21.04.2013 Views

Eckhard Bick - VISL

Eckhard Bick - VISL

Eckhard Bick - VISL

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Therefore, there is a case for introducing "real" (i.e. not primarily syntactic) semantic<br />

classes on top of “mappable” (i.e. syntax or morphology based), semantic distinctions.<br />

But which and how many semantic tag categories should be chosen? And how<br />

should they be organised with regard to each other? When answering this question, it is<br />

important to make the teleological distinction between defining meaning and<br />

distinguishing meaning. In a system that does not claim to understand text, but only to<br />

structure or translate it, the final semantics lies (only) in the eye of the beholder.<br />

Therefore, in order to make a sense distinction, it isn't necessary to define a given<br />

sense, or identify its referent, - it is enough to draw a line across the semantic map<br />

which separates two different senses of a word. If a parser can determine - in a given<br />

text - just which side of the line a given usage has to be placed, then a disambiguated<br />

tag will allow the system to pick the exact word sense from the lexicon – the sense need<br />

not be in the tag itself. Constraint Grammar is an ideal tool to make such choices in a<br />

context dependent and robust way. Like on the morphological and syntactic level, CG<br />

rules will not define but disambiguate from a given set of distinctions. Semantic tags,<br />

then, have to reflect the structure of the semantic landscape rather than the meaning of<br />

individual words. Consequently, for a CG to work well, the system’s set of semantic<br />

tags should be inspired by some ordering or classification principle for word senses.<br />

Word meanings – to mention some basic systems - can be ordered in autonomous<br />

groups (word fields, picture dictionaries), semantic hierarchies (monolingual thesauri,<br />

biological classification, Princetown WordNet) or word nets (multilingual relational<br />

systems like EuroWordNet). Finally, word meanings can be ordered by semantic<br />

decomposition:<br />

vivo<br />

-> morto = ‘NOT alive’<br />

viver = ‘BE alive’<br />

-> nascer = BECOME alive’<br />

-> morrer = ‘BECOME NOT alive’<br />

-> matar = ‘MAKE BECOME NOT alive<br />

I believe that the most cost efficient 217 way to draw distinction lines across the semantic<br />

landscape is by prototype similarity. Prototypes can be conceived as both class<br />

hyperonyms (‘animal’: dog, pig, lion, ‘plant’: oak, sunflower) and common well known<br />

set members of classes otherwise too “abstract” (‘knife’ for ‘cutting tools’: knife, sword,<br />

saber, ’book’ for ‘readables’: book, paper, magazine). Whatever the abstraction level,<br />

prototype similarity testing is done by asking ‘is it more like A or more like B?’ rather<br />

217 On the one hand, in terms of CG rule efficiency, on the other, in terms of the work load needed to enter the relevant<br />

information into the lexion.<br />

- 365 -

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!