CBSU bibliography search
To request a reprint of a CBSU publication, please click here to send us an email (reprints may not be available for all publications)
Spoken-Word Recognition As Optimal Bayesian Decision-Making
McQueen, J.M. & NORRIS, D.
In proceedings of 13th Annual Conference on Architectures and Mechanisms for Language Processing. August 24-27, 2007, Turku, Finland
Year of publication:
We recently developed Shortlist B, a Bayesian model of spoken-word recognition . It shares most theoretical assumptions with its predecessor, Shortlist A , but differs in one critical respect: Shortlist A is based on the notion of Activation, but Shortlist B is based on the assumption that listeners are optimal Bayesian decision-makers. Here, we test that assumption, and the model. In Experiment 1, Dutch listeners made lexical decisions to spoken lists of Dutch CVC words and nonwords. As in an English study , words varied in lexical frequency, and words and nonwords varied in measures of neighborhood density and frequency, as defi ned over the monosyllabic lexicon. Effects of all three variables, and interactions among them, were found. For example, decisions were faster to high- than to low-frequency words and to words in sparse than in dense neighborhoods (but only for low-frequency words), and decisions were more accurate for words with low- than with high-frequency neighbors (but only in sparse neighborhoods). Such effects follow from Bayesian principles. The probability of recognizing a word should vary as a function of its own prior probability, and the prior probabilities of its neighbors. The Bayesian approach makes another prediction, however: Neighborhood effects should depend on list composition. We tested this in Experiment 2, which was identical to Experiment 1 except that the CVCs were mixed with polysyllabic words and nonwords. Word frequency effects were the same (as predicted, prior probabilities of the targets themselves should have similar effects no matter what occurs on other trials), but the effects of neighborhood on word decisions found in Experiment 1 diminished (the infl uence of monosyllabic neighbors should become weaker when the probability increases that a trial might contain a polysyllabic word). Simulations of these data with Shortlist B will also be presented. We will argue that word recognition is a Bayesian process. Effects of word frequency, lexical neighborhood, and list composition all follow from the simple assumption that listeners attempt to make optimal decisions about the words they are hearing.  Norris, D. & McQueen, J.M. (submitted). Shortlist B: A Bayesian model of continuous speech recognition. http://www.mrc-cbu.cam.ac.uk/~dennis/NorrisMcQueenSLB.pdf  Norris, D. (1994). Shortlist: A connectionist model of continuous speech recognition. Cognition 52, 189-234.  Luce, P.A. & Pisoni, D.B. (1998). Recognizing spoken words: The neighborhood activation model. Ear & Hearing 19, 1-36.