Disfluency in Dialogue: Attention, Structure and Function
Nicholson, Hannele Buffy Marie
Spontaneous speech is replete with disfluencies: pauses, hesitations, restarts, and less than ideal deliveries of information. Disfluency is a topic of interdisciplinary research with insights from psycholinguistics, phonetics and speech technology. Researchers have tried to determine: When does disfluency occur?, Can disfluency be reliably predicted to occur?, and ultimately, Why does disfluency occur? The focus of my thesis will be to address the question of why disfluency occurs by reporting the results of analyses of disfluency frequency and the relationship between disfluency and eye gaze in a collaborative dialogue. Psycholinguistic studies of disfluency and collaborative dialogue differ on their answers to why disfluency occurs and its role in dialogue. One hypothesis, which I will refer to as Strategic Modelling, suggests that disfluencies are designed by the speaker. According to the alternative view, which I will call the Cognitive Burden View, disfluency is the result of an overburdened language production system. Throughout this thesis, I will contrast these two theories for an ultimate answer to why disfluency occurs. Each hypothesis attaches a functional role to a structural definition of disfluency and therefore in order to determine why disfluency occurs, I will contrast the structural and functional characteristics of disfluency. I will attempt to do this by analysing the dialogue behaviour in terms of speech goals and eye gaze behaviour a speaker is engaged in when they make certain types of disfluencies. A multi-modal Map Task paradigm was used in this thesis, in which speakers were asked to describe the route on a cartoon map to a distant confederate listener who provided either visual or verbal feedback. Speakers were eye-tracked during the dialogue and a record was kept of when the speaker attended to the listener’s visual feedback. Experiment 1 tested the visual feedback paradigm to establish its validity as a baseline condition. Speakers were found to make more disfluencies when they could interact with the visual feedback, suggesting disfluency is more common in interactive circumstances. Experiment 2 added verbal feedback to the experimental paradigm to test whether listeners react differently to the two modalities of feedback. Speakers made more disfluencies when the feedback was more complicated. Structural disfluency types were also observed to fulfil different functions. Finally, Experiment 3 manipulated the motivation of the speaker and found that Motivated speakers gazed more often and were more disfluent per opportunity than Control speakers suggesting that highly motivated subjects are more willing to engage in difficult tasks.