📄 Discourse Chunking: A Tool in Dialogue Act Tagging

Summary

The paper addresses dialogue act (DA) tagging, where most systems rely on words, n-grams or prosody and use little real discourse structure. Midgley proposes "discourse chunking," a lightweight way to divide a dialogue into three phases: a greeting phase, a negotiation phase made of repeated topic cycles, and a farewell phase. Boundaries come from hand-written rules triggered by DA tags that introduce a topic (INIT, SUGGEST, REQUEST_SUGGEST, EXCLUDE) and by the first BYE. Each utterance's position inside its chunk is rescaled to a 1–100 range. This position is added as a feature to a case-based reasoning tagger, a similarity-weighted k-nearest-neighbour classifier, trained and tested on the Verbmobil-2 corpus with its 32-tag set. Adding the chunk feature raised accuracy from about 54% to about 65%. The largest gains were on short, ambiguous replies such as "okay" and "yeah." The result suggests that cheap structural cues can stand in for expensive prosodic analysis. One caveat is that the test data used gold chunk labels.

Key points

• Data: Verbmobil-2. The case base has 59 hand-tagged dialogues (8398 utterances), and the test set has 20 dialogues (2604 utterances). Chunks were assigned to the case base automatically by rule.
• Tagger: k-NN with k = 7, where each neighbour's vote is weighted by its similarity to the test utterance.
  • Features include speaker change, word number, word and n-gram similarity, and the previous and second-previous DA tags, which are predicted rather than gold.
  • A discourse chunk position feature is also added.
• Feature weights were tuned by Evolutionary Programming:
  • population of 100;
  • the best 10 survive and mutate with Gaussian noise;
  • this repeats for 10 generations.
• Ten frequent BNC words serve as stopwords. They are dropped for word similarity but kept for n-gram similarity.
• Results: overall accuracy rose from 53.68% to 65.44%, against an estimated human ceiling of about 84%.
  • Gains appeared in nearly all tags, including tags that do not start chunks.
  • EXCLUDE, a chunk starter, got slightly worse.
  • One-word utterances improved noticeably.

Interesting ideas

• I1. Instead of annotating discourse structure separately, the author builds segments deterministically from labels that already exist (topic-introducing acts start a new segment), so no new annotation effort is needed.
• I2. Position within a segment is normalized to 1–100, and two positions are compared as a ratio. This makes "near the start" versus "near the end" comparable across segments of any length.
• I3. Ambiguous short replies ("yeah," "okay") are resolved by their place in a cycle: near a proposal they lean toward acceptance, and later they act as wind-down feedback.
• I4. To check whether a feature merely leaks label information, the author tests if gains are confined to the labels that define it or appear across all labels.
• I5. Stopwords are removed for bag-level word overlap but kept for sequence matching, because function words help tell similar word orders apart.

For Mova

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

• I1 (frames): prag could split a text into topic or episode segments by rule whenever it detects a proposal or topic-raising speech act, giving each later act an explainable structural context.
• I2 (world): event ordering could carry a normalized relative position inside each scene, helping decide whether a step opens or wraps up an episode.
• I3 (frames): the speech-act classifier could label bare affirmatives as ACCEPT or as acknowledgement depending on their distance from the last proposal, with that rule recorded as the explanation.
• I4 (learn): this is a cheap negative-control pattern; when a feature is derived from labels, compare per-class gains on derived versus non-derived classes before trusting the improvement.
• I5 (tokens): similarity or lookup utilities in crate en could keep function words for sequence-sensitive matching and drop them only for content-overlap scores.