 'chairman <strong>of</strong> <strong>the</strong><br />
House <strong>of</strong> Parliament' <strong>and</strong> winkelopheffing 'closing down <strong>of</strong> a shop',<br />
respectively, coincide. In o<strong>the</strong>r words, we expect to hear restructuring in<br />
exactly <strong>the</strong>se three items.<br />
Unfortunately, our auditive analysis indicates only attested<br />
combinations <strong>of</strong> restructuring in items 2 <strong>and</strong> 6: wegwerpaansteker<br />
'disposable lighter' <strong>and</strong> gemeente-inschrijving 'municipal registration',<br />
respectively. Obviously, rhythmic timing is not <strong>the</strong> decisive characteristic<br />
<strong>of</strong> perceived restructuring in allegro speech ei<strong>the</strong>r.<br />
6. Discussion <strong>and</strong> Conclusion<br />
In section 4, we presented our phonological account <strong>of</strong> <strong>the</strong> restructuring<br />
within <strong>the</strong> framework <strong>of</strong> OT. Our main conclusion is that phonetic<br />
compression cannot be <strong>the</strong> sole explanation <strong>of</strong> <strong>the</strong> different rhythm<br />
patterns. Although <strong>the</strong> results cannot really confirm our hypo<strong>the</strong>sis that<br />
<strong>the</strong>re are different grammars, i.e. constraint rankings for different rates <strong>of</strong><br />
speaking, <strong>the</strong>re seems to be something that relates to speech rate. The fast<br />
speakers display different grammars, i.e. constraint rankings, for different<br />
rates <strong>of</strong> speaking. In <strong>the</strong>ir <strong>and</strong>ante tempo, correspondence constraints<br />
prevail, whereas in allegro tempo markedness constraints dominate <strong>the</strong><br />
correspondence ones. These preferences resemble <strong>the</strong> preferences <strong>of</strong><br />
<strong>and</strong>ante <strong>and</strong> allegro music. In both disciplines clashes are avoided in<br />
allegro tempo by means <strong>of</strong> enlarging <strong>the</strong> distances between beats.<br />
In section 5, we attempted to confirm our phonological account with a<br />
phonetic analysis. Unfortunately, <strong>the</strong> phonetic correlates <strong>of</strong> stress -<br />
duration, pitch, intensity <strong>and</strong> spectral balance - do not show <strong>the</strong> expected<br />
<strong>and</strong> perceived differences in rhythm patterns in all pairs. Sluijter (1995)<br />
found out that duration is <strong>the</strong> main correlate <strong>of</strong> primary stress with spectral<br />
balance as an important second characteristic. In our analysis, however,<br />
nei<strong>the</strong>r differences in duration nor differences in spectral balance could<br />
identify secondary stress. Therefore, we have to conclude that our analysis<br />
supports earlier work by Shattuck Hufnagel et al (1994), Cooper <strong>and</strong> Eady<br />
(1986), Huss (1978) <strong>and</strong> Grabe <strong>and</strong> Warren (1995), who all claim that<br />
acoustic evidence for secondary stress cannot be found unambiguously.<br />
Although we did find some differences in duration, spectral balance or<br />
pitch, <strong>the</strong>se differences were not systematically found in all pairs in which
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 199<br />
we perceived rhythmic variability. Finally, we discussed rhythmic timing as<br />
a cue for variable patterns. However, <strong>the</strong> hypo<strong>the</strong>sis that <strong>the</strong> duration<br />
