Corley, M. 1 , MacGregor, L. 2 & Donaldson, D. 3
1 University of Edinburgh
2 MRC Cognition and Brain Sciences Unit
3 University of Stirling
The human language comprehension system must be robust to the fact that spontaneous speech is riddled with false starts, repetitions, and disfluent hesitations, at an estimated rate of 6 per 100 words (Fox Tree, 1995). Where predictions are made, they must be updated or abandoned as necessary in the face of these disfluencies. Here, we contrast two ways in which prediction could be affected by disfluency. The first is algorithmic", in the sense that the listener makes use of full information about shared knowledge and the speaker's state (e.g., Arnold et al., 2004, 2007) in order to predict what is most likely to be mentioned following a disfluency. However, this view is largely based on evidence from Visual World experiments in which one of a small set of candidate images will ultimately be named. An alternate, heuristic", view derives from a series of ERP experiments (Corley et al., 2007; MacGregor et al., 2009, 2010) in which the magnitude of an N400 response to unpredictable vs. predictable words is attenuated when the target words follow a disfluency, and targets which occur post-disfluency are more likely to be accurately identifed in a later recognition test. Based on this evidence, we argue that listeners do not make specific predictions where the evidence is not good (e.g., where the speaker is disfluent). As a consequence, more attention is allocated to the speech signal, rendering it more memorable, and preventing subsequent acoustic anomalies from further engaging attention (Collard et al., 2008). Once fluent speech resumes, listeners must get `back on track', and we offer a tentative memory-control account of a distinct post-disfluency late positivity (MacGregor et al., 2009, 2010). By focusing on spoken language, we emphasize the interplay of predictive, attentional, and memory processes in robustly comprehending the speaker's intended message.