📄 Embedded Implicatures? Remarks on the debate between globalist and localist theories

Summary

The paper asks whether localist accounts of scalar implicature can survive the experiments of Geurts & Pouscoulous (2009). Those experiments found that embedded implicatures are rarely drawn. Localist accounts compute implicatures inside embedded clauses with a silent exhaustivity operator O. Ippolito tests the strongest version of this view: Chierchia, Fox & Spector (2008), extended with a "strongest meaning hypothesis" in two forms, SMH1 and SMH2, that predicts which placement of O is preferred. She works through attitude predicates (believe, want, be certain, wish) and pays special attention to Neg-raising verbs. Both SMH versions predict preferences that participants did not show. The clearest failure is that the theory cannot separate believe from want, although both are Neg-raising and behaved very differently in the data. The usual escape route is to say that context overrides the prediction. That route is closed here, because the experimental items had no context. Globalist theories explain apparent embedded readings as a global implicature strengthened by assumptions in the common ground, so context-free tasks should rarely yield them. The paper argues that plausibility differences among these assumptions may explain the believe/want contrast, and it calls for further experiments.

Key points

• Geurts & Pouscoulous used two methods. In the inference task, people judge whether the version with the stronger scalar term is false. In the verification task, people match a sentence against a picture. The inference task gave higher implicature rates for plain, unembedded sentences, so the verification task was treated as the more reliable measure.
• The verification task tested "some" in downward-entailing, upward-entailing and non-monotonic environments. It found none of the local implicatures that localists predict.
• SMH1 compares all logical forms that differ only in where O sits. SMH2 compares each O-containing form only with the same form minus that O.
• For be certain, SMH1 predicts that the embedded reading should win, which clashes with the experimental pattern.
• The believe condition showed a non-negligible acceptance rate for the embedded reading, though it was still not the majority choice. The want condition showed a very low rate. This difference also weakens Gajewski & Sharvit's objection that Russell's globalist account succeeds only because believe is Neg-raising.

Interesting ideas

• I1. An apparent embedded reading can come from a weak global inference plus an extra assumption about the speaker or attitude holder, such as Sauerland's "epistemic step" (if John is assumed to know either way, "not believing all" becomes "believing not all"); this keeps the strong reading derivable and traceable.
• I2. The way you ask about an inference changes how often people report it: the inference-style question inflated rates compared with picture verification.
• I3. Ranking candidate readings by logical strength is a clean, explainable selection rule, and it can be tested systematically against judgments.
• I4. Neg-raising is a lexical property of predicates, but it does not on its own predict implicature behaviour; the plausibility of the disjunctive assumption ("either expects all or expects not all") also matters, and it differs between predicates such as expect and ought.
• I5. A theory that has to invoke context to fix its predictions is suspect when the tests themselves are context-free; context-free items isolate the default behaviour of the mechanism.

For Mova

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

• I1 (frames): crate prag could first compute the weak global implicature, then strengthen it only when an explicit competence or opinionatedness assumption is recorded, and log that assumption as the justification.
• I2 (learn): implicature evaluation should not rely only on "does X follow?" prompts to an LLM teacher; verification-style items checked against a constructed world state would give a less biased gold signal.
• I3 (frames): an SMH-like strength ordering could serve as a default disambiguation heuristic over candidate implicature readings, with known failure cases (e.g. want vs believe) kept as negative controls.
• I4 (world): the verb-class knowledge layer could mark Neg-raising predicates and attach a prior plausibility for the "opinionated holder" assumption, which prag then consults.
• I5 (answer): out-of-context test items for pragmatic readers can show whether a pragmatic rule fires by default or only when discourse forces it, as in "salad or dessert… but if you take both".