between prominent syllables is approximately equal in both <strong>and</strong>ante <strong>and</strong><br />
allegro speech was not confirmed by <strong>the</strong> auditive analysis <strong>of</strong> <strong>the</strong> data. It<br />
seems that rhythmic restructuring is more a matter <strong>of</strong> perception than <strong>of</strong><br />
production. At this point, <strong>the</strong> question remains: are we fooled by our brains<br />
<strong>and</strong> is <strong>the</strong>re no phonetic correlate <strong>of</strong> <strong>the</strong> perceived phonological stress shifts<br />
in <strong>the</strong> acoustic signal or do we have to conclude that <strong>the</strong> real phonetic<br />
correlate <strong>of</strong> secondary stress has yet to be found?<br />
Notes<br />
1 This paper is an extension <strong>of</strong> our paper "Restructuring <strong>the</strong> melodic content <strong>of</strong><br />
feet", which is submitted to <strong>the</strong> proceedings <strong>of</strong> <strong>the</strong> 9 th International <strong>Phonology</strong><br />
Meeting: Structure <strong>and</strong> melody, Vienna 2002. We wish to thank Grzegorz<br />
Dogil, Hidetoshi Shiraishi plus <strong>the</strong> participants <strong>of</strong> <strong>the</strong> 9 th International<br />
<strong>Phonology</strong> Meeting, Vienna 2002 <strong>and</strong> <strong>the</strong> participants <strong>of</strong> <strong>the</strong> 11 th Manchester<br />
<strong>Phonology</strong> Meeting, Manchester 2003 for <strong>the</strong>ir useful comments. We are also<br />
grateful to Sible Andringa, Nynke van den Bergh, Gerl<strong>of</strong> Bouma, John Hoeks,<br />
Jack Hoeksema, W<strong>and</strong>er Lowie, Dirk-Bart den Ouden, Joanneke Prenger,<br />
Ingeborg Prinsen, Femke Wester for participating in our experiment. We<br />
especially thank Wilbert Heeringa <strong>and</strong> Hugo Quené for supplying us with <strong>the</strong><br />
PRAAT scripts that we could use for our spectral balance <strong>and</strong> rhythmic timing<br />
analyses.<br />
2 For reasons <strong>of</strong> clarity, we abstract from constraints such as FOOTBINARITY<br />
(FTBIN) <strong>and</strong> WEIGHT-TO-STRESS PRINCIPLE in Table 2. Although <strong>the</strong>se<br />
constraints play an important role in <strong>the</strong> Dutch stress system (cf. Gilbers &<br />
Jansen, 1996), <strong>the</strong> conflict between OUTPUT-OUTPUT CORRESPONDENCE <strong>and</strong><br />
FOOT REPULSION is essential for our present analysis.<br />
3 With respect to <strong>the</strong> phonological analysis <strong>of</strong> <strong>the</strong> data, we suggest a r<strong>and</strong>om<br />
ranking <strong>of</strong> weighed correspondence <strong>and</strong> markedness constraints. By means <strong>of</strong><br />
weighing constraints we adopt an OT variant that more or less resembles <strong>the</strong><br />
analyses in OT’s predecessor Harmonic Grammar (cf. Legendre, G., Y. Miyata &<br />
P. Smolensky, 1990). Note that we do not opt for a co-phonology for allegro-style<br />
speech in our analysis. In a co-phonology, <strong>the</strong> output <strong>of</strong> <strong>the</strong> <strong>and</strong>ante-style ranking<br />
is input or base for <strong>the</strong> allegro-style ranking. We opt for a r<strong>and</strong>om ranking with<br />
different preferences for allegro <strong>and</strong> <strong>and</strong>ante speech, because our data show<br />
variable rhythmic structures at both rates. Both rankings evaluate <strong>the</strong> same input<br />
form.
