How listeners respond to speaker's troubles

How listeners respond to speaker’s troubles Patrick G. T. Healey, Mary Lavelle, Christine Howes , Stuart Battersby, Rose McCabe ph@eecs.qmul.ac.uk Queen Mary University of London, Cognitive Science Research Group, School of Electronic Engineering and Computer Science, London E1 4NS, UK Abstract Listeners normally provide speakers with simultaneous feedback such as nods, “yeah”s and “mhm”s. These ‘backchannels’ are important in helping speakers to talk effectively. Two factors are known to influence when a backchannel is produced; if the speaker is looking at the listener or if the speaker is presenting new information. We investigate a third factor: whether the speaker is having trouble speaking i.e. self-repair. If dialogue is an active collaborative process then listener’s responses should be especially critical when trouble is encountered. Using data from a corpus of three person dialogues we show that speaker’s rate of self-repair is a better predic- tor of listener responses than speech rate. We also show that listeners respond strongly to speaker troubles in- dependently of whether the speaker is looking at them. We argue that it is the points at which conversation threatens to go off-course that are most significant for coordination. Keywords: Gesture; repair; dialogue Introduction Listening in conversation is not a passive activity. As Goffman (1955) noted, what listeners do while being addressed has important consequences for the way that speakers produce their turns. Goffman distinguished be- tween two general kinds of listener feedback; displays of attention and understanding of what is said and the sig- nalling of interactional functions such as a desire to speak next. Yngve (1970) introduced the term ‘backchannel’ to describe these uses of simultanous feedback that provide speakers with concurrent information about how their turn is being received. In a series of experiements examining the effects of lis- tener response behaviours Bavelas and colleagues were able to show that the fluency and effectiveness of a speaker’s turns depends directly on the level of feedback they are getting from their addressees (J. B. Bavelas et al., 2000; J. Bavelas et al., 2006). People telling stories to listeners who are engaged in a distractor task speak less fluently and are less compelling than those whose listeners are attending more carefully. Given the importance of listener responses for success- ful interaction a key question is what prompts a listener to produce them? Many of the most common backchan- nel signals, such as nods and smiles, use the visual chan- nel which avoids potential competition with concurrent speech. One common finding in the literature is that ad- dressee responses are reliably correlated with speaker’s gaze. Goodwin (1979) observed that speakers will peri- odically check whether addressees are attending by look- ing at them and if they get no response may restart or switch to a new addressee mid-turn. J. B. Bavelas et al. (2002) found that listener responses in their ‘close call’ story telling task were significantly more likely to occur in a ‘gaze window’ i.e. when a speaker is looking at a listener than when they are not. A second common observation in the literature is that backchannels are also associated with the introduction of new information into a dialogue such as the introduction of a new referent or proposal that may warrant some signal of interim acknowledgement or acceptance before the speaker’s turn is completed (J. Bavelas et al., 2006; Clark & Wilkes-Gibbs, 1986; Yngve, 1970). In this case it is the information update that prompts the use of a backchannel to signal understanding ‘so far’ (Goodwin, In this paper we explore the effects of a third factor on listener responses: the degree of difficulty a speaker has in producing their turn. Few conversational turns are produced without some form of online revision or refor- mulation during their production. Sometimes referred to as disfluencies these self-repairs are indicative of some sort of trouble producing a turn. If conversation is a collaborative process in which each turn is co-produced (Goodwin, 1979; Clark, 1996) then this leads to the hy- pothesis that the points at which the speaker shows signs of getting into trouble ought to be especially critical for collaborative reponses. This paper tests this hypothe- sis by investigating the relationship between nodding, speech rate and repair rate in a corpus of three person dialogues. Methods Experimental work on listener backchannel responses has focussed only on dyadic, i.e. two person, interactions. However, natural interactions frequently involve more than two people (Goffman, 1981; Eshghi, 2009). For current purposes three-way interactions also have the practical advantage that they make it possible to com- pare two kinds of listener depending on who the speaker is looking at while they talk. Given the importance of speaker gaze to the production of backchannels this pro- vides a useful opportunity to compare the responsiveness