😉 Implicature
Implicature is the part of what a speaker communicates that goes beyond the literal content of the words. Hearers infer it by assuming the speaker is cooperative, as when "some students passed" is taken to mean that not all did. Unlike entailment, conversational implicature can be cancelled.
What it is
An implicature is a meaning that a hearer infers from the fact that a speaker chose a particular utterance in a particular context. It is not something the sentence asserts.
Suppose someone asks "Did you go shopping?" and the reply is "My car's not running." The reply says nothing about shopping, yet it clearly conveys "no." Such indirect answers have been analysed as particularized conversational implicatures. They are recovered through a shared expectation that the reply bears on the question asked. An implicit discourse relation also links the stated fact to the unstated answer; here the relation is an obstacle (Green & Carberry 1992).
Scalar implicatures are the most studied type. Using a weaker term on a scale, such as some rather than all, invites the inference that the stronger term does not apply.
Two properties set implicatures apart from neighbouring phenomena:
- They are cancellable. "Some, in fact all, passed" is not a contradiction.
- They behave differently from presuppositions. "Jo's cat" takes for granted that Jo has a cat, and that assumption survives under negation, questions and conditionals. Scalar inferences do not survive in the same way. The IMPPRES diagnostic set was built to test whether models capture this contrast (Jeretic et al. 2020).
Pragmatics surveys and benchmarks list implicature next to presupposition, reference and deixis as separate phenomena (Sravanthi et al. 2024; Ma et al. 2025).
The key idea
The classical mechanism comes from Grice (1975). Hearers assume speakers follow a cooperative principle, which is spelled out as maxims of quantity, quality, relation and manner. Sometimes an utterance appears to violate or flout a maxim. The hearer then infers whatever extra meaning would restore cooperativity. Other implicatures, scalar ones among them, arise because the speaker is assumed to be observing a maxim, here Quantity. Had a stronger statement been true, a cooperative speaker would have made it.
A formal approach called Rational Speech Acts (RSA) makes this reasoning explicit and quantitative. It treats pragmatics as listeners and speakers recursively modelling one another (Cohn-Gordon et al. 2019; Andreas & Klein 2016). In its common form:
L0(w | u) ∝ ⟦u⟧(w) · P(w)
S1(u | w) ∝ exp( α · ( log L0(w | u) − C(u) ) )
L1(w | u) ∝ S1(u | w) · P(w)
The symbols are as follows:
- w is a possible state of the world or an intended referent, and u is an utterance.
- ⟦u⟧(w) is the literal meaning: 1 if u is true of w and 0 otherwise.
- P(w) is the prior probability of w.
- L0 is a literal listener, who simply conditions on truth.
- S1 is a pragmatic speaker. It prefers utterances that lead L0 to the intended w and that have low cost C(u). The parameter α is a softmax rationality parameter. The higher it is, the more reliably the speaker picks the utterance with the highest utility, and as α grows without bound it always does.
- L1 is a pragmatic listener, who reasons about why S1 would have said u.
Scalar implicature falls out of this setup. Take a world where all students passed. There, S1 would more likely say "all" than "some," because "all" identifies the world better. Hearing "some" therefore shifts L1 toward the not-all worlds.
The same pattern appears in decision-theoretic agents with nested beliefs. These are modelled as a decentralized partially observable Markov decision process (Dec-POMDP), in which agents act on private observations to maximize a shared reward. When such agents optimize that reward, scalar implicatures and mutual exclusivity arise without any dedicated pragmatic rules. Mutual exclusivity here means the tendency to map distinct messages onto distinct referents (Vogel et al. 2013).
A complementary view grounds Grice's relation maxim in a question under discussion (QUD). Here the context is a scoreboard holding shared information and a stack of accepted questions. A move counts as Relevant exactly when it addresses the current QUD (Roberts 2012).
History and founding papers
The concept goes back to Grice (1975). Horn's work on scales and Levinson's (2000) account of generalized, default-like implicatures form one family of theories. Sperber and Wilson's relevance theory, which stresses context-driven inference, forms another. LLM studies modelled on human processing experiments still contrast a Default model and a Context-driven model drawn from these two families (Cho & Kim 2024). Hirschberg (1985) treated scalar implicature in terms of partially ordered sets of alternatives. Later computational work on generation used a similar order-based view (Reiter 1990).
One early computational strand concentrated on natural language generation. Reiter showed how a generator could avoid triggering false implicatures in referring expressions. Describing the only table in a room as "brown wooden," for example, suggests that those details matter. Reiter recast the maxims as preference rules whose maximal descriptions can be found in polynomial time. By contrast, strict Full Brevity, which always demands the shortest distinguishing description, is NP-hard (Reiter 1990). Dale and Reiter then compared four computational readings of the maxims. They recommended the Incremental Algorithm, which adds attributes one at a time following a fixed, domain-specific ranking. It is fast, it never backtracks, and, like human speakers, it sometimes includes redundant properties (Dale & Reiter 1995).
