📄 A Practical Nonmonotonic Theory for Reasoning about Speech Acts

Summary

The paper asks how to formally describe what changes in the beliefs of a speaker and a hearer after an utterance, since the outcome depends heavily on context. Earlier monotonic accounts either posit a context-free core effect (which counterexamples break) or need an explicit record of what agents do not believe. Perrault's default-logic theory, built on Reiter's normal defaults, handles the conventional side of speech acts but has three flaws. It lets a speaker persuade himself of an unsupported claim just by asserting it. It assumes beliefs always persist, which blocks integration with action and belief revision. Its belief-transfer rule cannot be run efficiently. Appelt and Konolige reformulate the theory in hierarchic autoepistemic logic, a layered refinement of Moore's autoepistemic logic. In it, ordered subtheories give some defaults priority over others, so competing outcomes (the hearer is convinced, resists, or suspends judgment) can each be selected by context. The logic has efficient proof methods. The theory was implemented and used inside the GENESYS utterance-planning system.

Key points

• The target is Perrault's theory, which combines weak S5 belief operators with these rules: Persistence, Memory, Observability, Belief Transfer, and a Declarative default.
• Counterexample to Perrault: a speaker who lacks belief in P but utters it will, under axioms (1)–(5), end up believing P.
• Why simple fixes fail: rules saying "ignorance persists" would stop anyone from ever learning from speech. Switching to non-normal defaults loses guaranteed extensions and any proof procedure.
• The new formalism: hierarchic autoepistemic logic uses a hierarchy of subtheories. Lower levels feed higher ones through an inheritance condition, which allows the speaker's own belief default to be blocked while the hearer's default still fires.
• Further scope:
  • Hearer belief revision is modelled by choosing which subtheory initial beliefs are copied into.
  • Nested mutual reflection is captured by a schema [(s,h)n] that alternates the two agents' beliefs.
  • Imperatives and directives are only outlined.
  • The theory is implemented in GENESYS.

Interesting ideas

• I1. Express preferences between conflicting defaults by placing them on ordered layers, rather than adding ad hoc exception clauses to every default; this makes priority an explicit, inspectable design choice.
• I2. Treat the effect of an utterance as defeasible, and keep the speaker's and hearer's updates as separate defaults, so that one can be blocked without the other (the liar convinces the listener but not himself).
• I3. Model belief revision as a choice of where prior beliefs are placed relative to new evidence: strong priors outrank the utterance, while weak priors sit below it and get overridden.
• I4. Handle nested beliefs ("S thinks H thinks S believes…") with one uniform schema applied level by level, instead of writing separate rules for each depth.
• I5. Choose a logic partly for its computational tractability, because a theory that cannot be run efficiently is of limited practical use.

For Mova

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

• I1 (frames): speech-act and implicature rules in prag could be stacked in priority tiers, so a stronger cue (e.g., a known lie) cleanly overrides the default "assertion ⇒ belief" and the trace shows which tier won.
• I2 (world): store per-agent belief states for story characters, where "X said P" adds P to the listener's beliefs by default but not automatically to the speaker's, which matters for fables with deception.
• I3 (world): attach a strength level to entity states, so a new statement in the text either revises a prior fact or is recorded as disputed, with an explicit reason in either case.
• I4 (answer): QA about what a character thinks another character knows could reuse one recursive rule, keeping false-belief questions in tales explainable.
• I5 (frames): when designing pragmatic inference, prefer rule formats with bounded, directed application over open-ended forward chaining over all beliefs about beliefs.