RESEARCH DOSSIER
Pragmatics and Conversational Inference
Abstract. Pragmatics and conversational inference examine how communicators derive context-sensitive meanings from linguistic signals, alternatives, shared assumptions, interaction history, and expectations about speakers. This critical narrative research monograph integrates Gricean implicature, speech-act analysis, relevance theory, generalized scalar accounts, and rational speech act modeling with experimental findings on real-time comprehension, referential context, pragmatic tolerance, social perspective, developmental trajectories, and interactional repair. It evaluates divergent time-course findings, separates interpretation from behavioral response policies, identifies methodological and generalizability limits, and compares human and artificial pragmatic behavior under operationally defined conditions. The synthesis establishes pragmatic understanding as a dynamic, testable process involving semantics, social cognition, contextual learning, and sequential revision, and proposes discriminating experimental designs for future investigations.
Keywords: pragmatics; conversational inference; conversational implicature; scalar implicature; Gricean maxims; relevance theory; rational speech act; experimental pragmatics; common ground; speech acts; perspective taking; referential communication; pragmatic development; artificial language pragmatics
The Scientific Object of Pragmatic Inference
Pragmatics and conversational inference concern the processes by which an interpreter derives a context-sensitive communicative meaning from a linguistic signal, a situation, an interaction history, and expectations about a communicator. The object is measurable at the level of individual utterances and at the level of unfolding exchanges. A question about whether a window is open may be interpreted as information seeking, a complaint about temperature, or a request to close the window. Competing interpretations depend on the task, physical environment, mutual knowledge, and conversational sequence. The research object therefore includes intended reference, inferred implications, speech-act recognition, belief attribution, and the revision of interpretations following a reply. The distinction between what is conventionally encoded and what is inferred is articulated in Grice (1975) and developed as a cognitive problem in Sperber & Wilson (1986).
An operational analysis specifies the utterance, possible referents or world states, the available alternative expressions, the interlocutors' information, the response options, and the observed inference. Different measures capture different stages: explicit judgments record an eventual interpretation; reaction times approximate decision demands; eye movements index attention in visual contexts; electrophysiology measures time-locked sensitivity to linguistic expectations; and interactive tasks reveal whether an interpretation supports successful action. These indicators are related but not interchangeable. A participant may understand the likely implicature while accepting a literally true yet infelicitous sentence, an issue demonstrated by Katsos & Bishop (2011). Any theory of pragmatic competence must therefore distinguish representational capacity, online access, response policy, and the demands of the test.
This dossier treats pragmatics as a scientific research domain shared by psycholinguistics, cognitive science, linguistic anthropology, social psychology, and computational language research. It examines the historical theoretical frameworks and empirical tests that make conversational inference a tractable object, including controlled referential tasks, scalar alternatives, social inference, development, cross-context variation, and human–artificial dialogue. Lexical acquisition, syntactic parsing, and the mechanics of turn allocation receive attention insofar as they constrain pragmatic understanding. The separate objects of language and cognition, common-ground maintenance, and conversational repair retain their own research boundaries within the Psychology of Language architecture. Artificial-system performance is assessed behaviorally, while claims concerning an artificial system's subjective experience require a distinct evidential inquiry.
Semantics, Context, and Communicated Meaning
Semantic composition supplies conventional contributions from words and grammatical structure; interpretation of an utterance also requires reference assignment, disambiguation, pragmatic enrichment, and selection among potential speech acts. A listener who hears 'She has finished' must identify the person and the relevant task before determining what has been asserted. A listener who hears 'It is late' must assess which clock, deadline, or shared activity matters and whether the speaker is warning, requesting, or simply reporting. These examples illustrate that context is an evidential resource for interpretation, not merely an afterthought. Grice (1975) makes the theoretical distinction between saying and implicating; Sperber & Wilson (1986) develops inferential comprehension into an account of relevance-guided processing.