200 Maartje Schreu<strong>der</strong> <strong>and</strong> Dicky Gilbers<br />
4 The perceived loudness depends on <strong>the</strong> frequency <strong>of</strong> <strong>the</strong> tone. The Phon entity<br />
is defined using <strong>the</strong> 1kHz tone <strong>and</strong> <strong>the</strong> decibel scale. A pure sinus tone at any<br />
frequency with 100 Phon is as loud as a pure tone with 100 dB at 1kHz<br />
(Rietveld <strong>and</strong> Van Heuven, 1997: 199). We are most sensitive to frequencies<br />
around 3kHz. The hearing threshold rapidly rises around <strong>the</strong> lower <strong>and</strong> upper<br />
frequency limits, which are respectively about 20Hz <strong>and</strong> 16kHz.<br />
References<br />
Boersma, Paul, <strong>and</strong> David Weenink (1992-2002). PRAAT, phonetics by<br />
computer. Available at http://www.praat.org, University <strong>of</strong><br />
Amsterdam.<br />
Burzio, Luigi (1998). Multiple Correspondence. Lingua, 104: 79-109.<br />
Cooper, W., <strong>and</strong> J. Eady (1986). Metrical phonology in speech production.<br />
Journal <strong>of</strong> Memory <strong>and</strong> Language, 25: 369-384.<br />
Couper-Kuhlen, Elizabeth (1993). English speech rhythm: form <strong>and</strong> function in<br />
everyday verbal interaction. Benjamins, Amsterdam.<br />
Cummins, Fred, <strong>and</strong> Robert Port (1998). Rhythmic constraints on stress timing<br />
in English. Journal <strong>of</strong> <strong>Phonetics</strong>, 26(2): 145-171.<br />
Eefting, Wieke, <strong>and</strong> Toni Rietveld (1989). Just noticeable differences <strong>of</strong><br />
articulation rate at sentence level. Speech Communication, 8:<br />
355-351.<br />
Gilbers, Dicky, <strong>and</strong> Wouter Jansen (1996). Klemtoon en ritme in Optimality<br />
Theory, deel 1: ho<strong>of</strong>d-, neven-, samenstellings- en<br />
woordgroepsklemtoon in het Ne<strong>der</strong>l<strong>and</strong>s [Stress <strong>and</strong> rhythm in<br />
Optimality Theory, part 1: primary stress, secondary stress,<br />
compound stress <strong>and</strong> phrasal stress in Dutch]. TABU, 26(2): 53-<br />
101.<br />
Gilbers, Dicky, <strong>and</strong> Maartje Schreu<strong>der</strong> (to appear). Language <strong>and</strong> Music in<br />
Optimality Theory. Proceedings <strong>of</strong> <strong>the</strong> 7th International<br />
Congress on Musical Signification 2001, Imatra, Finl<strong>and</strong>.<br />
Extended manuscript available as ROA-571.<br />
Grabe, Es<strong>the</strong>r, <strong>and</strong> Paul Warren (1995). Stress shift: do speakers do it or do<br />
listeners hear it? In: Connell, Bruce <strong>and</strong> Amalia Arvaniti (eds.).<br />
<strong>Phonology</strong> <strong>and</strong> phonetic evidence. Papers in Laboratory<br />
<strong>Phonology</strong> IV.<br />
Hart, Johan, René Collier, <strong>and</strong> Antonie Cohen (1990). A perceptual study <strong>of</strong><br />
intonation. An experimental-phonetic approach to speech<br />
melody. Cambridge University Press, Cambridge.
The Influence <strong>of</strong> Speech Rate on Rhythm Patterns 201<br />
Huss, V. (1978). English word stress in <strong>the</strong> postnuclear position. Phonetica, 35:<br />
86-105.<br />
Kager, René (1994). Ternary rhythm in alignment <strong>the</strong>ory. ROA-35.<br />
Legendre, Geraldine, Yoshiro Miyata, <strong>and</strong> Paul Smolensky (1990). Harmonic<br />
Grammar - A formal multi-level connectionist <strong>the</strong>ory <strong>of</strong><br />
linguistic well- formedness: An application. In: Proceedings <strong>of</strong><br />
<strong>the</strong> Twelfth Annual Meeting <strong>of</strong> <strong>the</strong> Cognitive Science Society,<br />
884-891.<br />
Lerdahl, Fred, <strong>and</strong> Ray Jackend<strong>of</strong>f (1983). A Generative Theory <strong>of</strong> Tonal<br />
Music. The MIT Press, Cambridge, Massachusetts, London,<br />
Engl<strong>and</strong>.<br />
Liberman, Mark (1975). The Intonational System <strong>of</strong> English. Garl<strong>and</strong>, New<br />
York <strong>and</strong> London.<br />
McCarthy, John J. (1986). OCP Effects: Gemination <strong>and</strong> antigemination.<br />
Linguistic Inquiry, 17: 207-263.<br />
Neijt, Anneke, <strong>and</strong> Wim Zonneveld (1982). Metrische fonologie - De<br />
