Abstract - Chris Howes October 18th 2012
Finishing each other's . . .Responding to incomplete contributions in dialogue
Christine Howes, Patrick G.T. Healey, Matthew Purver, Arash Eshghi
The full paper as presented at COGSCI 2012 can be found here
Introduction and Method
A distinguishing feature of dialogue is that contributions can be fragmentary or incomplete. Such incomplete utterances may be later completed by another interlocutor (Purver et al, 2009). These cross-person compound contributions (CCs) are a paradigmatic feature of dialogue and have been hypothesised to be more likely in predictable contexts (Lerner, 1991) but the contributions of different sources of predictability has not been systematically investigated. We present an experiment which is, to the best of our knowledge, the first to ever systematically attempt to induce continuations in an ongoing dialogue.
Using the DiET chat tool, we artificially truncate genuine contributions in ongoing text-based dialogues providing other participants with the opportunity to provide completions. This intervention is introduced systematically, in real time with the truncation point manipulated to vary the lexical and syntactic predictability of what comes next using an entropy calculation. Pragmatic predictability – whether the intervened turn contributed to an ongoing topic under discussion or not, and whether the turn could have been considered complete at the truncation point – was subsequently annotated by the authors.
Results and Discussion
The results look at what type of response (if any) participants produce to the apparently incomplete turn. They show that what is critical to the likelihood of one's interlocutor supplying a continuation is the accessibility of common ground. While people are sensitive to syntactic predictability, this alone is insufficient to prompt a completion. Participants make use of syntactic predictability only if the context is sufficiently constrained. Though people do respect syntactic constraints when producing continuations, truncation at different syntactic points in the sentence does not cause any difference in difficulty in producing them. Clarification requests, in contrast, are more likely, and more likely to be formulated as continuations, when the syntactic category of the upcoming material is more predictable; this suggests that while the grammar is a mutually available resource, it is not used in the same way by all interlocutors, with syntax able to be exploited to localise the source of a potential misunderstanding.
Another of the main findings is that people are sensitive to potential turn endings. These may be syntactic but they are not necessarily so. Some cases which appear to be syntactically incomplete can be responded to as if they are complete, provided that the continuation is highly predictable. If there are indeed cases which are interpreted as complete when they are not – as if the hearer is supplying the missing material internally, but does not necessarily produce it, this has implications for any grammatical or dialogue model. Incomplete syntactic strings must be not only successfully analysed, but also assigned potentially complete semantic representations.
Lerner, G. H. (1991). On the syntax of sentences-in-progress. Language in Society
Purver, M., Howes, C., Gregoromichelaki, E., and Healey, P. G. T. (2009). Split utterances in dialogue: A corpus study. In Proceedings of the 10th Annual SIGDIAL Meeting on Discourse and Dialogue (SIGDIAL 2009 Conference), London, UK.
Back to IMC Seminar Series