| dc.description.abstract | This thesis provides a quantitative examination of multiparty casual talk, a speech genre which is fundamental to human social life. Multiparty casual conversation has been reported to comprise chat and chunk phases, where chat is interactive talk involving several participants and chunk is talk where one speaker dominates, holding the floor for a stretch of time. This thesis examines the current understanding of multiparty casual conversation, and reports several corpus studies performed in order to form a more accurate picture of its overall characteristics, and to compare and contrast its constituent chat and chunk phases in terms of composition and of conversation dynamics. This work aims to gain a quantitative understanding of the differences between chat and chunk phases, previously described by example, by analysis of speech, silence, laughter, conversation dynamics, and disfluency.
A collection of audio and video recordings of long (c. 1 hour) multiparty casual conversations (Casu- alTalk) was segmented, and annotated for speech, silence, laughter, and a subsection was annotated for disfluency. A novel chat and chunk annotation scheme was devised, tested, and applied to the corpus. The corpus was then segmented into novel floor state labels which reflected the conversation state at any instant in time.
The examination of the general structure of the casual conversations in CasualTalk supported accounts in the literature of such talk’s structure of chat and chunk phases. The annotation process showed that chat and chunk phases can be annotated with a high degree of inter-annotator agreement. Experiments on the compositions of chat and chunk phases showed that these do differ significantly in their distribution of speech and silence, proportion of laughter and overlap, and to a lesser extent in disfluency. For dynamics, there were clear differences in the liveliness of chat and chunk, both in terms of existing compression measures and a novel floor state change rate metric. The dynamics around turn change and retention also differed significantly. Analysis based on the identification of simple and complex within and between speaker transitions showed that more than half of all transitions are complex, and thus can be misclassified unless there is careful thresholding of the duration of single party speech bounding turns. Chat and chunk segments were compared to the task-based TEAMS Corpus, revealing significant differences in the distribution of silence and overlap, pointing to the need to consider and analyse different genres or speech exchange systems separately to form an accurate picture of their dynamics.
The results found will be useful in the design of more human-like dialog systems for a range of applications, particularly those requiring users to converse with the system in entertainment, education, or healthcare settings. | en |