representatie van klemtoon in Ne<strong>der</strong>l<strong>and</strong>se monomorfematische<br />
woorden. [Metrical phonology – The representation <strong>of</strong> stress in<br />
Dutch monomorphemic words] De nieuwe Taalgids, 75: 527-<br />
547.<br />
Prince, Alan, <strong>and</strong> Paul Smolensky (1993). Optimality Theory: constraint<br />
interaction in generative grammar. Ms., ROA-537.<br />
Quené, Hugo, <strong>and</strong> Robert F. Port (2002). Rhythmical factors in stress shift.<br />
Paper presented at <strong>the</strong> 38th Meeting <strong>of</strong> <strong>the</strong> Chicago Linguistic<br />
Society, Chicago.<br />
Rietveld, Toni, <strong>and</strong> Vincent van Heuven (1997). Algemene Fonetiek. [General<br />
<strong>Phonetics</strong>]. Dick Coutinho, Bussum.<br />
Schreu<strong>der</strong>, Maartje, <strong>and</strong> Dicky Gilbers (submitted). Restructuring <strong>the</strong> melodic<br />
content <strong>of</strong> feet. In: Proceedings <strong>of</strong> <strong>the</strong> 9th International<br />
<strong>Phonology</strong> Meeting 2002, Vienna, Austria.<br />
Shattuck Hufnagel, Stephanie, Mari Ostendorf, <strong>and</strong> Ken Ross (1994). Stress<br />
shift <strong>and</strong> early pitch accent placement in lexical items in<br />
American English. Journal <strong>of</strong> <strong>Phonetics</strong>, 22: 357-388.<br />
Sluijter, Agaath (1995). Phonetic Correlates <strong>of</strong> Stress <strong>and</strong> Accent. HIL<br />
dissertations 15, Leiden University.<br />
Sluijter, Agaath, <strong>and</strong> Vincent van Heuven (1996). Spectral balance as an<br />
acoustic correlate <strong>of</strong> linguistic stress. Journal <strong>of</strong> <strong>the</strong> Acoustical<br />
Society <strong>of</strong> America, 100(4): 2471-2485.
List <strong>of</strong> Addresses<br />
Drs. Markus Bergmann<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635982, M.Bergmann@let.rug.nl<br />
Drs. Tamás Bíró<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Computational<br />
Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3636852, Birot@let.rug.nl<br />
Dr. Dicky Gilbers<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635983, D.G.Gilbers@let.rug.nl<br />
Dr. Charlotte Gooskens<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Sc<strong>and</strong>inavian<br />
Languages <strong>and</strong> Cultures<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635827, C.S.Gooskens@let.rug.nl<br />
Dr. Dr. Tjeerd de Graaf <strong>and</strong> Drs. Nynke de Graaf<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635982, T.de.Graaf@let.rug.nl<br />
Drs. Angela Grimm<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635920, A.Grimm@let.rug.nl<br />
Dr. Ing. Wilbert Heeringa<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Computational<br />
Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635970, W.J.Heeringa@let.rug.nl
204<br />
Pr<strong>of</strong>. Dr. Vincent J. van Heuven<br />
University <strong>of</strong> Leiden, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Linguistics<br />
Van Wijkplaats 4, 2311 BX Leiden, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 71 5272105, V.J.J.P.van.Heuven@let.leidenuniv.nl<br />
Nienke Knevel<br />
p/a University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635983, N.B.Knevel@student.rug.nl<br />
Dr. Jurjen van <strong>der</strong> Kooi<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Frisian<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635966, J.van.<strong>der</strong>.Kooi@let.rug.nl<br />
Pr<strong>of</strong>. Dr. Ir. John Nerbonne<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Computational<br />
Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635815, J.Nerbonne@let.rug.nl<br />
Drs. Maartje Schreu<strong>der</strong><br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635920, M.J.Schreu<strong>der</strong>@let.rug.nl<br />
Drs. Hidetoshi Shiraishi<br />
University <strong>of</strong> Groningen, Faculty <strong>of</strong> Arts, Department <strong>of</strong> Linguistics<br />
Oude Kijk in 't Jatstraat 26, 9712 EK Groningen, The Ne<strong>the</strong>rl<strong>and</strong>s<br />
+31 50 3635982, H.Shiraishi@let.rug.nl<br />
Dr. Ivilin Stoianov<br />
University <strong>of</strong> Padova, Department <strong>of</strong> General Psychology<br />
Via Venezia 8, 35100 AS Padova, Italy<br />
+39 049 8276676, Ivilin.Stoianov@unipd.it