Do Computational Models of Reading Need a Bit of Semantics?

Summary. Coltheart, Rastle, Perry, Langdon, and Ziegler [1] claim that “the psychology of reading has been revolutionized by the development of computational models of visual word recognition and reading aloud”. They attribute this to the fact that a computational model is a computer program – an algorithm – “that is capable of performing the cognitive task of interest and does so by using exactly the same information-processing procedures as are specified in a theory of how people carry out this cognitive activity” [1, p. 204]. According to this view, the computational model is the theory, not a simple instantiation of a theory. In this paper we argue that computational models of reading have indeed helped in dealing with such a complex system, in interpreting the phenomena underlying it, and in making sense of the experimental data. However, we also argue that it is crucial for a model of reading to implement a computational semantic system that is as yet a missing component of all computational models. We provide two reasons for such a move. First, this would allow explaining some phenomena arising from the interaction of semantics and lexical variables. We will review the following empirical findings: faster response times to polysemic words [2] and slower response times to synonyms [3]; the leotard [4] and turple effects [5]; and the asymmetry of the neighbourhood density effect in free and conditional reading [6]. Second, such an “enriched” model would be able to account for a richer set of tasks than current computational models do. Specifically, it would simulate tasks that require access to semantic representation to be performed, such as semantic categorization and semantically-based conditional naming. We will present a computational instantiation of a semantic module that accounts for all the described phenomena, and that has helped in generating predictions that guides on-going experimental activity.