Finishing each other’s … Responding to incomplete contributions in dialogue

0
467

Finishing each other’s . . . Responding to incomplete contributions in dialogue Christine Howes, Patrick G. T. Healey, Matthew Purver, Arash Eshghi {chrizba, ph, mpurver, arash}@eecs.qmul.ac.uk Queen Mary University of London Interaction, Media and Communication Research Group School of Electronic Engineering and Computer Science, London E1 4NS, UK Abstract Hypothesis 1 Cross-person completions likely at transition relevance places A distinguishing feature of dialogue is that contributions can be fragmentary or incomplete. Such incomplete ut- terances may be later completed by another interlocu- tor. These cross-person compound contributions (CCs) have been hypothesised to be more likely in predictable contexts but the contributions of different sources of pre- dictability has not been systematically investigated. In this paper we present an experiment which artificially truncates genuine contributions in ongoing text-based dialogues, to investigate the effects of lexical, syntac- tic and pragmatic predictability of the truncation point on the likelihood of one’s interlocutor supplying a con- tinuation. We show that what is critical is the actual and presumed accessibility of common ground, and that while people are sensitive to syntactic predictability, this alone is insufficient to prompt a completion. Keywords: Dialogue; compound contributions; com- mon ground. are more Second, completions should tend to occur at syntacti- cally projectable points (e.g. compound turn construc- tional units Lerner, 1991). Hypothesis 2 Cross-person completions are likely when they are syntactically predictable. more A third source of predictability comes from the degree to which the speaker and hearer share, or can be assumed to share, common ground relevant to the CC. If the topic of the utterance is already in the common ground then the content of the completion is more predictable. Hypothesis 3 Cross-person completions are more likely when they address topics that are part of the common ground. Introduction The effects of these different forms of predictability are directly tested here for the first time using a text chat experiment performed with the DiET experimental platform. The evidence points towards shared knowledge being a key factor with other sources of predictability also contributing. It is well known that contributions to dialogue are often fragmentary or in some sense unfinished Fern´andez and Ginzburg (2002). These incomplete utterances may be subsequently completed, either by the original speaker following some response or interruption from an inter- locutor, or, by another person (Purver et al., 2009). These compound contributions (CCs) are a paradig- matic feature of dialogue, and cross-person CCs in par- ticular are a key indicator of coordination between in- terlocutors. Although naturally occurring cross-person CCs and their interpretations have been studied (Lerner, 1996; Purver et al., 2009), there has not previously been a systematic, experimental, attempt to investigate the factors that influence how a completion for an incom- plete utterance may be produced. Intuitively, people’s willingness to finish another person’s incomplete utter- ance will depend (at least) on how predictable the rest of the utterance is. There are several sources of possible predictability. Expansions are CCs which add material (e.g. an ad- junct) to an already complete syntactic element; com- pletions are CCs which complete an incomplete element. Conversation analytic (CA) discussions of CCs suggest that they should preferably occur at transition relevance places (TRPs), points that are foreseeable by the partic- ipants. Expansions are CCs with split points at TRPs, and are more common in spoken dialogue (Howes et al., 2011) so ought to be more likely than completions. Method In this experiment, to see what factors influence how people respond to unfinished turns and their likelihood of producing a continuation, a number of genuine single contributions in dyadic text-based conversations were ar- tificially split into two parts, using the DiET chat tool. The DiET chat tool The Dialogue Experimental Toolkit (DiET) chat tool is a text-based chat interface into which interventions can be introduced into a dialogue in real time. These in- terventions can take a number of forms; turns may not be relayed, additional turns may be added, as in Healey et al. (2003), in which spoof clarification requests are added to the dialogue, or turns may be altered prior to transmission. As these manipulations occur as the dia- logue progresses, they cause a minimum of disruption to the ‘flow’ of the conversation.