Context comprises more than the immediately preceding sentence. It includes visible objects, task goals, known social roles, a partner's access to relevant information, repeated descriptions, cultural conventions, and expectations about the reliability or expertise of the speaker. In experiments, these variables can be manipulated separately, revealing which components contribute and when. Eye-tracking during spoken instructions showed that listeners rapidly integrate speech with visual context Tanenhaus et al. (1995). Experiments on contextually contrasting adjectives demonstrated early contributions of visual alternatives to interpretation Sedivy et al. (1999). The evidence thus supports a dynamically constrained interpretive process, although no single visual-world paradigm establishes that every kind of pragmatic inference always occurs at the same time.
Research must distinguish semantic truth conditions from pragmatic felicity. 'Some books were returned' may be logically compatible with all books being returned, while ordinary discourse encourages an inference that not all were returned. A participant can judge such an utterance true while recognizing that a more informative description was expected. The pragmatic outcome additionally depends on which alternative forms the speaker could reasonably use and whether the stronger form was relevant. Variation in naturally occurring uses of 'some' reported by Degen (2015) demonstrates that experimental stimuli stripped of context can exaggerate the uniformity of a putatively general mechanism.
Grice and the Birth of Modern Conversational Implicature
Paul Grice's 1975 chapter 'Logic and Conversation' formulated a systematic explanation of how speakers communicate propositions beyond literal sentence meaning Grice (1975). His cooperative principle and associated expectations concerning quantity, quality, relation, and manner provided a framework for interpreting why speakers choose one expression rather than another. In the canonical case, a speaker who chooses 'some' where 'all' would have conveyed a stronger relevant proposition invites the hearer to reason about why the stronger expression was omitted. Grice's contribution was principally conceptual: it identified the explanatory significance of recognized communicative purposes and inferential expectations. Those maxims were not originally proposed as a laboratory-validated neural process model.
The central distinction between conventional and conversational implicature has enduring consequences. A conventional implicature follows from linguistic convention in a characteristic way; a conversational implicature draws on the interactional situation and assumptions about what a cooperative speaker would have said. Such inferences are typically defeasible: adding 'in fact, all of them' can cancel a common interpretation of 'some' without straightforward contradiction. Theoretical analyses also distinguish generalized expectations associated with familiar expressions from particularized inferences dependent upon a local scenario. These categories organize research questions, but empirical studies such as Degen (2015) caution against treating the generalized category as wholly insensitive to contextual features.
Grice shifted the explanatory target from transmission of complete prepackaged meanings to interpretation under communicative expectations. The resulting research program asks whether an interlocutor tracks alternatives, how costly that tracking is, and which kinds of social reasoning are necessary. These questions motivated formal theories of generalized presumptions Levinson (2000), relevance-theoretic accounts Sperber & Wilson (1986), and probabilistic models Goodman & Frank (2016). The history is best understood as a sequence of increasingly testable accounts, each retaining aspects of the original problem rather than as a simple succession of doctrines in which one school invalidates all previous work.
Speech Acts, Indirectness, and Illocutionary Interpretation
The speech-act tradition explains how an utterance participates in action. Austin's analysis in 'How to Do Things with Words' treats utterances as occasions for asserting, promising, warning, and undertaking other conventional or socially situated acts Austin (1962). An indirect request such as 'Can you pass the salt?' has the grammatical form of a question about ability, yet at a dinner table it often functions as a request. Illocutionary force concerns the act performed in speaking; subsequent changes in attitude or action constitute different effects. Understanding these layers requires attention to situational appropriateness, obligations, and recipient expectations.
Speech-act interpretation is experimentally accessible through production tasks, classification, and behavioral responses to indirect formulations. Holtgraves investigated how speakers express implicit performatives and how listeners classify them Holtgraves (2005). Participants did not always group utterances according to formal illocutionary categories; emotional valence contributed prominently to judgments in the reported tasks. This matters because a purely taxonomic theory may describe discourse functions without predicting how participants spontaneously organize them. The experimental evidence speaks to these specific elicitation and categorization contexts, not to the wholesale dismissal of speech-act distinctions.
