Pat Healey (Queen Mary University of London), 'Making a Contribution. Experiments on how people construct conversational turns'
- They X-rayed me, and took a urine sample, took a blood sample. Er, the doctor...
- Chorlton, mhm, he examined me [. . .]
(From the British National Corpus)
B's question clarifies the reference of 'doctor' before A has completed their turn. This highlights the fact that contributions to dialogue are often built by more than one participant and that this process is incremental. This raises two basic questions about language processing in dialogue:
- How are the increments of contributions organised?
- How does collaboration with a conversational partner affect the processing of those increments?
The organisation of increments affects how parts of a contribution can be accessed or amended. A word- or string-based processing model predicts that the distance between the target of a clarification request (A's "the doctor") and the point of interruption (where B says "Chorlton?") is a key factor in processing. A `flat' semantic model, in which phrases introduce discourse referents into an unstructured record (e.g. Kamp & Reyle, 1993) predicts no effect of interruption point. A model based on constituent structure, either syntactic (e.g. Chomsky, 1957) or semantic (e.g. Kempson et. al. 2001), predicts effects of structure at the point of interruption. Collaboration can affect processing because A and B must suspend construction of the current turn, engage in a clarification sub-dialogue, and resume the original turn (Chevalley and Bangerter, 2010).
We report a 'chat tool' experiment that explores the effects of interruptive clarification requests on live text-based dialogues. The technique involves inserting 'spoof' clarification requests, that appear to originate from another participant, while a turn is being constructed. We track the structure of turns in real-time as they are typed and this enables us to manipulate the interruption point. We then insert probe CRs – verbatim repeats of the target NP plus a question mark (Reprise Fragments) – into the dialogue. Using this approach we test the effects of two factors:
- Insertion Point: The spoof clarification request is inserted either a) within an incomplete constituent (e.g. after a determiner) or b) at the end of a complete constituent (e.g. after a completed NP).
- Apparent Origin: (Between-subjects) The probe CR appears to originate either from the actual conversational partner or from a chatbot ('Q-Bot').
The results provide evidence for collaborative processing effects: the same question in the same dialogue context gets different answers depending on its origin. Specifically, the interrupted turn more likely to be re-constructed, i.e. restarted or recycled more the for the active participant. We argue that this provides evidence of participant-specific indexing of context in dialogue. The results also provide evidence of sensitivity to constituent structure. The structure at the interruption point affects how easy it is to answer the question. Within-constituent interruptions are more disruptive (cause more restarts and more back tracking) even though they are nearer the target. We argue that this provides evidence that increments in turn production are organised in terms of structured (grounded?) constituents and points to the need for a structured, word-by-word incremental model of dialogue context (e.g. Cann, Kempson and Purver, 2007). We conclude that people work to construct an agreed, public, record of 'what has been said' in a conversation. When someone outside the conversation interrupts with a question they just answer it. However, when their partner interrupts they work to reconstruct the problematic contribution (turn), often reformulating the whole clause.