📄 Global Approach to Scalar Implicatures in Dynamic Semantics

Summary

The paper addresses a long-running debate in pragmatics: are scalar implicatures (SIs) computed over a whole utterance (globalism, the Neo-Gricean view) or over embedded syntactic constituents (localism, as in Chierchia 2002 with a negation operator and Fox 2006 with an exhaustivity operator, exh)? Yeom argues that SIs are always global. Cases that look local are effects of context and background knowledge acting on a global computation.

The argument uses minimal-pair examples in which syntactically parallel sentences receive different implicatures. It also uses examples whose plausible SI corresponds to no clause-type constituent at all. To model this, the author recasts sentence meaning as a DRS that includes a possible-world variable, interpreted against information states (sets of world–assignment pairs) through support conditions. An SI is computed after interpretation by negating a stronger alternative DRS. Inside that alternative, individual conditions may be dropped, and its variables may be bound to referents in the main DRS, much as presuppositions project. This matters because it explains apparently local readings without syntactic operators. On this account, SI calculation depends on content and likelihood rather than tree structure.

Key points

• Sauerland's (2004) two-step procedure is the Neo-Gricean baseline. The speaker first yields uncertainty inferences (primary implicatures, ¬K). These are then strengthened to K¬ (secondary implicatures) if consistent with the statement and the other primaries.
• The key contrast pair is about drinking beer or wine and being allowed to drive versus watching TV or playing games and failing math. In the first, the global SI is natural: no one who did both was allowed to drive. In the second, it is not, because doing both makes failing more likely.
• For the example stating that certain boys, each having gone through a portion of the books, passed the test, the plausible SIs are that no boy read every book or that the boys who passed had not read them all. The second refers to an entity with no matching syntactic constituent, which the author treats as evidence against localist analyses.
• The formal machinery has four parts:
  • DRS = ⟨var, con⟩, with a world variable v.
  • A model ⟨W, D, F, G⟩.
  • Support conditions for predicates, disjunction and conditionals, where a conditional is treated like ¬antecedent ∨ consequent.
  • Projection of alternative variables, modeled on van der Sandt (1992) presupposition projection, with sentence content treated as presupposed in the spirit of Stalnaker (1978, 2006).
• Three factors are named as deciding the surviving SI: the semantic strength of alternatives, the information already present in the information state, and background knowledge that sets a scale of likelihood.

Interesting ideas

• I1. Compute implicatures only after the full meaning representation is built, comparing whole alternative representations instead of attaching operators to syntax nodes.
• I2. Filter a candidate implicature by checking whether its strengthening runs with or against a likelihood scale from background knowledge, treating the disjunct as an upper or a lower limit for the outcome.
• I3. When building the negated stronger alternative, allow individual conditions to be dropped, so the result can be a contextually relevant SI rather than the literal maximal one.
• I4. Bind the variables of the alternative representation to discourse referents already in the main representation, the way presuppositions project, so an SI can talk about "the boys who passed".
• I5. Use minimal pairs with identical structure but opposite real-world likelihood as a diagnostic for whether a pragmatic module relies on structure or content.

For Mova

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

• I1 (frames): crate prag could generate scalar alternatives over the finished frame graph and compare their entailment strength, keeping the UD tree out of the implicature logic.
• I2 (world): the compiled knowledge layer could supply rough "more X makes Y more/less likely" directions, so a candidate SI that conflicts with them is rejected and the rejection is logged as an explanation.
• I3 (frames): prag could enumerate condition-subsets of the alternative frame and pick the strongest one consistent with the world state, recording which conditions were dropped and why.
• I4 (world): implicature frames could reuse existing entity IDs from the text world via the coreference machinery instead of introducing fresh existentials.
• I5 (learn): such pairs could serve as negative controls in implicature evaluation, flagging any rule that gives both sentences the same SI.