On the interpretation side, Green and Carberry's account of indirect replies, outlined above, used plan operators and defeasible discourse relations. This extended coverage beyond scalar cases (Green & Carberry 1992). Roberts' QUD framework first circulated as a 1996 working paper and was published in 2012. It became the standard reference for QUD analyses. It also sketched how scalar implicature, among other phenomena, could fit into a unified formal pragmatics (Roberts 2012).
How it is used today
A major current use is probing and benchmarking language models. IMPPRES contains about 25.5K template-generated sentence pairs. Tests on it involved BERT trained on MultiNLI, a large natural language inference (NLI) dataset. This BERT reliably labels "not all" as entailed by "some." For most other scales, however, the models collapse the scale members into synonyms (Jeretic et al. 2020).
GRICE is a synthetic, grammar-generated dialogue set. On its conversational reasoning task, BERT reached 71.06% test accuracy against 97.50% for humans. Adding a module that explicitly recovers implicatures raised a memory network from 67.32% to 73.12% (Zheng et al. 2021).
PUB assembles about 28k multiple-choice items. Ten of its fourteen tasks target implicature and the rest cover presupposition, reference and deixis. Humans still outperform the models overall (Sravanthi et al. 2024). Other studies examine how fine-tuning strategy affects implicature resolution in large models (Ruis et al. 2022). Some test implicature in Chinese sitcom dialogue (Yue et al. 2024), and others argue for open-ended evaluation and preference tuning (Wu et al. 2024). Broader pragmatic suites likewise note that standard NLU benchmarks largely ignore implicature and speaker intent (Sileo et al. 2022).
RSA-style reasoning is used in grounded generation:
- Reference games. A pragmatic speaker that reranks neural caption samples succeeded in 81% of human-played games, against 69% for a contrastive baseline (Andreas & Klein 2016).
- Incremental reasoning. Word-by-word RSA captures anticipatory implicatures that arise from contrastive modifiers (Cohn-Gordon et al. 2019).
- Other applications. RSA has been applied to empathetic dialogue (Kim et al. 2021) and to lightweight controllable text generation (Wang & Demberg 2024).
- Sequential games. Models that combine planning with pragmatic inference predict human moves in sequential games (Khani et al. 2018).
- Survey view. A review of grounded tasks places reasoning about alternatives among the core roles of pragmatics (Fried et al. 2023).
Implicature also supports applied text analysis:
- Implicit sentiment. Defeasible rules over events that benefit or harm entities propagate the writer's sentiment through a graph. This raises precision by roughly 10 points over an explicit-only classifier (Deng & Wiebe 2014). The same rules can also serve as soft constraints in integer linear programming. Used that way, they raise sentiment F-measure by more than 10 points over local detectors that score each entity separately, from 0.4384 to 0.6421 on items with correct spans (Deng et al. 2014).
- Adversarial testing. Implicature-based edits feature among the pragmatic tricks that broke sentiment classifiers. Across all edit categories, not only implicature, the team's 99 minimal pairs broke systems at an average rate of 72.6% (Mahler et al. 2017).
- Implicit toxicity. Structured pragmatic inference chains have been proposed for detecting implicitly toxic language (Chen & Wang 2025).
Open questions
Do models compute implicatures or only imitate them? Strong results on "some/not all" coexist with failures on other scales (Jeretic et al. 2020). Models also err differently from people. For example, they read figurative agreement as sarcasm and stumble on items that people find easy (Sravanthi et al. 2024).
Is scalar implicature a default or a context-driven inference? The question persists for models just as it does for people. In one study, BERT ignored the type of QUD while GPT-2 struggled when the QUD required the implicature (Cho & Kim 2024).
Can explicit pragmatic reasoning be distilled into a single network? A direct speaker trained on RSA outputs kept little of RSA's advantage (Andreas & Klein 2016).
How well do models handle inferences that hinge on discourse-level pragmatic knowledge? Judging event factuality and choosing determiners under the Maximize Presupposition! principle both depend on context that models often fail to use (Jiang & Marneffe 2021; Sieker & Zarrieß 2023).
How much does the hearer's identity matter? Sociocultural background measurably shapes which inferences succeed (Shaikh et al. 2023).
Finally, how can pragmatic reasoning scale computationally? Exact nested-belief planning is intractable, so approximations are needed (Vogel et al. 2013). Evaluation still depends heavily on human judgement (Fried et al. 2023).
Key papers
Related concepts
- Discourse analysis — Discourse analysis is the study of how stretches of language beyond the single sentence form coherent wholes. It covers the relations that…
- Cognitive linguistics — Cognitive linguistics is an approach to language that treats grammar and meaning as products of general human cognition, including…
- Psycholinguistics — Psycholinguistics is the study of the mental processes by which people produce, comprehend and acquire language. It combines experimental…
- Coreference resolution — Coreference resolution is the task of finding the expressions in a text that refer to the same entity as one another, such as names…
More papers
Topics: 😉 Implicature · 💬 Speech act theory
Shelves: 🗣️ Linguistics & languages
This page in Markdown → https://md.research.elamur.ai/free/en/concepts/implicature/