Armstrong, B. C. (2012).  The Temporal Dynamics of Word comprehension and Response Selection: Computational and Behavioral Studies.  Doctoral Dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy, Psychology Department, Carnegie Mellon University.  

Download:

Author's self-archived version (zipped .pdf)  (411 pages)

Abstract

Most words are semantically ambiguous in that their meaning depends on the context in which they occur (e.g., <river> vs. <money> BANK). Developing a theory of how the meanings of semantically ambiguous words are comprehended has proven difficult because of discrepancies in the effects of relatedness of meaning observed across tasks. Further, existing accounts are underspecified, narrow in the scope of issues they address, and mutually inconsistent.

The current work proposes a theory of semantic ambiguity resolution in which the discrepant effects are explained by the temporal settling dynamics in semantics and how these dynamics interact with the semantic representations of ambiguous words. This account was instantiated using a distributed connectionist network that incorporated biologically processing constraints. The network shows that the semantic activity evoked at different points in time is consistent with the effects observed in different tasks. The account is further supported by behavioral studies of lexical decision, to evaluate whether differences in processing time, as opposed to qualitative task differences, are responsible for the different ambiguity effects observed across tasks. In these experiments, task difficulty—and the presumed amount of semantic processing—was manipulated both by altering the wordlikeness of the nonword foils and by altering the visual contrast of the stimuli. The selection of optimized word stimuli was enhanced by the development of an automatic stimulus selection algorithm which allowed for a large number of confounding variables to be controlled for, including a measure of the relative frequency of an ambiguous word’s meanings that was collected using a new norming method. The results of the lexical decision experiment show that the contrast manipulation caused large increases in overall latencies and produced semantic ambiguity effects consistent with later semantic processing. This coordinated computational and behavioral work suggests that properties of settling dynamics within a distributed network explain the discrepancies observed across tasks, and generate predictions that can guide future research.

Furthermore, this work points to the importance of understanding how the semantic, orthographic, and phonological representations interact with the response selection system to generate the patterns of effects observed in different tasks. The second portion of the dissertation develops a model of response selection that employs a similar set of domain-general learning, processing and representation principles to those that were used to model semantic ambiguity effects. This work was challenged by previous computational and behavioral investigations using a numerosity judgement task that revealed numerous disagreements between the connectionist models and the behavioral data. New behavioral data collected in an extension of the original numerosity judgement paradigm show that some of these findings do not replicate and were likely due to several idiosyncratic aspects of the original experiment. Connectionist simulations of this extension of the original task succeed in capturing key elements of these new data, including some that are not captured by other models. This work provides the foundation for developing models that integrate the word comprehension system and the response selection system to understand and predict the effects of ambiguity in different tasks, and beyond.

keywords: word comprehension; semantic ambiguity; temporal processing dynamics; models of response selection; computational/connectionist modeling; lexical decision; stimulus degradation; numerosity judgment; selecting optimized stimuli; norming meaning dominance