📄 Using Structural Constraints for Speech Act Interpretation

Summary

Allen and Hinkelman challenge the literal meaning hypothesis. That hypothesis holds that sentence structure fixes one context-free "literal" speech act, and that any indirect meaning is reached later by plan reasoning from that starting point. The authors argue this approach cannot capture phrasing subtleties. For example, Can you pass the salt works as a request, while a paraphrase that explicitly asks the hearer to report on their ability to pass the salt is only a question about ability. The approach also cannot absorb cues like intonation, and it wastes effort re-deriving common indirect forms from first principles. In their alternative, structural interpretation rules match a combined syntactic structure and logical form. The rules output an interpretation space with two parts: limits on what the intended act may be, and specific conventional readings it is likely to be. A plan reasoner then filters these candidates against act constraints, preconditions and effects. Only when ambiguity remains does it fall back on general plan recognition based on Kautz's algorithm. A prototype built in COMMONLISP within RHET runs all the paper's examples. The authors claim it is faster on conventional forms, more accurate, and open to prosodic input. They report no quantitative evaluation.

Key points

• The paper targets the literal meaning hypothesis used by Searle and by earlier Allen & Perrault work, where syntactic mood mapped to a surface act:
  • imperative → surface request,
  • declarative → surface inform,
  • interrogative → request to inform.
• It relies on Searle's five act classes: Representatives, Directives, Commissives, Expressives, and Declaratives.
• Older systems compiled shortcuts such as REQUEST decomposing into SURFACE-REQUEST(INFORMIF(CANDO)). These shortcuts still began from literal meaning, so they could not tell paraphrases apart.
• A sample "can you" rule fires on a yes/no question with auxiliary can, subject you and an action verb. It yields two items: {REQUEST of that action, generic SPEECHACT}. If filtering leaves one specific reading, the system accepts it.
• The implementation uses COMMONLISP and the RHET environment, with plan recognition following Kautz. The prototype covers only a limited knowledge base of acts and rules.

Interesting ideas

• I1. Have the structural stage emit a bounded set of candidate readings plus a coarse class constraint, rather than one committed meaning, and let context prune that set. This keeps cheap structure-driven guesses while leaving room for reasoning.
• I2. Encode conventional indirect forms as pattern rules over parse plus logical form, keyed on specific features like auxiliary, subject person, mood and verb type. This makes "Can you X?" and "Tell me whether you can X" diverge naturally.
• I3. Filter candidate acts by testing their constraints, preconditions and effects against context, and run full plan inference only when filtering leaves ambiguity. This gives a fast path for common cases and an expensive path for hard ones.
• I4. Treat every cue source, including punctuation and (in speech) intonation, as just another rule family that restricts the same interpretation space. New evidence then plugs in without redesign.
• I5. The authors suggest attaching frequency weights to interpretation rules but warn that the weights are hard to set, since "can you" sentences have many genuine yes/no uses (e.g. eating three pizzas in ten minutes).

For Mova

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

• I1 (frames): prag could output a speech-act space per clause (e.g. {request(lift rock), any-act}) derived from the UD tree, then narrow it with world-of-text checks.
• I2 (frames): match UD features (aux lemma can, nsubj you, Mood=Int) to propose request frames, keeping each rule as an explainable trace.
• I3 (world): check whether the addressee is the agent and whether the action is feasible in the text's world before keeping a directive reading.
• I4 (frames): a final "?" on a declarative ("You are going to the party?") could add a question candidate through the same mechanism.
• I5 (learn): estimate rule weights from silver data and test them on negative controls such as genuine ability questions.