📄 Speech Acts as a Basis for Understanding Dialogue Coherence

Summary

The paper asks what makes a dialogue hang together. The authors argue that utterances, like physical actions, seem coherent when an observer can link them to a shared overall goal. They propose treating speech acts as planning operators. Each operator has preconditions and effects, and those effects fall mainly on the models that speaker and hearer keep of each other's beliefs and goals. Understanding an utterance then means recognizing the plan behind it. The authors also assume that a speaker intends any inference the hearer can draw from shared knowledge.

The theory covers only two acts, REQUEST and INFORM, and it is limited to task-oriented, two-party cooperative dialogues. Most examples come from an information clerk at a train station. The paper describes:

• belief models of other agents,
• planning operators for asserting, commanding and asking,
• how illocutionary force is identified, including indirect speech acts,
• how referring expressions are planned against the hearer's beliefs.

This gives an early, explicit computational account of cooperative answers, indirect requests and elliptical replies, all grounded in beliefs and plans.

Key points

• Speech acts are modelled as planning operators with preconditions and effects, in the operator tradition of STRIPS (Fikes & Nilsson). Plans can be built, simulated, executed and debugged.
• Agent models use "believe" in Hintikka's sense, not "know" as true belief. This lets SELF represent that OTHER disagrees or is ignorant. The paper contrasts this with Moore (1977), whose know-based models cannot capture such disagreements.
• "OTHER knows whether P" is represented as a disjunction of two beliefs: OTHER believes P, or OTHER believes not-P. The authors call this essential for planning yes/no questions.
• REQUEST and INFORM are taken as prototypes of Searle's directive and representative classes. Questions are represented as a REQUEST to INFORM.
• Planning a definite reference requires the speaker to believe the chosen description picks out the right object in the hearer's model of the speaker. Mutual belief extends this to deeper embedded models.

Interesting ideas

• I1. Cooperative answering works by detecting obstacles in the other person's inferred plan: a helpful reply to a request for the nearest service station is that it is closed. Inferring the goal behind a question lets a system add the warning a literal answer would omit.
• I2. Indirect speech acts such as "Can you pass the salt?" are recovered by plan inference from the literal act to the intended one. Indirect force is derived rather than listed, so it generalizes to new phrasings.
• I3. Knowing-whether is stored as a disjunction of possible beliefs, and three belief states are kept apart: believing a specific fact, believing some value exists, and knowing what the value is. This cleanly separates "knows that", "believes there is" and "knows which".
• I4. Identifying illocutionary force also gives a way to understand elliptical replies and utterances that acknowledge, correct or clarify earlier ones. Fragments get their reading from the act they continue.
• I5. The detail a referring expression needs depends on what the hearer is believed to know. "In the drawer with the hammer" works only if the hearer knows where the hammer is.

For Mova

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

• I1 (answer): The QA reader could infer why a question is asked and flag world-state facts that block the asker's evident goal, instead of only filling the asked slot.
• I2 (frames): crate prag could encode REQUEST/INFORM with preconditions and derive indirect requests by explicit inference steps. This keeps each reinterpretation explainable.
• I3 (world): Character belief states in fables could use these three distinct belief forms, so the system can answer who knows the gate versus who merely thinks there is one.
• I4 (frames): Short replies in dialogue texts could be attached to the pending speech-act frame rather than parsed in isolation, filling missing roles from the open REQUEST.
• I5 (back): Generation could check that each definite description resolves uniquely in a modelled reader's world, choosing anchors the reader already knows.