Exploring the Role of Laughter in Multiparty Conversation

Conversation is widely studied through corpus analysis, often concentrating on ‘task-based’ interactions such as information gap activities (map-tasks [1], spot the difference [2], ranking items [3]) and real or staged business meetings [4], [5]. This task-based dialogue (on which spoken dialogue technology is based [6]) relies heavily on verbal information exchange. However, the immediate task in natural conversation is often not so clear and the purpose of some interaction may be best described as social bonding. Laughter is universally observed in human interaction. It is multimodal: a stereotyped exhalation from the mouth in conjunction with rhythmic head and body movement [7]. It is part of the gesture call system, older than language [8], predominantly social rather than solo, and aiding social bonding [9]. It punctuates speech [10], and manifests in a range of forms [11]. We investigate laughter in situ, using corpora of nonscripted (spontaneous) multiparty interaction: the task-oriented AMI meetings corpus [5], and the conversational TableTalk [12], d64 [13], and DANS corpora. We address laughter and topic change, multimodal aspects of laughter, and the interplay of laughter and bio-signals. In earlier work on topic change in AMI and TableTalk we found that laughter, and especially shared laughter, is likely near topic change in both corpora, with a stronger effect in TableTalk, and that the number of people laughing together grows with proximity to topic change in TableTalk [14], [15]. These results on multiparty interaction reflect the literature on laughter in two-party dialogue [16], [17], which points towards discourse functions for laughter as a topic termination mechanism. To investigate whether these findings reflect a general phenomenon we extend this temporal analysis to the DANS Corpus. We speculate that laughter may function as a strategy to instigate a topic change, or as a marker of topic exhaustion providing a buffer against an embarrassing silence. We are examining laughter in terms of speaker role (who speaks/laughs first and last, etc.) and turn-taking activity to better understand its function. Our work on multimodality investigates the perception of audio and visual laughter cues by naïve annotators, to investigate whether they can reliably spot unimodally. We are also exploring the interplay of laughter and electro-dermal activity (EDA), linked to levels of emotional arousal [18] and to cognitive load [19]. Social chat has been linked to implicit processing, which is reported to involve lower cognitive load [20], while laughter has been observed to be more frequent in social than in task-based dialogue [21].