📄 A Tool for Extracting Conversational Implicatures

Summary

The paper addresses the gap between what a conversational turn literally says and what the speaker means: Grice's conversational implicatures. The authors use English Twitter dialogues from the Microsoft Research Conversation (MRC) corpus and build a comparable Arabic Twitter corpus. They then convert Lymba's logic-based textual-entailment prover COGEX into a non-monotonic, abductive engine.

Each turn becomes a deep semantic first-order logical form, with indexicals such as I and you bound to the actual participants of that turn. The prover draws on several kinds of knowledge:

• lexical-chain axioms mined from eXtended WordNet (Arabic WordNet for Arabic),
• rules that compose pairs of semantic relations,
• common-sense knowledge,
• Grice's maxims recast as default macro-axioms.

Results are stored in an eight-component implicature model, and the conversation's common ground is updated after every turn. On manually annotated English utterances, the system derives entailments and implicatures with an F-measure of 70.42%. The work shows that implied meaning can be represented explicitly and traced through inspectable inference steps.

Key points

• The implicature model is an 8-tuple {S_U, S_I, H_TM, H_TE, H_I, C, GM, K}. It covers the speaker's utterance and intentions, the hearer's understanding, entailments and implicatures, the context, the maxim status (none violated, clash, or flouted) and the needed common-sense knowledge.
• The data are MRC (about 1.3 million open-domain Twitter conversations) and a new Arabic corpus of 4,000 conversations among 2,067 users. The Arabic threads were assembled backwards from reply tweets.
• Gold annotation covers implicature-model components for 253 English and 75 Arabic utterances. The Relevance maxim is exploited most often, and Manner floutings are another major source.
• The system uses 86 hand-derived, language-independent semantic calculus axioms of the form R1 ∘ R2 → R0, each with accuracy measured on a large corpus. Lexical chain axioms are generated on demand.
• COGEX was changed in three ways: dialog-aware logical forms for English and Arabic, semantic axiom generation, and support for non-monotonic reasoning.

Interesting ideas

• I1. Grice's maxims are encoded as default rules that fire on utterance-type predicates; for example, Quality plus an exclamatory turn yields "speaker shows strong feeling." This turns vague pragmatic principles into explicit, traceable inference steps.
• I2. A small set of relation-composition rules lets the reasoner bridge concepts that have no direct link, such as part-whole transitivity or parts inheriting their whole's location, and each rule's reliability is measured empirically.
• I3. Before building the logical form, indexicals are resolved to the concrete speaker and hearer of each turn, and the common ground is updated incrementally after every turn.
• I4. Surface cues such as capitalized characters, utterance type and elongated interjections ("Eeeeeee") become explicit predicates in the semantic representation, so the reasoner can use them as evidence of attitude.
• I5. The authors argue for keeping several competing implicature readings alive and letting later turns disambiguate them, rather than committing early.

For Mova

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

• I1 (frames): crate prag could express its implicature and irony rules as defeasible maxim-triggered axioms keyed on speech-act type, so every derived implicature cites the maxim that produced it.
• I2 (world): the world builder could hold a compact, accuracy-scored table of role/relation compositions to infer holdings and locations, discarding rules whose measured precision is low.
• I3 (frames): coreference for dialogues and fable dialogues could bind I/you per turn before frame extraction, keeping a running common-ground store that later turns read from.
• I4 (tokens): the tokenizer and tagger could emit flags for all-caps, letter elongation and interjections as features that prag consumes for emotion and irony detection.
• I5 (answer): readers answering questions about dialogues could hold ranked alternative interpretations and retract cancelled implicatures when subsequent text contradicts them.