📄 Developing a Typology of Dialogue Acts: Some Boundary Problems

Summary

The authors, a team from the University of Tartu, describe a dialogue act scheme for tagging Estonian spoken conversation. Their longer-term goal is a computer that takes part in dialogue while respecting the conventions of human talk. They found existing schemes too coarse, for example because they treat all communication management or all acknowledgement as one function, so they built their categories from conversation analysis. Two principles organise the scheme. First, acts that open an adjacency pair, and so oblige the partner to respond, are separated from acts that do not. Second, dialogue management is split into acts that keep talk flowing and repair acts that resolve communication trouble. The result is eight groups containing 137 acts. Using double-annotated institutional dialogues (information calls and travel-bureau talk) from the Estonian spoken corpus, they examine hard boundary cases between questions and directives. Inter-annotator kappa ranges from 0.59 to 0.79, which shows the definitions still need sharpening. The work lays groundwork for automatic dialogue act recognition and dialogue system design.

Key points

• The eight groups are rituals, restructuring acts, contact control, repair, questions/answers, directives, opinions, and other (non-AP) acts, with 137 acts in total.
• Each group includes an "other" subtype so that every utterance can receive a tag.
• The sub-corpus has 114 texts, each tagged by two annotators and then merged. It contains 5815 tags, among them 308 directive first parts, 258 directive second parts, 633 questions and 1081 answers.
• Questions are grouped by the reaction they expect:
  • information: wh-questions and open yes/no questions;
  • agreement or refusal: closed yes/no questions and questions that offer an answer;
  • choice: alternative questions.
• Kappa is 0.59 for face-to-face travel-bureau dialogues and 0.79 for information calls. Planned next steps are frame-based act definitions and statistical act recognition.

Interesting ideas

• I1. The scheme's top-level split asks whether an act creates an obligation to respond (adjacency pair first part), rather than starting from forward- versus backward-looking acts; this ties each response to its trigger and exposes insertion sequences between them.
• I2. Repair is kept apart from information exchange even though both often look like questions, so surface form alone never fixes the act type.
• I3. Open and closed yes/no questions share identical form, so the authors tell them apart by their position in the dialogue (main-line series versus specifying follow-ups).
• I4. Refusing and lacking information are merged into one tag because both lead to the same continuation; categories are defined by their consequences for the dialogue rather than by speaker motive.
• I5. Agreement is reported separately per activity type, which shows how much harder face-to-face talk is to annotate than short service calls.

For Mova

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

• I1 (frames): Speech-act frames could carry a "pending reply" slot, so prag can link answers to their questions and flag rhetorical questions that open no obligation.
• I2 (frames): A question-shaped clause in dialogue could be classified as a clarification request or a real information query before frames are built, which avoids treating repair as new content.
• I3 (answer): When answering yes/no questions about tales or dialogues, the reader could check discourse position to decide whether to reply "yes/no" or supply the requested details.
• I4 (world): The world model could record answer outcomes by their effect, such as "query closed, unanswered", which keeps state updates simple and explainable.
• I5 (learn): Silver-data agreement between witness parsers could be broken down by genre, so weak domains are visible instead of being hidden inside one average.