📄 Exploiting Conversational Implicature for Generating Concise Explanations

Summary

The paper tackles verbosity in machine-generated explanations. A content planner produces a complete set of propositions (rules and facts) justifying an expert-system decision, and Horacek asks how far that set can be pruned while the listener can still recover the full message without drawing false conclusions. The domain is OFFICE-PLAN, a system that assigns employees to office rooms under constraints. Knowledge is encoded in a sorted type hierarchy plus IRS-style logical rules. Three "contextually motivated" rules then formalise parts of Grice's relevance maxim, using meta-predicates over the dialogue state such as topic, newinfo and not-false. An anticipation feedback loop tries candidate subsets from smallest to largest. It simulates the hearer's inferences and accepts the first subset that is complete with respect to the original content and produces no contradiction, meaning no false implicature. A worked example with three user stereotypes shows the same underlying explanation reduced to different surface contents. The work matters because it treats conciseness as reasoning about what the hearer will infer, not as a style filter.

Key points

• The application is OFFICE-PLAN. The example question asks for the reason person A was given room B instead of room C, and the full explanation has five propositions, including the rule that group leaders must be in single rooms.
• There are three C-Rules:
  • C-Rule 1 infers a contrasting fact about another topic entity of the same top class that the explanation leaves unmentioned.
  • C-Rule 2 infers that a questioned entity belongs to the class in a stated rule's premise.
  • C-Rule 3 infers that a rule the user already knows is relevant from a stated fact.
• A candidate subset is valid if every original proposition is either in the subset, in the user model, or inferred, and no contradiction arises. Rules the user does not know are always included in every hypothesis.
• There are three user stereotypes:
  • The local employee needs only the rule.
  • The novice also gets the rule, from which the rest is inferred.
  • The expert gets only the fact "A is a group leader".
• The system was implemented in CommonLisp on a SUN4. The paper gives no quantitative evaluation, only the worked example.

Interesting ideas

• I1. Generate-and-test conciseness: the system enumerates subsets of the content from the smallest up, simulates the hearer's inferences on each, and keeps the first one that recovers everything without contradiction. This gives a principled stopping criterion for compression.
• I2. Implicature as rules over dialogue-state meta-predicates: the paper's rules refer to which entities are topics of the question and which facts are new information. Contrastive implicature then falls out, as in "B is single" leading to "so C is not".
• I3. Checking that an explanation is free of false implicatures: a shortened output is rejected if the hearer's likely inferences contradict the full content, not merely if information is missing.
• I4. User-model stereotypes as known-fact and known-rule sets: the same content is pruned differently depending on what the addressee is assumed to know, with presupposed facts from the question treated as known to all.
• I5. A brevity preference for triggering rather than stating: if the hearer knows a rule, uttering the triggering fact is preferred over restating the rule.

For Mova

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

• I1 (back): The text → graph → text compressor could drop graph facts only when re-reading the shortened text through Mova's own pipeline recovers them, using the reader as the hearer model.
• I2 (frames): The prag crate could encode scalar or contrastive implicature as rules that check which question topics were addressed and which were left unmentioned, deriving explainable negated facts.
• I3 (answer): Short answers to fable or test questions could be checked for unintended implications before being emitted, which keeps answers terse but safe.
• I4 (back): Generation could take a reader profile, such as a pupil versus a teacher, and omit facts and rules that the profile marks as known or as presupposed by the question.
• I5 (answer): Explanations of maths word-problem solutions could cite the specific quantity fact and let a known rule, such as how a measure converts, stay implicit for advanced users.