Making predictions from speech with repairs: Evidence from eye movements

Abstract

When listeners hear a spoken utterance, they are able to predict upcoming information on the basis of what they have already heard. But what happens when the speaker changes his or her mind mid-utterance? The present paper investigates the immediate effects of repairs on listeners' linguistic predictions. Participants listened to sentences like the boy will eat/move the cake while viewing scenes depicting the agent, the theme, and distractor objects (which were not edible). Over 25% of items included conjoined verbs (eat and move) and 25% included repairs (eat- uh, move). Participants were sensitive to repairs: where eat was overridden by move, fixations on the theme patterned with the plain move condition, but where there was a conjunct, fixations patterned with eat. However, once the theme had been heard, there were more fixations to the cake in all conditions including eat, showing that the first verb maintained an influence on prediction, even following a repair. The results are compatible with the view that prediction during comprehension is updated incrementally, but not completely, as the linguistic input unfolds.

Keywords:

• Language production
• Speech
• Repairs
• Eye-tracking

Acknowledgements

The author is grateful to Andrew Wilson, Gemma Devlin, and Anika Fiebich for help with the creation of materials and collection of data. Yuki Kamide kindly provided the images used in these experiments, and Christoph Scheepers answered a number of queries concerning the manipulation of eye-tracking data. Pia Knoeferle and two anonymous reviewers made a number of useful suggestions concerning the manuscript. Versions of this work were presented at the AMLaP 2006 and CUNY 2007 conferences.

Notes

¹I am grateful to an anonymous reviewer for drawing this to my attention.

²In an attempt to determine whether pitch or duration played a role in repair detection, we compared durations of repair verbs (at V2) to those in conjuncts. We also automatically extracted pitch information using the pitch detection algorithm from the Edinburgh Speech Tools Library (King, Black, Taylor, Caley, & Clark, 2003). There were no differences neither in mean duration (see Table 1), nor in mean pitch (repairs: 139 Hz; conjuncts: 156 Hz).