effet du nombre des graphèmes en Anglais - Aix Marseille Université

230Appendice IClearly, the ability to predict a new phenomenon (and the conditions under which it must appear or mustnot appear) is one of the higher criteria for model evaluation (Gigerenzer, Hoffrage, & Kleinbölting, 1991 ;Jacobs & Grainger, 1994). One of the features of connectionist models in general, and the IAM in particular,is that they are rich enough to allow emergence of effects that have not yet been observed.As an example for this, consider the neighborhood frequency effect in visual word recognition. Playingwith a variant of the original IAM, one of the present authors discovered that such an effect is possiblewithin the model system. Looking at the activation function for the word "blur", he observed an attenuationof the rise of this function during the early phase (a cross-over between the functions for "blur" and "blue").This is because "blur", a low frequency word, shares all but one letter with "blue", a high frequency word.The activation functions for "blur" type words reach a criterion level of activation (arbitrarily defined forresponse generation in the model) more slowly than low frequency words that have no high frequency orthographicneighbors (e.g., "idle" ; see Figure 2). What is more important, is that further simulations with theIAM showed that a selection of low frequency words with many high frequency neighbors (e.g., heal) did notdiffer from low frequency words with a single high frequency neighbor in terms of the number of cycles requiredto reach criterion activation levels. This particular simulation result was important with respect to ourapplication of strong scientific inference in model testing. In contrast to the IAM, serial search / verificationmodels of visual word recognition (Forster, 1976 ; Paap, Newsome, McDonald, & Schvaneveldt, 1982) predicta further decrement in performance to such stimuli. The pattern predicted by the IAM was observed byGrainger et al. (1989) and Grainger (1990). Although more recent research has complicated the neighborhoodfrequency story (e.g., Sears, Hino, & Lupker, 1995), the important point is that the IAM simulations, usingthe same stimuli as in the human experiments, accurately predicted the observed pattern for that particularstimulus set.Thus the IAM, which, as all connectionist models, has structural and processing features that were builtin specifically to create known empirical phenomena (e.g. the resting level parameter that creates the frequencyeffect ; cf. Dell, 1988), predicted an unknown effect that has now been observed under a variety ofconditions (Grainger & Jacobs, 1996). Let us note as an aside that this provides an encouraging example forsolving the recurring epistemological issue of a theory-centered approach as seemingly opposed to a resultcenteredapproach (Greenberg, Solomon, Pyszcinski, & Steinberg, 1988 ; Greenwald & Pratkanis, 1988 ;Moser, Gadenne, & Schröder, 1988), by showing how a model can specify the conditions under which previouslyunobtainable results occur. This clearly is a theory-centered demonstration for one of the complementarytwo "result-centered" approaches (i.e., the design approach), advocated by Greenwald, Pratkanis,Leippe, and Baumgardner (1986) in their attack on theory-centered approaches to psychology.ACTIVATION0.80.60.40.20-0.22015isleidleidly2aResponse cycle : 17ACTIVATION0.80.60.40.20blurResponse cycle : 19blue20152bCYCLES1050505490495500WORD NUMBER-0.2959085WORD NUMBER800510 CYCLESFigure 2 : Activation functions in reduced parts of the English four letter lexical space. Panel 2a and 2b showtwo simulations obtained with the target words "IDLE" and "BLUR", respectively. Both these words have orthographicneighbors ("IDLY", "ISLE" and "BLUE", respectively) and have the same frequency (15 per million), butonly "BLUR" has a higher frequency neighbor. Consequently, "IDLE" reaches the response threshold two cyclesearlier than "BLUR" (i.e., 17 vs. 19 cycles).MODEL STRUCTURETHE MROM. It is useful to give a short description of the MROM here (see Grainger & Jacobs, 1996,for more details). The MROM is an extension of the IAM incorporating the design principle of multipleread-out. The principle of multiple read-out states that a response in a given experimental task is generated(read-out) when at least one of the codes that is appropriate for responding in that task reaches a critical activationlevel. This principle is particularly relevant to our explanation of performance in the LDT. With respectto this particular task, we hypothesize that unique word identification is not the only process that canlead to a correct "yes" decision in the lexical decision task, and that an extra-lexical process controls the productionof "no" responses. In the functional context of the lexical decision task, word-nonword discriminationrequires that participants use a reliable source of information that allows them to make rapid and accuratejudgments concerning the "word-likeness" of stimuli (e.g., their familiarity / meaningfulness, Balota &