We report a dialogue task which investigates how the mechanisms of miscommunication contribute toward referential coordination. Participants communicate via a text-based instant messaging tool which is used to identify turns that were edited prior to sending. These turns are transformed by the server into artificial selfcorrections, and sent to the participants. The patterns observed in the dialogues show that these interventions have a beneficial effect on referential coordination.
A central finding in research on dialogue is that
interlocutors rapidly converge on referring
expressions
The development of systematicity is not
simply due to the coordination problem of creating a
novel referring expression: once referring
expressions have been used successfully, they continue
to develop
Cumulatively, these findings suggest that
processing that occurs in dialogue places important
constraints on the semantics of referring
expressions. However, there is currently no consensus
about how best to account for how convergence
develops. The iterated learning model of
Further, a series of experiments (Healey and
Mills, 2006; Mills and Healey, 2008) suggest
that the development of abstraction can be driven
by participants encountering and resolving
problematic understanding. In these experiments,
participants played an online version of the
maze game
A: Go to the 3rd row 1st box B: row? (produced by the server) A: yeah from the top When participants received these interventions, they produced less abstract descriptions. However, once the interventions stopped, participants subsequently used more abstract descriptions than participants who had received no interventions (Mills, 2015).
In a subsequent experiment (Healey, Mills, Eshghi, 2013) , this methodology was used to automatically detect naturally occurring clarification requests and transform them into more severe signals of miscommunication. For example in the following conversation between two participants A and B, B’s clarification request “5th?” is intercepted and transformed into “what?” and sent to A.
A: go to the 5th row 2nd square
B: 5th? (intercepted by server)
B: what? (transformed turn sent to B)
A: yeah from the top
Notice that this transformation reduces the
diagnostic specificity of the clarification request; A has
less evidence of B’s level of (mis)understanding.
Since there is an expectation that a conversational
partner should provide diagnostic information
that is sufficient to resolve misunderstanding
Taken together, these results suggest that (1) When interlocutors encounter problematic understanding, they initially decrease the level of abstraction of their referring expressions, allowing them to identify and diagnose the nature of the misunderstanding, and (2) Once the problem has been resolved, this subsequently allows the participants to coordinate on even more abstract and systematized referring expressions.
However, these experiments have focused solely on ”trouble” that is signalled in clarification requests about the content of another’s turns, i.e. in ”other-initiated” repair (Schegloff, 2007). It is currently unclear whether negative evidence in self-repair might also have an effect on the development of abstract referring conventions. 2
To investigate in closer detail how negative evidence might contribute toward convergence, we report a variant of the maze-task. Here too, participants communicate with each other via an experimental chat tool which automatically transforms participants’ private turn-revisions into public self-repairs that are made visible to the other participant. For example, if a participant, A types: A: Now go to the square on the left, next to the big block on top
and then before sending, A revises the turn to: A: Now go to the square on the left, next to the third column The chat server automatically detects the left-most boundary of the edited portion of the turn and inserts a hesitation marker (e.g. “umm” or “uhhh” immediately preceding the revision), followed by the text that was deleted. This would yield the following turn, sent to B: :
A: Now go to the square on the left, next to the big block on top umm..I meant next to the third column
Two self-repair formats were used:
Interventions were performed symmetrically on both members of a dyad. No participants reported detecting the experimental manipulation. Examining the transcripts showed that participants who received these transformed turns used more abstract Cartesian location descriptions than participants in a baseline condition. This pattern was already apparent after 5 minutes in the task. Task performance followed a different pattern initially participants who received these interventions performed worse completing fewer mazes and requiring more moves to solve each maze. However, by the end of the task, participants who received the interventions performed at the same level as participants in the baseline condition. Crucially, participants who received the transformed turns continued to use more abstract descriptions. 4
The patterns observed in the maze game dialogues show that the interventions have a beneficial effect on semantic coordination. However, it is currently unclear how the constituent components of the self-repairs contributed: It could be that this effect is due entirely to the hesitation markers. Conversely, it is possible that this effect is due solely to participants reading the deleted text. If so, it is possible that the deleted text provides additional information about the other’s level of (mis)understanding. It could also be that the deleted text makes the dialogue less coherent, forcing participants to compensate for the perturbation caused by the interventions.
Since participants encountered multiple interventions per trial, it is not possible to distinguish between the effects of the individual components. However, in aggregate we argue that the artificial self-repairs having a beneficial effect of amplifying naturally occurring signals of miscommunication: the artificially generated disfluencies and reformulations are used by participants as cues that their partner is having difficulty coordinating on the semantics of referring expressions. Consequently, participants expend more effort to address these problems and once these problems have been identified and resolved, dyads are able to converge quicker on more stable and more abstract referring schemas.