📄 The independence of dimensions in multidimensional dialogue act annotation

Summary

Utterances in natural dialogue frequently do several jobs at once. This has led to multidimensional dialogue act annotation. The DIT++ scheme sorts communicative functions into ten dimensions: Task, Auto-feedback, Allo-feedback, OCM, PCM, Time Management, Contact Management, Turn Management, Dialogue Structuring and SOM. It treats a dimension as an aspect of interaction that has its own dedicated acts and can be addressed separately from the others.

Petukhova and Bunt test whether this independence actually holds in data. They use three manually segmented corpora tagged with DIT++: DIAMOND, AMI and OVIS. They run four tests:

• cross-dimension co-occurrence frequencies;
• phi coefficients between dimension pairs;
• whether each dimension is ever the only one an utterance addresses;
• frequencies of a dimension given the previous dimension.

Each analysis is done twice, once with and once without implicated and entailed functions. No strong positive or negative association appears, and every dimension sometimes occurs alone. The authors therefore treat the ten dimensions as effectively independent. This supports modular annotation schemes, where dimensions can be dropped, added or re-tagged without disturbing the rest.

Key points

• Corpora:
  • DIAMOND: Dutch, two-party, instructional, 1,408 utterances.
  • AMI: English, multi-party, task-oriented, 3,897 utterances.
  • OVIS: Dutch, human-computer, public transport information, 3,942 utterances.
• Phi values on AMI never reach ±.7.
  • Weak positive links: Turn–Time Management (.6) and OCM–Turn Management (.4).
  • Weak negative links: Task–Auto-feedback (−.5, without entailed/implicated functions), and Task–Contact and Auto–Allo-feedback (−.6 each, with them).
• Share of utterances addressing only one dimension:
  • Task 28.8%, Auto-feedback 14.2%, Turn Management 7.4%, Discourse Structuring 1.9%, Allo-feedback 0.7%.
  • OCM 0.5%, Time Management 0.3%, SOM 0.3%, PCM 0.2%, Contact Management 0.1%.
• Some pairs never co-occur, such as Contact+Time and PCM+Discourse Structuring. However, the tag definitions contain nothing that forbids these combinations.
• Counting entailed/implicated functions raises co-occurrence. Answers, accepts and rejects entail positive feedback, and questions often carry a turn-releasing function.

Interesting ideas

• I1. Independence of two label layers is defined operationally: one layer can be filled whether or not the other is, and a value in one never fixes the value in the other; this gives a checkable criterion for schema design.
• I2. Annotate explicit functions and inferred (entailed or implicated) functions as separate tiers, then compare statistics with and without the inferred ones; this shows how much structure comes from inference rather than surface form.
• I3. A "does it ever occur alone?" test exposes categories that only exist as add-ons to others; a dimension never seen solo is suspect.
• I4. Counting pairs of (current tag, previous tag) reveals purely reactive categories, such as answers, which cannot exist without a preceding question.
• I5. Zero co-occurrence is not taken as proof of dependence; the authors first check the definitions for a real combination restriction.

For Mova

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

• I1 (frames): Apply this criterion to the speech-act layers in prag to confirm that each layer (e.g. task act vs. feedback vs. turn) is a separate tier, not a disguised duplicate.
• I2 (frames): Keep implicature output in prag as its own layer, so a thank-you's implied positive feedback is traceable and can be switched off for diagnostics.
• I3 (learn): When evaluating label sets for silver data, flag labels that never appear without a companion label as candidates for merging or redefinition.
• I4 (answer): In QA on fables and tales, use previous-act statistics to check that answer-type frames are anchored to an actual question in the dialogue before treating them as replies.
• I5 (learn): Before turning an empty cell in a co-occurrence matrix into a hard repair rule, require a definitional reason, so data sparsity is not mistaken for a constraint.