Expectation-based language comprehension and production

Roger Levy

Roger Levy

(UC San Diego)

Play Video

Date: January 28, 2015


Using language to communicate is central to what makes us human. Elucidating the knowledge, expectations, and cognitive resources that allow us to communicate so effectively is one of the most fundamental problems in the study of mind. For much of the contemporary history of psychology and linguistics, motivated by the ideas of figures including Chomsky, Miller, and Fodor, work on this problem has conceptualized language processing as centrally about modular structure-building operations and the memory resources required to carry them out. Here I describe an alternative approach that conceptualizes language processing as centrally about rational, goal-driven inference and action. First, I outline a state-of-the-art theory of expectation-based incremental language understanding, in which comprehenders integrate diverse information sources from preceding context to guide interpretation of current input. This theory unifies three key, seemingly disparate topics in the domain of language understanding — ambiguity resolution, prediction, and syntactic complexity effects — and finds broad empirical support in data from both controlled experiments and naturalistic language understanding. Second, I describe several apparent empirical puzzles for this theory that ultimately lead us to revisit one of the implicit foundational assumptions in all theories of language understanding: that of modularity between the processes of word recognition and of inter-word, utterance-level comprehension. I generalize the expectation-based theory to a fully bidirectional, interactive theory of word recognition and utterance comprehension, and show how this generalized theory solves the apparent puzzles and leads to a range of new, empirically verified predictions. Finally, I touch briefly on the consequences of this view for language production: why do speakers choose to structure their utterances the way they do? The expectation-based theory novelly predicts that speakers use the options afforded them by their native language to come as close as possible to a uniform distribution of information content throughout their utterance. We confirm this prediction through statistical analysis of speaker choices regarding optional word omission in naturalistic speech.

Created: Thursday, January 29th, 2015