Settling dynamics in distributed networks explain task differences in semantic ambiguity effects: Computational and behavioral evidence

Settling Dynamics in Distributed Networks Explain Task Differences in Semantic Ambiguity Effects: Computational and Behavioral Evidence Blair C. Armstrong (blairarm@andrew.cmu.edu) Department of Psychology and the Center for the Neural Basis of Cognition, Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh, PA 15213 USA David C. Plaut (plaut@cmu.edu) Department of Psychology and the Center for the Neural Basis of Cognition, Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh, PA 15213 USA the patterns of performance observed for homonymous, polysemous, and unambiguous words in different tasks. For example, lexical decision studies typically report faster responses to polysemous words and either no or minimal differences between unambiguous and homonymous words (Azuma & Van Orden, 1997; Rodd et al., 2002; Hino et al., 2006). In contrast, semantic categorization studies show roughly the opposite pattern of results: words with less relation among their meanings (i.e., tending towards homonymy) are typically responded to more slowly than words with highly related meanings or unambiguous words (Hino et al., 2006). Hino et al. (2006) argued against a semantic coding based explanation of the task differences given that all tasks share the same semantic coding process. Consequently, they suggest that the observed task differences “are likely not due to the semantic-coding process as that process is conceptualized within parallel distributed processing (PDP) models” (p. 266); rather, they must be the result of how the decision-making component of different tasks taps into the semantic code. Without denying that decision processes may differ across tasks, we propose that apparent contradictions in the results from different types of behavioral experiments can in fact be explained primarily by how the semantic-coding process unfolds over time, as conceptualized in a PDP network. Specifically, the nonlinear dynamics of parallel distributed processing systems are such that different trends can manifest themselves at different time points during processing (Kawamoto, 1993). Thus, the apparent task differences may result from the different degrees of semantic precision required to complete each task. In particular, very coarse semantic information may be sufficient to decide that a letter string is a word, whereas semantic categorization requires deriving a sufficiently precise semantic representation to verify category membership. To assess the validity of our proposed account, we implemented a connectionist model aimed at predicting the degree of semantic precision realized for unambiguous, polysemous, and homonymous words as a function of the time-course of processing. We also carried out a lexical decision experiment in which we varied the difficulty of the task to show that when the configuration of the decision system is constant, increasing the degree of required semantic Abstract Developing a theory of semantic ambiguity resolution (i.e., selecting a contextually appropriate interpretation of a word with multiple meanings such as BANK) has proven difficult because of discrepancies in the effects of relatedness of meaning observed across tasks. Hino, Pexman, and Lupker (2006) suggested that these task differences could not be attributed to a general semantic coding process as this process is shared across the tasks, but instead must be due to differences in the configuration of a decision making system. We argue that these task differences can be explained in terms of the settling dynamics of semantic coding within a distributed network. We support our account with a connectionist model of the semantic coding process and a lexical decision experiment in which we vary the difficulty of the task. The results show that increasing the degree of semantic coding alone produces results similar to those observed in different tasks. Keywords: semantic ambiguity; word comprehension; processing dynamics; computational/connectionist modeling; decision making; lexical decision. Deriving the meaning of a word presents a challenge in part because many words do not convey the same meaning in all of the contexts in which they are encountered. A classic, oft- cited example of this phenomenon is the word BANK, which refers to the border of a river in some contexts, and to a financial institution in others. Words such as BANK whose meanings are substantially modulated by context are referred to as being semantically ambiguous (alternatively, lexically ambiguous), and by some accounts represent the majority of words in English and other languages (Klein & Murphy, Central to developing a theory of semantic ambiguity resolution is understanding the impact of the relatedness among the meanings of an ambiguous word – a question which has been studied in substantial detail recently (Azuma & Van Orden, 1997; Rodd, Gaskell, & Marslen-Wilson, 2002; Hino, Pexman, & Lupker, 2006). These studies typically show different performance for polysemous words with related meanings (e.g., / PAPER) relative to unambiguous words with only a single meaning (e.g., CHALK) and homonymous words, with unrelated meanings (e.g., BANK). However, arriving at a comprehensive account of semantic ambiguity resolution has nevertheless proven difficult because of the discrepancies in