effet du nombre des graphèmes en Anglais - Aix Marseille Université

234Appendice IFigure 6 gives an illustration of what one might call the "phonemic" space in the MROM-P. It showsactivation functions of the phonological units at the onset, nucleus and codas positions obtained for the targetword "BLUE".PARAMETER TUNING. In our parameter tuning approach, we also followed the stratagem of nested modeling.Since this demands that MROM is an integral part of MROM-P, we faced the problem of having tofind new parameters for the phonological substructures of the model, while keeping the original parametersof MROM as constant as possible. However, because adding new interactive substructures to the originalIAM leads to a new global dynamical system, the original parameter set had to be adjusted (the old one leadto catastrophic model behavior). Table 3 gives the parameter set fixed for the present simulation studies onthe basis of the parameter tuning studies discussed below. We acknowledge that more work has to be done inorder to precisely determine the role of each parameter for the model dynamics. Furthermore, althoughconnectionist modelers in general seem to ignore the issue, we acknowledge that a central aim for futurework in the field must be to solve the nontrivial problem of the identifiability of complex A-type models ingeneral, and models of the IA family, in particular.SIMULATION METHOD. In the following studies, we simulate effects obtained in the LDT. Because thetask-specific mechanisms of the LDT have well been specified in the MROM (Grainger & Jacobs, 1996),the same design principles (i.e., principles that determine the behavior of a class of models and the observeddependent variables) were adopted here. However, given that the present version of MROM-P is still prototypical-implying that the present study provides no criteria-oriented falsification test of the MROM-P, butrather a test of its appropriateness as a "working model" of phonological coding- (see discussion section), wedid not carry out full-blown simulations. This would have necessitated, for example, use of stochastic responsethresholds and of a large number of simulated subjects equal to the number of subjects in the realexperiments (Grainger & Jacobs, 1996). Consequently, rather than response (time) distributions, we simplyobtained activation functions for each item used in the simulated experiments. For each experimental condition,the mean of these activation functions was calculated across the different items. The resulting activationfunctions (and derived mean RT bar charts) are presented as illustrations of how the model can quantitativelyaccount for the experimental data./u/1/bl/1ACTIVATION0.50/sl//pl//kl//gl/ /fl/2030ACTIVATION0.50/ju//$//5/2030403020ONSET NUMBER100010CYCLES403020NUCLEUS NUMBER100010CYCLES1/*/ACTIVATION0.50/ls//l//g/2030403020CODA NUMBER100010CYCLESFigure 6 : Activation functions of phonological units at the onset, nucleus and coda positions obtained for thetarget word "BLUE".Concerning the task-specific read-out procedure, the MROM-P follows the design principle of theMROM in that "No" responses are produced by monitoring the global orthographic activity generated by thestimuli, and "Yes" responses are generated by looking at the orthographic unit activity (for simplicity, wedid not consider the possible role of global orthographic activity on "Yes" trials here ; cf. Figure 3 above).The rationale for this is given in Ferrand and Grainger (1996) who discuss the qualitative predictions ofMROM-P for a masked priming LDT. In accord with the assumptions of Grainger and Jacobs (1996), theresults of Ferrand and Grainger suggest that in an LDT using pseudohomophones, participants use read-outfrom the orthographic lexicon, because read-out from the phonologic lexicon would lead to too many falsepositive errors.