📄 Rare Dialogue Acts in Oncology Consultations

Summary

The paper argues that certain dialogue acts (DAs), which large conversational corpora treat as marginal, are routine in conversations between cancer patients and health professionals. The data are 37 recorded dialogues between Macmillan Cancer Care nurses and their patients. These were collected by the Psychological Medicine Group at the University of Manchester for communication-skills research. The authors have not yet fully annotated the corpus. Instead, they present annotated excerpts and set them against the frequency figures reported for the Switchboard corpus tagged with SWBD-DAMSL. Five DA types stand out: collaborative completions, summaries/reformulations, thanks, apologies, and long, negotiated closings. Most of these fall in Switchboard's rarest band. In the oncology data they appear regularly and serve two purposes: securing shared factual and emotional understanding, and building rapport. The authors trace this to the health-critical, emotionally charged nature of the consultations and compare it with the read-back confirmations used in Air Traffic Control. The work matters because it shows that DA distributions depend on the domain. Tagsets and statistics built on casual talk can underrepresent exactly the acts that matter in high-stakes dialogue.

Key points

• Corpus: 37 nurse–patient dialogues, each 200–1200 utterances long (typically 300–600). The patients' names have been changed for confidentiality.
• Baseline is Switchboard with SWBD-DAMSL (205,000 utterances):
  • Statements, Continuers and Opinions together cover 68% of utterances.
  • The original 220 tags were collapsed to 42 because many were too rare.
  • Of those 42, 32 are under 1% frequency and 25 are under 0.5%.
• Switchboard counts for the acts discussed:
  • Collaborative completions: 699 (about 0.3%).
  • Summarize/reformulate: 919 (about 0.5%).
  • Thanks: 67.
  • Apologies: 76.
  • Conventional-closing: 2486 (about 1%, the tenth most frequent tag).
• Completions and summaries come in factual and emotional variants:
  • A factual example is the nurse supplying a drug name the patient is struggling to recall.
  • An emotional example is the nurse voicing the feeling the patient is struggling to express.
  • Repetition and explicit questions are used alongside them to confirm understanding.
• The authors call their findings impressionistic. Quantified verification awaits annotation of a larger sample.

Interesting ideas

• I1. A tag that is too rare to model in one corpus can be central in another, so collapsing rare categories based on a single reference corpus discards domain-relevant distinctions.
• I2. The same surface move can mean opposite things depending on the setting. Finishing another speaker's sentence can be a hostile interruption or a supportive act, and it can target facts or feelings.
• I3. Echoing what the other speaker said, as in Air Traffic Control read-backs, signals that information was received and agreed rather than adding anything new.
• I4. Some acts are hard to spot directly. Apologies can be implicit (pitying the listener for having to hear one's troubles), and the boundary between thanks and appreciation is fuzzy.
• I5. Reformulations made by participants are natural meaning-preserving paraphrases of earlier content, and closing sequences gather them into a recap.

For Mova

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

• I1 (frames): keep the speech-act inventory in prag fine-grained, including completion, reformulation, thanks and apology, even if a general training corpus barely contains them.
• I2 (frames): label a completion with its function (supportive vs. pre-emptive) and its target (factual vs. emotional), using context rather than form alone.
• I3 (world): treat a repeated or reformulated proposition as confirmation that a fact is shared, updating the participants' common ground instead of creating a duplicate event or state.
• I4 (frames): add implicature rules that map sympathy-for-the-listener expressions onto an apology frame, and model thanks and appreciation as a graded family.
• I5 (back): use in-dialogue summary/reformulation pairs as reference material for testing whether the text → graph → text compressor preserves meaning.