Beyond Repair – Testing the Limits of the Conversational Repair System

We report on an experiment on the effects of inducing acoustic understanding problems in task-oriented dialogue. We found that despite causing real problems w.r.t. task performance, many instances of induced problems were not explicitly repaired by the dialogue participants. Almost all repairs referred to the immediately preceding utterance, with problems in prior utterances left unacknowledged. Clarification requests of certain forms were in this corpus more likely to trigger reformulations than repetitions, unlike in different settings.