Segmental duration as a cue to sentence structure

Abstract

In order to parse speech in real time, listeners should use any informative cues available. Here, we investigate the role of segmental duration. Previous work has found statistically significant differences in the mean durations of analogous segments across different lexical/syntactic structures. However, a difference in means does not necessarily mean that the distributions of these durations make individual token durations sufficiently informative to be a useful cue. The goal of this work is to use production data to quantify how informative segmental duration is about syntactic/lexical structure. Our model is based on an ideal listener model, where we assume listeners have implicit knowledge of segmental duration distributions for active and passive sentences. Given these distributions, the model can infer the posterior probability that a particular token belongs to one distribution or the other. After implementing our model in a Bayesian classifier, our results indicate there is indeed sufficient information contained in individual token durations so as to be useful in real-time sentence processing. Furthermore, we modeled listener behavior in a gating task with syntactically ambiguous sentences truncated before disambiguating morphosyntax and achieved 74% accuracy in predicting syntactic outcome, similar to accuracy reported in behavioral studies (62%–84%).