effet du nombre des graphèmes en Anglais - Aix Marseille Université

242Appendice INotwithstanding, a nontrivial question regarding falsifiability can be answered here : Is the MROM-Pfalsifiable, at all? Given the repeatedly appearing critique of connectionist models as being too powerful, andtherefore not falsifiable in any easy way, this is not an idle question. Following other theoreticians, wetherefore have proposed that any A-type model should come with a clear answer to the following question :What cannot be or happen, if the model is correct? In other words, which effects or phenomena does themodel exclude? An example is given in Grainger and Jacobs (1996). They show that if the MROM is correct,a facilitatory effect of orthographic neighborhood density (as measured by Coltheart's N) is possible inboth the yes/no and go/no-go variants of the LDT, but not in the perceptual identification task. Thus, replicableexperimental demonstrations of a facilitatory N effect in the perceptual identification tasks would falsifythe MROM.Another example is given above (see Figure 8). According to the MROM, as included in the presentMROM-P, a phonological pseudohomophone effect in the LDT is not possible, because it possesses nophonological processing units whatsoever. Thus, although the MROM was explicitly designed to deal withorthographic processing in the LDT and other reading tasks that do not include pseudohomophonic stimuli,the simulation data in Figure 8, for example, present a falsification of the (non-phonological) MROM. Evidently,simply falsifying a model by using conditions that are outside of its explicitly stated validity space(domain of application) is not necessarily useful. As we have demonstrated in this chapter, using theMROM as a null-model against which to test models of phonological coding is a more useful variant of"falsification studies".At any rate, in a field that lacks universal laws, we cannot expect models to have universal validity (cf.Newell, 1990). On the other hand, we can hardly want to continue with models that can accurately accountonly for a single effect, as measured by a single variable in a single task, but whose validity stops there (cf.Jacobs & Grainger, 1994 ; Newell, 1990 ; Roberts & Sternberg, 1993).The MROM-P is also falsifiable in several nontrivial respects. Like the MROM, it allows to makequalitative predictions that can be tested in a straightforward way. An example is discussed in Ferrand andGrainger (1996). They used a pre-quantitative version of MROM-P -they called it a bimodal extension of theMROM- to make qualitative predictions concerning the existence and direction of priming effects in amasked priming LDT manipulating prime type (homophones, pseudohomophones, or unrelated controls)and list composition (pseudohomophones, legal pseudowords, or illegal nonwords). The strongest qualitativeprediction of MROM-P, i.e., the one most easily falsifiable, was that it predicts a null effect with homophoneprimes in the presence of illegal nonwords 9 . The rationale for this is that i) the presence of illegalnonwords encourages participants to use the ∑ criterion, since such nonwords can easily be discriminatedfrom words on the basis of summed lexical-orthographic activity. ii) homophone primes generate high levelsof orthographic inhibition when read-out is from the orthographic M criterion. The facilitatory effectsdue to increased use of the ∑ criterion (i.e., the fast-guess mechanism producing decreases in RT) will becanceled by the inhibitory effects due to homophone primes. A null-effect is the predicted result. In Ferrandand Grainger's (1996) experiment, this was the case.DISCUSSION AND OUTLOOKTo summarize : While a complete, criteria-oriented evaluation of the MROM-P is not possible at present,the results of our partial evaluation stand the general test criterion that we had fixed as our objective,that is, whether the present MROM-P is an appropriate "prototype" for developing a general model ofphonological coding in visual word recognition. The model presented here is definitely a prototype, not inthe sense of representing an ideal, but in the sense of being a "working model". If one accepts the principlesof model development we adhere to, it has some virtues. Within the constraint of nested modeling, it representswhat we think to be the simplest possible localist connectionist network that allows an account oftwo?? critical empirical effects indicating the influence of phonological processes in what is still the mostwidely used reading task in experimental psychology and psycholinguistics, i.e. the LDT.Moreover, the MROM-P, as our other work involving A-type modeling, is essentially a heuristic devicein the sense discussed in Grainger and Jacobs (present volume) : It provides a heuristic, algorithmic descriptionof phonological coding, but -needless to say- it falls short of presenting a computational theory in thesense of Marr (1982). It is not difficult to admit this : Few theoreticians in the field of cognitive sciencehave achieved (or come close to) a computational theory (Marr, 1982 ; see also Jacobs, 1994 ; Pylyshyn,9 A word on null-effects and their significance for theory building is in order, because many psychologists arefirm believers in the virtues of null-hypothesis testing (but see Gigerenzer & Murray, 1987 ; Rouanet, 1996 ;Van Orden, Aitchison, & Podgornik, 1996). Whenever one possesses models permitting quantitative predictionswith reasonable precision, the prediction of a null-effect is actually a strong prediction to make. Perhapsthe most famous example is the prediction of the null-effect concerning the speed of light in Michelsonand Morley''s experiments by Einstein's special theory of relativity (Spielberg & Anderson, 1985). This isnot to say that any known psychological A-type model can be compared with Einstein's theory. We simplywant to make clear that the existence of theoretical tools allowing quantitative predictions concerning empiricaleffects frees us from the use of null-hypothesis testing as exclusive inferential method. Thus, contraryto standard practice, accepting the null-hypothesis can become a valid inference whenever one has sufficientfaith in the validity and precision of a formal model or theory.