📄 Speech Acts as a Basis for Understanding Dialogue Coherence

Summary

Perrault, Allen and Cohen ask what makes a dialogue hang together. Their answer is that utterances, like physical actions, are steps in plans. A dialogue reads as coherent when a listener can reconstruct the goal those steps serve. Each agent keeps a model of the world, including its beliefs about the other agent's beliefs and wants. Speech acts are planning operators with preconditions and effects, and those effects land mainly on these mutual models. The authors restrict themselves to two acts, REQUEST and INFORM, and treat questions as a request to be informed. They also limit the setting to cooperative, task-oriented dialogues, mostly with a train-station information clerk. They state requirements for representing belief, want, "knowing whether" and mutual belief. The program OSCAR plans speech acts and is used to test the act definitions. The paper also sketches how recognising the speaker's plan explains:

• indirect requests,
• elliptical replies and acknowledgements,
• helpful answers that address obstacles the asker has not mentioned,
• the choice of referring expressions.

Key points

• Only REQUEST and INFORM are modelled, as prototypes of Searle's directive and representative classes. A question is a REQUEST that the hearer INFORM.
• Models are built on belief (following Hintikka), not on knowledge as true belief, so that disagreement can be represented. "OTHER knows whether P" is a disjunction of two beliefs, and the authors call this essential for planning yes/no questions.
• Different situations need different depths of belief nesting:
  • Disagreement needs two levels.
  • Lying needs at least three.
  • Mutual belief is unbounded in principle, but Cohen (1978) stores it finitely with a circular structure, usually in three levels.
• OSCAR is meant to predict all and only the speech acts a person would appropriately produce in a given situation.
• Successful definite reference depends on the hearer's model of the speaker, not on the facts. Their adaptation of Donnellan's martini example shows this: a woman whose glass Jones believes holds water can still be picked out as "the woman holding the martini".

Interesting ideas

• I1. Give speech acts preconditions and effects on the participants' belief models, so that explaining an utterance becomes a matter of recognising the plan behind it, not matching surface forms.
• I2. Represent "knows whether P" as a disjunction of two belief states, so the system can record that someone has an answer without itself knowing it.
• I3. Treat a cooperative answer as finding obstacles in the plan the asker probably holds. Their example is replying "It's closed" when asked where the nearest service station is.
• I4. Choose a referring expression by checking that it picks out the intended object in the hearer's model, not in the speaker's own picture of reality.
• I5. Nest beliefs to a fixed, purpose-driven depth: two levels for disagreement, three for deception.

For Mova

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

• I1 (frames): crate prag could store each detected act as an operator with explicit preconditions and effects, which yields a traceable reason for every act label.
• I2 (world): the world of the text could hold such disjunctive beliefs for characters, which matters for tales where one character knows a secret the reader has not yet learned.
• I3 (answer): when answering QA on fables, Mova could infer the asker's goal and check whether it is blocked, then report the blocker instead of only the literal fact.
• I4 (back): generation could test each candidate description against the reader's assumed knowledge state, and choose, for example, between "the drawer with the hammer" and a positional description.
• I5 (world): this gives a principled cap on nesting when tracking who believes what about whom, enough to model tricksters in fables without unbounded recursion.