Testing Assumptions in Computational Theories of Aphasia Wheeler Ruml, Alfonso Caramazza, Jennifer R. Shelton, and Doriana Chialant Harvard University We present the performances of 13 aphasic patients on a picture-naming task and attempt to model these data using computer simulations. We systematically manipulate the assumptions underlying several interactive, two-step, spreading-activation models, including the proposals of Dell et al. (1997), Foygel and Dell (2000), and Rapp and Goldrick (in press). Using a numerical regres- sion procedure and multiple views of each model’s possible output, we find that peripheral pragmatic assumptions play a role equal to that of theoretically more central model components. None of the models we consider can account for all of the patients, leading us to conclude that one or more of the assumptions underlying each model is flawed. We argue that there are strong limitations on the conclusions that can legitimately be drawn from such simulation studies but that close analysis of individual patients can allow sound testing of potentially more accurate models. © 2000 Academic Press Key Words: computational modeling; aphasia; lexical access; computational neuropsychology. The promise of computational models of hu- man language processing is widely recognized. Not only does the act of constructing a simula- tion force one to specify one’s theory precisely, but the resulting model can be quantitatively tested against empirical data. Furthermore, the ease of simulation allows one to experiment with models that deviate from normal behavior and thereby to form theories about the interac- tions between brain damage and language pro- cessing. Data from aphasic patients can then be used to test the adequacy of the combined model of normal processing and damage in aphasia. One could even imagine using simula- tion results to provide insight into the break- down occurring in specific patients. Examples of recent computational investigations of low- level language processing include the word- reading models of Plaut et al. (1996) and Shal- lice et al. (1995) and the word production models of Levelt et al. (1999) and Dell et al. (1997). In practice, a computational model is often constructed with the aim of testing claims about one or two specific theoretical issues, such as the role of interaction between levels of repre- sentation during word production. But the col- lection of theoretical principles that one wishes to put to empirical test does not usually describe a complete mechanism suitable for simulation. Details beyond the scope of the theory at stake must be filled in, such as the exact semantic relations between words in the model. And de- tails supposedly within the purview of the the- ory must often be left out for the sake of reduc- ing computation time, such as the full inventory of a typical human lexicon. In this paper, we systematically examine the role played by these seemingly minor assumptions by evaluating three closely related models of word produc- tion. We present data from thirteen aphasic pa- tients on a picture-naming task and attempt to account for their performance using each of the We thank Gary Dell, Randi Martin, and two anonymous reviewers for their many helpful comments, Nadine Martin for help in scoring some patient responses, Brenda Rapp and Matthew Goldrick for providing detailed information re- garding their model’s lexicon, Angelos Kottas for running some preliminary experiments, and Michele Miozzo and the Harvard Cognitive Neuropsychology Laboratory for many stimulating discussions regarding this research. Support was provided in part by the National Science Foundation under grants CDA-94-01024 and IRI-9618848, and by the Na- tional Institutes of Health under grant NS-22201. Please address correspondence concerning this article to Wheeler Ruml, Maxwell Dworkin Laboratory, Harvard University, 33 Oxford Street, Cambridge, MA 02138. E-mail: ruml@eecs.harvard.edu. 217 0749-596X/00 $35.00 Copyright © 2000 by Academic Press All rights of reproduction in any form reserved. Journal of Memory and Language 43, 217–248 (2000) doi:10.1006/jmla.2000.2730, available online at http://www.idealibrary.com on