📄 Performatives in a Rationally Based Speech Act Theory

Summary

The paper asks why a first-person present-tense utterance such as "I request you to open the door" counts as a request just by being said, while "I lie that p" defeats itself and "frighten" cannot be used this way at all. Cohen and Levesque answer with their rationally based theory of speech acts, which rests on a modal logic covering what agents believe and aim for, the actions they perform, and how these unfold over time (BEL, GOAL, HAPPENS, DONE, plus dynamic-logic action expressions). Performatives are treated as ordinary declarative utterances, not first as assertions. Each sentence mood is tied to a "core attitude" of the speaker. Illocutionary acts are defined as attempts that depend on the speaker's own beliefs and goals, which the speaker can introspect. A sincere performative request is therefore guaranteed to be a request. A performative lie is provably impossible. The analysis answers Searle's 1989 claim that performative-as-declarative accounts must fail. It also extends Bach and Harnish without needing a separate class of "declarations." The authors argue this matters for dialogue systems, database QA, translation of force markers, and agent communication languages.

Key points

• The formal base is a first-order modal logic:
  • BEL has weak S5 semantics and GOAL has system K semantics.
  • Assumptions include that beliefs imply choice, that goals are known (introspection), and that agents remember their past beliefs.
• Commitment is modelled in layers:
  • A persistent goal (P-GOAL) is kept until the agent believes it is achieved, believes it is impossible, or believes an escape condition q no longer holds.
  • Intention to act (INTEND1) and intention to achieve a proposition (INTEND2) are special cases of P-GOAL. They rule out achieving something accidentally or unknowingly.
• Utterance moods carry default attitudes:
  • A declarative signals that the speaker believes the sentence is true.
  • An imperative signals that the speaker wants the addressee to do an act fulfilling the sentence's conditions.
  • Observers adopt these attitudes unless they hold contrary beliefs, and this reasoning can be nested over several levels.
• The paper states two desiderata and meets both:
  • A sincere performative request is both a request and an assertion, and neither is derived from the other.
  • Theorem 3 shows that saying "I lie that …" can never constitute a lie.
• Verbs fit the analysis differently:
  • Verbs naming pure attempts work performatively.
  • Verbs that need an outcome to hold after the utterance cannot be used performatively. Examples are perlocutionary verbs like "frighten" and acts like lying.
  • Institutional performatives, such as a judge pronouncing a couple married, are handled by domain axioms that reflect the institution rather than rationality.

Interesting ideas

• I1. Define a speech act as an attempt fixed by the speaker's beliefs and goals, not by its outcome; this cleanly separates acts that guarantee themselves from acts that need uptake.
• I2. Attach a default "core attitude" to grammatical mood (declarative → speaker believes it, imperative → speaker wants the addressee to act), and let observers cancel it when they hold contrary beliefs.
• I3. Model commitment as a persistent goal with explicit drop conditions: achieved, impossible, or the escape reason has become false.
• I4. Derive the request reading and the assertion reading separately from the utterance event, instead of deriving one force from the other as an indirect act.
• I5. Express institutional effects (marrying, sentencing) as separate domain axioms keyed on speaker role and circumstances, kept apart from the general laws of rational action.

For Mova

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

• I1 (frames): crate prag could mark each speech-act frame as either "attempt-only" or "requires outcome," and accept explicit performatives only for the attempt-only class.
• I2 (frames): mood read from the UD tree could set a defeasible speaker-attitude slot, which irony or lie detection then overrides explicitly rather than guessing from words.
• I3 (world): goals of story characters could carry these drop conditions, so the world model can explain why a character abandoned or kept pursuing an aim.
• I4 (frames): one utterance node could yield several parallel act frames, each with its own explainable derivation chain and preconditions.
• I5 (world): the compiled knowledge layer could hold role-gated rules that turn a performative utterance into a new state, such as "married," in the text world.