📄 Implications for Generating Clarification Requests in Task-Oriented Dialogues

Summary

The paper asks how a spoken dialogue system should formulate clarification requests (CRs) so that they show what went wrong in understanding. Rieser and Moore annotated the human-human travel-booking dialogues of the Carnegie Mellon Communicator Corpus. They used an extended version of the multidimensional Rodriguez and Schlangen (R&S) scheme, in which form features (mood, completeness, relation to antecedent, boundary tone) are kept apart from function features (source, severity, extent, answer, satisfaction). Source follows Clark's four grounding levels.

They compared their counts with R&S's study of the German Bielefeld Corpus and with Purver, Ginzburg and Healey's BNC study. The two task-oriented corpora showed no significant differences, but both differed clearly from everyday conversation. This suggests that dialogue type, modality and channel quality shape when and at which level people clarify.

In task dialogues, people tend to clarify immediately and to offer a hypothesis for confirmation rather than ask for repetition. Chi-square tests found significant form–function correlations, which can serve as rules for generating natural CRs that signal the problem source.

Key points

• Data:
  • 31 Communicator dialogues: 2098 turns and 19,395 words of transcribed telephone speech.
  • 98 CRs in total, which is 4.6% of turns.
• Extensions to the R&S scheme:
  • Relation-antecedent now separates add from repet-add.
  • Source gains the values belief and ambiguity refinement.
  • Severity gains cont-disamb, following Gabsdil.
• Most common form: partial declarative questions, at 53.1% of CRs. Rising boundary tone accounts for about 74%.
• Function distribution:
  • Source: np-reference 40%, acoustic 31%.
  • Severity: about 73% ask for confirmation of a hypothesis.
  • Extent: 80% point to a specific element.
• Form–function tests: Pearson's χ² with Yates correction.
  • Source correlates with relation-antecedent and mood.
  • Severity and answer correlate with mood.
  • Extent correlates with completeness.
  • Acoustic problems lead to literal repetition. Reference problems lead to reformulation or repetition-with-addition.

Interesting ideas

• I1. Describe a clarification act along two separate axes, function (why) and form (how), then learn the mapping between them; this keeps the generator's decision step independent of its surface realisation step.
• I2. Tie the source of a problem to grounding levels (contact, acoustic, reference, intention) so that a request reveals which level failed; an error label then carries diagnostic information rather than just "not understood".
• I3. Prefer offering the most plausible hypothesis, drawn from dialogue history and world knowledge, over asking for repetition; humans do this to stay efficient while still checking.
• I4. Annotate a request's meaning partly by what happens next, using the addressee's reply and whether the asker renews the request, instead of fixing its reading in advance.
• I5. Compare behaviour across corpora that differ in genre and channel quality to see which regularities hold and which are domain-specific.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (back): When Mova cannot resolve something, it could first pick a function (source plus severity), then choose a surface pattern from a function→form table, so each generated question stays explainable.
• I2 (frames): Failed coreference or role assignment could be labelled with the level where it broke, such as np-reference versus intention, and passed on as an explanation instead of a silent fallback.
• I3 (world): When an entity or date is underspecified (e.g., a day without a month), the world model could fill in the most reasonable value and mark it as a confirmable assumption rather than leaving a gap.
• I4 (learn): Ambiguous silver annotations could be resolved by downstream consequences, for example whether the follow-up text confirms an interpretation, as a contextual check on witness disagreement.
• I5 (learn): Evaluation could compare rule statistics across fables, school texts and dialogue, flagging rules whose distribution shifts as genre-dependent rather than universal.