📄 On Trying to Do Things with Words Another plan-based approach to speech act interpretation

Summary

Classic plan-based models of speech acts treat an utterance as a planning operator whose effects follow once the act is performed. Such models therefore cannot describe an assertion that convinces nobody or a request that goes unheeded. Perrault's fix makes the step from performing an act to its effects defeasible. The authors argue this has two costs. Speech acts lose any real definition. A speaker with no prior opinion can also talk herself into believing whatever she says.

Their alternative inserts an explicit "trying" layer between intention and action. A default rule moves from trying to do A to doing A. A strict rule then links doing A to its preconditions and effects. When those consequences contradict existing beliefs, the default is withdrawn. This keeps strict operator definitions, such as a fully specified Assert, while still covering failed acts. Speaker, hearer and overhearer all interpret the utterance through the same procedure.

The paper also separates "intending to try" from "intending to do". This resolves puzzles in Bratman's theory of intention. The approach was prototyped as MEDIAS.

Key points

• Actions are encoded as quintuples:
  • action description
  • preconditions
  • add-list
  • delete-list
  • body
• These components are read through conditional generation (after Goldman and Pollack): the body actions generate the action only when the preconditions hold.
• Perrault's theory has two parts:
  • axioms for memory, persistence and observability;
  • defaults for a declarative rule and for belief transfer.
• The authors' critique of Perrault: it leaves assertion undefined, and it lets an agent convince herself of P just by uttering it. Appelt and Konolige patched the second problem with a heavier nonmonotonic logic.
• Assert P is defined strictly:
  • precondition: the hearer takes the speaker to be sincere and competent;
  • add-list: the hearer believes the speaker believes P;
  • body: uttering "P.".
• Insincerity and incompetence are judged by separate rules over temporally related situations. Belief transfer is restricted so agents do not automatically adopt what others think they believe.
• MEDIAS was built with the expert system shell HARES. It covers assertions and yes/no questions, and tells information-seeking questions apart from information-probing ones.

Interesting ideas

• I1. Put the default on the step from attempt to success, not on the step from success to effects, so the act's definition stays strict and only its occurrence is defeasible.
• I2. Run one interpretation procedure from every viewpoint (speaker, hearer, overhearer), each with its own belief store, so a lie yields different but consistent belief states for each participant.
• I3. Treat the speaker's sincerity and competence as derived judgements from explicit belief-timing rules, rather than as fixed flags.
• I4. Treat "trying A" as an action in its own right, whose only guaranteed result is the attempt, so an agent can rationally aim to try what it expects to fail.
• I5. Block self-referential belief loops, such as believing P because you think others think you believe P, as an explicit inference guard.

For Mova

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

• I1 (frames): A speech-act frame in prag could always record an "attempted assert/request" and mark success as a separate retractable assumption, with a traceable reason when it is defeated.
• I2 (world): The text world could keep per-character belief layers updated by the same rules, which would help answer fable questions like "who believed the fox?".
• I3 (frames): Irony and deception detection in prag could flag an assertion as insincere when the narrator's world state shows the speaker believes the opposite.
• I4 (world): Event ordering could store tried-but-failed events (e.g., "tried to lift the log") as real events with unrealised effects, so that states are not wrongly updated.
• I5 (answer): QA readers chaining nested beliefs could apply such guards to avoid circular conclusions, and expose the blocked step in explanations.