effet du nombre des graphèmes en Anglais - Aix Marseille Université

Appendice I 229To our knowledge, neither the pseudohomophone effect, nor the feedforward and feedback consistencyeffects in the LDT have been given any formal account by a computational model so far. Preformal (verbalor boxological) accounts of both effects exist in the literature, but we will not consider them here (for argumentsconcerning the strengths and weaknesses of, and the complementarity between, different model formats,see Jacobs & Grainger, 1994). Note however that any model that does not assume automatic activationof -and feedback from- phonological representations in the visual LDT does not lead us to expect anyeffect of feedforward or feedback consistency, while models of the resonance/interactive activation family (fora classification of models, see Jacobs & Grainger, 1994) suggest such an effect rather naturally, that is,without going through pains of adding auxiliary assumptions (Stone et al., 1997 ; Ziegler et al., in press b,c). The present simulation studies will tell us whether this intuition matches the computational evidence.MODEL PRESENTATIONMODEL HISTORYThe scientific adventure of interactive activation models of cognitive processing has a rich history. Theconceptual ingredients that characterize this family of models can be traced back to many authors in differentfields such as biological cybernetics, artificial intelligence, and psychology (e.g. Arbib & Caplan, 1979 ;Erman & Lesser, 1975 ; Grossberg, 1976, 1980 ; Levin, 1976 ; Marslen-Wilson & Welsh, 1978 ; Morton,1969 ; Rumelhart, 1977 ; for a detailed history see Rumelhart & McClelland, 1986). As far as word recognitionis concerned, the adventure started for us with the publication of the two papers by McClelland andRumelhart (1981) and Rumelhart and McClelland (1982), in which all the different conceptual ingredientswere synthesized in an original and formal way, allowing direct applications to psycholinguistic studies.Why did we choose this model format and type? Before the interactive activation model (IAM), basicallytwo model formats were used in the word recognition literature : verbal or V-type models (any model that isexpressed verbally or graphically without making use of closed-form or algorithmic formulations) andmathematical, or M-type models (models that use closed-form expressions to represent the modeled sectionof reality). The IAM introduced algorithmic, or A-type models (models that are implemented in form of asimulation program, including production systems and neural nets of the localist or distributed families) tothe field.Apart from well-known innovative aspects that distinguished the IAM from its precursors (Jacobs &Grainger, 1992), it offered three possibilities that neither V-type, nor M-type models could provide as awhole. First, it possessed dynamics , and thus offered two important possibilities : i) time-dependent predictions,and ii) interval-scaled modeling of RT as a dependent variable 5 . McClelland and Rumelhart (1981) andRumelhart and McClelland (1982) had only exploited the first of these possibilities. We were also muchinterested in the second, given that RT is the major dependent variable in psycholinguistic research (mainlydue to the popularity of the LDT and naming task). Second, the IAM possessed a (toy) lexicon which madeitem-specific (fine-grained) predictions possible. To us, this seemed a logical necessity in a field that has toevaluate its empirical effects with respect to subject- and item-specific data . In addition, if one believes inthe virtues of strong scientific inference (see Grainger & Jacobs, present volume), fine-grained analyses are anecessity : Falsificationism and strong inference are not the only research strategies, but they seem to be thebest ones whenever theorists are concerned with specific assumptions that can be tested at the level of finegrainedanalyses (Jacobs & Grainger, 1994 ; Massaro & Cowan, 1993). Third, contrary to other models, theIAM rather than being definitive, possessed rich structural potential and appeared to include the promise ofinteresting further developments. For a reasonable application of the stratagems of nested and canonicalmodeling (Grainger & Jacobs, 1996), structural potential is a necessary (but not a sufficient) condition. TheIAM, as the prototype of a canonical resonance model (Stone & Van Orden, 1994), allows the testing ofsystem and design principles, to which one can attribute explanatory credit and blame independently of otheraspects of the model (Grainger & Jacobs, present volume).In sum, by its original combination of formal preciseness, structural-computational richness, and computationaltransparency (a feature that distinguishes it from most PDP models), the IAM intuitively seemedthe right model at the right time to allow falsifiable quantitative predictions, and the discovery of new phenomenavia simulations, the so-called neighborhood frequency effect, discussed below.PREDICTING A NEW PHENOMENON : THE DISCOVERY OF THE NEIGHBORHOODFREQUENCY EFFECT"Only theories tell us what can be observed" (variation on a theme by Einstein)5 For a critique of interval-scaled RT models, see Van Orden and Goldinger (1994), or Uttal (1990). For a replique,see Jacobs and Grainger (1994).