Incremental Processing of Laughter in Interaction

In dialogue, laughter is a frequent non-verbal signal that can precede, follow, or overlap its antecedent—the laughable. Furthermore, the time alignment between the laughter and the laughable is dependent on who produces the laughable, whether laughter overlaps or not with speech and the communicative act performed. Laughter can interrupt either one’s own or one’s conversational partners’ utterances and, like other well-studied features of dialogue such as repair and split utterances, this interruption does not necessarily occur at phrase boundaries. Similarly, much like repair and other feedback like backchannels, laughters can be categorised as forward-looking or backward-looking. Given these parallels, we propose an analysis of how laughter can be processed and integrated using a Dynamic Syntax (DS) model, which already has well-motivated accounts of repair, split utterances, and feedback. We present a corpus study of laughter in dialogue, as well as a model using Dynamic Syntax and Theory of Types with Records (DS-TTR). Analogously to pronouns and ellipsis, our approach uses underspecification to account for laughter types that are different in processing terms as anaphoric or cataphoric, and demonstrates how laughter is processed incrementally as an utterance unfolds. Our analysis covers ≈87% of the annotated corpus data.