effet du nombre des graphèmes en Anglais - Aix Marseille Université

Appendice I 2431989), and it can be argued that the fields of memory or reading research are still not fully prepared for suchan enterprise (Humphreys, Wiles, & Dennis, 1994 ; Jacobs, 1994). This is not to say, however, that theyare not set up for some fair amount of theoretical unification, as argued below.Neither the recent special section of the JEP :HPP (1994) on modeling visual word recognition, nor anyother literature we have come to know since provides a computational model that can formally account forthe pseudohomophone and bi-directional consistency effects in the LDT we have simulated here. However,there is, at least, one other model that has the potential to provide such an account, the dual route cascadedor DRC model (Coltheart & Rastle, 1994). Given the fact that both the present model and the DRC belongto the family of localist connectionist models (and therefore are more easily comparable), but differ with respectto one crucial structural feature -the presence of a GPC rule mechanism in the DRC- we may have thepromise of some exciting strong inference studies ahead. However, for such studies to be efficient, certainmethodological issues have to be solved, as argued below.Future work concerning the MROM-P will involve adding phonological representations to multilingualmodels such as the bilingual interactive activation model presented by Dijkstra and Van Heuven in the presentvolume. This raises the interesting question of how multilingual speaker-readers deal with the differentsets of spelling-to-sound correspondences in each language. If one postulates, as in the present work, thatthe phonological coding component of the MROM-P is automatic and strategically non-modifiable, how dobilingual readers deal with the potential interference caused by automatically generating all correspondencesin both languages? The notion of language node introduced by Grainger and Dijkstra (1992, see also Dijkstra& Van Heuven, present volume) provides one solution to this problem. Top-down inhibition from the unattendedlanguage node to the corresponding word units (i.e., all words in the unattended language) wouldblock resonance between irrelevant phoneme units and word units in the unattended language. Clearly muchexciting theoretical and empirical work is yet to be done in the multilingual domain.GOING BEYOND MROM-P : A GENERAL CHALLENGE FOR COGNITIVE MODELERS IN"WORD-NERD" WORLD"Science, even more profoundly than politics, is the art of the possible. It does only what can be donenext" (Newell, 1990, p. 26).The 1994 special section of the JEP :HPP entitled "Modeling visual word recognition" gives an impressionof the empirical and theoretical richness of this classical field of experimental, cognitive psychology.The question we want to raise here is whether it is about time for some serious effort of theoretical unification?Who thinks that this is an idle question is invited to take a look at Table 1 in our editorial of that specialsection (Jacobs & Grainger, 1994), which gives a selective overview of 15 models of visual word recognitionstarting with Morton's (1969) logogen model. Originally, this piece of taxonomic work startedwith the ambition to give a synopsis -as complete as possible- of models of word recognition to be used forthe tasks of theoretical unification and development of standards for model comparison and evaluation. Thatwas clearly too ambitious! Even the published version is far too complex to have a fair chance of being usedin the way we wanted it to (previous versions of the table, including more than 40 different models of thepast 30 years were worse!).Notwithstanding, we continue to think that the answer to the above question is a clear "yes". Perhapspressure for theoretical unification in the word recognition literature is not as high as in the general field ofcognitive science. As Newell (1990, p. 25) states : "In my view, it is time to get going on producing unifiedtheories of cognition-before the data base doubles again and the number of visible clashes (i.e., betweentheory and data) increases by the square or cube". On the other hand, if Newell (1990) can convincingly presentfour harbingers of unified theories of cognition, why should we not be able to agree on a limited numberof harbinger models that have the potential to become unified models of word recognition? After all, weare only dealing with a small part of the cake of cognition (although perhaps with one of the most complexparts). Clearly, we tend to think that models of the steadily growing IA family (see present volume) areamong the harbinger candidates, and since we have given our arguments for this in Jacobs and Grainger(1994), we will not reiterate them here (see Grainger a Jacobs, present volume). Suffice it to say that thepresent results showing the generalizability of the original IAM to conditions involving phonological processingare clearly in favor of our view.However, to facilitate theoretical integration in the "word-nerd" field, the following challenge has to bemet : "Agree on a minimal set of standards for model comparison and evaluation". This challenge has twofacets. First, it involves agreement on a minimal set of standard effects and tasks, any model of word recognitionthat competes for "harbingership" should be able to predict in a way that can enter into strong inferencecompetition. In Jacobs and Grainger (1994 ; Table 1) we made a minimalistic proposal of four sucheffects. Whether they were the right ones, or whether they have to be augmented by other effects is a questionthat can only be solved by ongoing published scientific debate.The second, and more problematic facet of this challenge is to agree on a minimal set of criteria formodel evaluation, and on a standard way of applying them. Among other things, this implies the trickyproblem that complex A-type models have to be made comparable, at least on a number of critical featuresor dimensions (e.g., the "currency" problem discussed by Massaro & Friedman, 1990). This is a nontrivialissue (cf. Estes, 1975), as model builders who have problems of keeping different variants of their ownmodel comparable over the years of development will admit (for some examples, see Jacobs & Grainger,