Inducing homonymy effects via stimulus quality and (not) nonword difficulty: Implications for models of semantic ambiguity and word recognition

Inducing homonymy effects via stimulus quality and (not) nonword difficulty: Implications for models of semantic ambiguity and word recognition 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 Abstract semantic features of homonyms, polysemes, and unambigu- ous words leads to competitive and co-operative settling dy- namics. These dynamics explain the ambiguity effects as a re- sult of sampling from the semantic code at different points in time (Armstrong & Plaut, 2008). Early on, co-operative dy- namics amongst the overlapping features of polysemes give rise to a polysemy advantage, whereas later competitive dy- namics amongst the inconsistent features of homonyms give rise to a homonymy disadvantage. In past connectionist mod- eling work, we (Armstrong & Plaut) have confirmed these predictions and shown that activation in semantics alone is sufficient to account for these two effects, and predicts both effects at some intermediate time-point (see Figure 1). Never- theless, stronger support for this account would involve show- ing that it correctly predicts a result that is not predicted by the decision-making account. Reports of a processing advantage for polysemes with re- lated senses (e.g., / PAPER) in lexical decision and a processing disadvantage for homonyms (e.g., / BANK) in semantic categorization have prompted the development of conflicting accounts of these phenomena. Whereas a decision-making account (Hino, Pex- man, & Lupker, 2006) suggests these effects are due to quali- tative differences between the tasks, accounts based on tempo- ral settling dynamics (Armstrong & Plaut, 2008) suggest that processing time is the critical factor. To compare these ac- counts, we manipulated nonword difficulty and stimulus qual- ity to make lexical decision difficult and attempted to produce the same homonymy disadvantage as in semantic categoriza- tion. We found that stimulus degradation succeeded to this end, and nonword difficulty only consistently slowed nonword responses. This provides evidence both for settling dynamics accounts of semantic ambiguity in particular, and for interac- tive orthographic-to-semantic processing and the construction of more integrated models, in general. Keywords: semantic ambiguity; settling dynamics; decision making; lexical decision; models of word recognition; non- word difficulty; stimulus degradation. Developing a mechanistic account of how words associ- ated with multiple interpretations (e.g., / BANK) are recognized is central to understanding the repre- sentations and processing mechanisms underlying word com- prehension. Recently, there has been a major upheaval in the ambiguity literature, as researchers have discovered that long held ambiguity effects are not associated with all am- biguous words universally. Rather, these effects appear to be critically modulated by the relatedness amongst the in- terpretations of the ambiguous word. Further complicating matters, there have been reports that the effects of related- ness are also not consistent across tasks. For instance, rel- ative to unambiguous controls, polysemes with highly re- lated senses (e.g., / PAPER) show a processing advantage in lexical decision (Rodd, Gaskell, & Marslen-Wilson, 2002), whereas a processing disadvantage has been reported for homonyms (e.g., BANK) in semantic categorization (Hino, Pexman, & Lupker, 2006). Two contrasting accounts have been proposed to explain these disparate results. One suggests that the post-semantic decision-making component of the two tasks is qualitatively different in lexical decision and semantic categorization and causes these different effects (Hino et al., 2006). Another ac- count suggests that varying numbers and overlap amongst the Figure 1: Reproduction of simulation results reported by Armstrong and Plaut (2008). The plot shows the average number of semantic units with activations above 0.7 in a connectionist network for polysemous, unambiguous, and homonymous words as a func- tion of time (in unit updates). Early on, the model shows a polysemy advantage (Slice A), late during processing it shows a homonymy disadvantage (Slice C), and in between it shows both effects (Slice B). These two accounts clearly make very different predic- tions for the patterns of performance that should be observed within and between tasks. The decision-system account im-