📄 Speech Acts of Assertion in Cooperative Informational Dialogue

Summary

The paper asks how a dialogue system can take part in cooperative, information-sharing conversation, where both sides aim to grow a pool of mutually agreed knowledge. Kononenko treats speech acts of assertion (SAAs) as the main tool for this and builds a formal, purely theoretical model. There is no corpus and no experiment. Each participant's knowledge base has two parts: their own beliefs about the world, and their beliefs about the partner's knowledge. Every proposition carries a graded verisimilitude score (v-evaluation) between 0 and 1. Rules describe how information moves between these parts and when it becomes shared knowledge M. The speaker's view of what the listener lacks falls into one of four ignorance types. These types fix the communicative goal, the change the speaker wants to make in the listener's mind, and so the type of assertion. The result is a typology of basic SAAs (Statements versus Evaluations, with subtypes) and non-elementary SAAs. It comes with a scheme of dialogue flow showing their different roles. The work matters as an early cognitive account linking assertion types to belief updates.

Key points

• Knowledge is split into the speaker's own state X and a nested model X:Z, meaning what X thinks Z knows. Shared knowledge M holds propositions present in X, Z, X:Z and Z:X. A stronger version also requires the v-values to match.
• Each proposition gets a v-evaluation on the scale 0 ≤ v ≤ 1. Ranges near 1 or 0 mean certain, ranges near 0.5 mean fully uncertain, and values in between mean leaning one way. A meta-value "?" marks a value believed to be unknown to the partner.
• Four ignorance situations (A–D) correspond to four operations on the listener's knowledge:
  • A: the proposition is not active for Z; the operation is inclusion.
  • B: Z has the proposition but lacks X's evaluation; the operation is exclusion.
  • C: Z faces a disjunction of alternatives; the operation is selection.
  • D: Z holds a wrong alternative; the operation is correction.
• There are two basic SAA types:
  • Statements (S-SAA) introduce a new proposition plus its evaluation. Subtypes are Certain, Uncertain and Indefinite.
  • Evaluations (E-SAA) add only an evaluation to a topic already open. Subtypes include Acceptance, Approval, Confirmation, Denial and Doubt.
• In the dialogue scheme, a topic opens with a Statement and closes once both values agree and enter M. If neither side changes its value after one exchange, the dialogue moves into argumentation.

Interesting ideas

• I1. Each assertion is packaged into three sections: what is already shared (M), what the speaker thinks the hearer believes, and the speaker's own view. This makes the act's preconditions explicit and checkable.
• I2. The difference between Statements and Evaluations rests on whether new propositional content arrives or only a new certainty value on an existing topic.
• I3. Graded certainty (0–1, plus an "unknown to partner" meta-value) replaces true/false/unknown, with coarse intervals mapped to named states.
• I4. Nested belief models (X, X:Z, X:Z:X) are paired with explicit transfer rules, and a fact counts as mutual only when all levels agree.
• I5. A dialogue goal is framed as one of four edit operations on the listener's knowledge: include, exclude, select among alternatives, or correct.

For Mova

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

• I1 (frames): prag could tag each assertive clause with these three slots, so a speech-act label always comes with a traceable belief-state justification.
• I2 (frames): this gives a clean test for deciding whether a reply ("Really?", "I doubt it") opens a new frame or only re-scores an existing one.
• I3 (world): world-of-text propositions could carry a per-character certainty interval, letting hedges like "it seems" or "I don't think" update a character's belief strength explainably.
• I4 (answer): fable and tale QA could use this to answer "does the fox know the crow knows…" questions by walking nested belief stores instead of guessing.
• I5 (back): generation could pick the surface form of an assertion (plain claim, denial, "it's X, not Y") from the required edit operation on the reader's assumed state.