Computational Linguistics Colloquium

Thursday, 6 December 2012, 16:15
Conference room, Building C7.4

The anticipation of the end of the current speaker's turn in natural conversation

Listeners in a conversation need not only process the content of speakers' utterances and prepare responses, but they also need to anticipate the moment in time that speakers finish their turn, allowing them (the listener) to start their own contribution right on time.

But how are listeners able to be so accurate in projecting the end of a speaker's turn? What information do they use? Some authors have suggested that syntactic information provides the main cue, while others have argued that the intonational contour (F0, melody) is the listener's main source of information for accurate ‘projection’.

In my talk, I will describe an on-line experiment in which we tested these hypotheses by independently manipulating the presence of lexico-syntactic content and the intonational contour of utterances. Importantly, these recordings were culled from natural conversations.

When hearing the original recordings, subjects could indeed anticipate turn endings with the same high degree of accuracy found in real conversations. When we entirely removed the intonational contour from the utterances (leaving words and syntax intact), there was no change in subjects accuracy of end-of-turn projection - they were just as accurate as with the original recordings. But when we made the words unrecognizable, while leaving intonational contour intact, subjects' performance deteriorated significantly. These results establish that the lexico/syntactic content of an utterance is necessary (and possibly even sufficient) for projecting the moment of its completion, and thus for regulating conversational turn-taking. By contrast, and perhaps surprisingly, intonational contour is neither necessary nor sufficient for end-of-turn projection.

I will also address some of the consequences of these findings for the computational modeling of projective turn-taking in artificial systems.

If you would like to meet with the speaker, please contact Pirita Pyykkönen-Klauck.