The role of syntactic prediction in auditory word recognition

Abstract

Context is widely understood to have some influence on how words are recognized from speech. This dissertation works toward a mechanistic account of how contextual influence occurs, looking deeply at what would seem to be a very simple instance of the problem: what happens when lexical candidates match with auditory input but do not fit with the syntactic context. There is, however, considerable conflict in the existing literature on this question. Using a combination of modelling and experimental work, I investigate both the generation of abstract syntactic predictions from sentence context and the mechanism by which those predictions impact auditory word recognition. In the first part of this dissertation, simulations in jTRACE show that the speed with which changes in lexical activation can be observed in dependent measures should depend on the size and composition of the set of response candidates allowed by the task. These insights inform a new design for the visual world paradigm that ensures that activation can be detected from words that are bad contextual fits, and that facilitatory and inhibitory mechanisms for the syntactic category constraint can be distinguished. This study finds that wrong-category words are activated, a result that is incompatible with an inhibitory syntactic category constraint. I then turn to a different approach to studying lexical activation, using information-theoretic properties of the set of words consistent with the auditory input while neural activity is recorded in MEG. Phoneme surprisal and cohort entropy are evaluated as predictors of the neural response to hearing single words when that response is modeled with temporal response functions. This lays the groundwork for a design that can test different versions of surprisal and entropy, incorporating facilitatory or inhibitory syntactic constraints on lexical activation when the stimuli are short sentences. Finally, I investigate a neural effect in MEG previously thought to reflect syntactic prediction during reading. When lexical predictability is minimized in a new study, there is no longer evidence for structural prediction occurring at the beginning of sentences. This supports the possibility of a tighter link between syntactic and lexical processing.