📄 Book Reviews: Presumptive Meanings: The theory of generalized conversational implicature

Summary

This item is a review by Nancy Green, published in Computational Linguistics in 2001, of Stephen C. Levinson's book Presumptive Meanings (MIT Press, 2000). The book develops a theory of generalized conversational implicature (GCI). GCIs are default, cancellable inferences that a hearer draws from the form and words a speaker chooses. Levinson argues that explaining them requires a third layer of meaning. This layer, utterance-type meaning, sits between sentence meaning and speaker meaning.

He sorts GCIs under three heuristics, Q, I and M, which descend from Grice's maxims. He illustrates them with a blocks-world example and ranks them to resolve conflicts. He also rejects the strict pipeline in which semantics finishes before pragmatics begins. In his model, default pragmatic inference helps fix the proposition expressed, and so can affect truth conditions.

Green values the book's rich description and its relevance to debates about linguistic architecture. She notes, however, that it offers no formal or computational model. She also finds its coverage of computational work out of date.

Key points

• The book is Levinson (2000), MIT Press, xxxiii+480 pp, $35. The review appeared in Computational Linguistics vol. 27, no. 3.
• The three heuristics are:
  • Q: anything left unsaid is assumed not to hold. For example, from "blue pyramid on the red cube", infer that there is no cone there.
  • I: a plain description is read stereotypically. For example, the pyramid rests directly on the cube.
  • M: an unusual or marked wording signals an unusual situation. For example, "cuboid block… supported by" suggests the block is not quite a cube and the support is not direct.
• The theory refines Gazdar's (1979) projection mechanism with a priority order: Q-inferences beat M-inferences, and M-inferences beat I-inferences.
• The model has a two-stage pragmatics:
  • "Gricean pragmatics I" enriches underspecified semantics before model-theoretic interpretation. It handles disambiguation, narrowing of word sense and reference.
  • "Gricean pragmatics II" then derives particularized implicatures and yields speaker meaning.
• Green points to several computational studies the book overlooks, among them the 1996 work of Di Eugenio and Webber, the 1997 paper by Elhadad, McKeown, and Robin, and Stone and Webber's 1998 contribution.

Interesting ideas

• I1. Default meanings tied to an expression type get their own layer, separate from both literal content and speaker intent, so they can be withdrawn without disturbing either.
• I2. The Q heuristic depends on a salient set of contrasting alternatives: mentioning one member implies the others do not apply.
• I3. The I heuristic fills in the stereotypical reading of simply worded statements, such as treating "on" as direct support.
• I4. The M heuristic treats an unusually wordy or rare phrasing as a cue that the situation is non-stereotypical.
• I5. A fixed precedence among inference types (Q over M over I) resolves conflicts between defaults deterministically.

For Mova

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

• I1 (frames): crate prag could hold GCI-style defaults in a distinct, defeasible annotation layer attached to frames, each with an explanation of which heuristic produced it.
• I2 (world): the world of the text could keep explicit contrast sets (colours, shapes, scalar terms) so the system can derive negative facts and explain them as "not mentioned among the alternatives."
• I3 (world): compiled knowledge could carry stereotype defaults per relation or noun, applied to entities and states unless the text contradicts them.
• I4 (frames): prag could compare a phrase against its plain competitor (e.g. "supported by" vs "on") and block the I-default when the marked form is used.
• I5 (frames): implicature conflicts could be settled by this explicit ranking rather than by scores, keeping each cancellation traceable.