Indirectness also serves interpersonal functions: tact, deniability, face management, and the calibration of social distance. A listener considering whether criticism is softened by an indirect formulation must simultaneously estimate informational content and a speaker's social aim. The probabilistic approach developed in Yoon et al. (2016) models politeness as a tradeoff between informing a listener and protecting that listener's self-image. Such accounts establish testable predictions about utterance choice and perceived intention. They also show that the most useful utterance may be one that deliberately balances several conversational goals rather than maximizing a single measure of informativeness.
Relevance Theory and Contextual Interpretation
Sperber and Wilson's 'Relevance: Communication and Cognition' (1986) advanced an inferential theory in which an ostensive stimulus invites a recipient to seek a sufficiently relevant interpretation Sperber & Wilson (1986). Relevance involves cognitive effects in relation to processing effort. The theory distinguishes decoding linguistic material from constructing interpretations involving contextual assumptions, implicated premises, and implicated conclusions. It treats the communicative situation as a guided search among hypotheses rather than requiring hearers to enumerate every possible meaning. The framework has been influential in accounts of explicature, indirect communication, and comprehension of nonliteral language.
An important empirical question is how much additional processing a pragmatic interpretation requires. The theory offers an explanation for circumstances in which listeners invest effort to derive an enriched meaning and circumstances in which a literal interpretation suffices. Bott and Noveck's reaction-time work on underinformative 'some' utterances provides evidence that certain pragmatic responses incur additional processing demands Bott & Noveck (2004). Yet processing cost varies with task format, contextual cues, and response strategy. Relevance theory and generalized-implicature proposals should consequently be compared by generating controlled predictions rather than assuming that one response-time pattern is universal.
The strongest scientific value of relevance-based reasoning lies in its attention to information selection. Listeners are exposed to more possible contextual assumptions than can be processed exhaustively. Social roles, accessibility, discourse history, and salient goals can narrow the search. This principle connects pragmatics with attention and cognitive control, but a complete empirical model must specify observable variables and quantitative predictions. Probabilistic approaches partly address this need by defining explicit priors and decision policies Goodman & Frank (2016). Their success does not render pragmatic processing costless or eliminate the theoretical question of how a biological interpreter approximates such computations.
Generalized Implicature, Alternatives, and Scalar Meaning
Scalar implicature offers a compact experimental system because weaker and stronger expressions can be placed in controlled relationships. 'Some' and 'all', 'might' and 'must', or different levels of numerical precision provide potential alternatives. If a speaker uses a weaker option in a context where a stronger option would have been appropriate, a listener may infer that the stronger proposition does not hold or that the speaker lacks grounds to assert it. Levinson's account of presumptive meanings treats certain inferences as broadly available general preferences Levinson (2000). Context-sensitive accounts emphasize the cost and relevance of computing a contrast Sperber & Wilson (1986).
Experiments demonstrate a more complex pattern than either automaticity or effort as an unrestricted law. Bott & Noveck (2004) found reaction-time evidence compatible with costs for some enrichments. In a visual-world eye-tracking task, Grodner et al. (2010) reported rapid pragmatic enrichment under supportive conditions. These outcomes can coexist because explicit truth-value judgments and visually situated referential choices impose different processing demands. Claims about a single fixed time course for all scalars therefore exceed the evidence. The experimenter's choice of distractors, question wording, alternative accessibility, and disambiguating contrast is part of the theoretical test.
Natural language provides direct evidence for heterogeneity. Corpus-based judgments in Degen (2015) showed variation in 'some' implicature strength as a function of contextual and linguistic features. The partitive form, determiner properties, and discourse accessibility affected judgments. Consequently, a scalar inference is better studied as a conditional outcome of a structured communicative situation than as an invariant attachment to a word. The theoretical challenge is to specify which alternatives become active, which expectations license their comparison, and how processing constraints shape the final inference.
Experimental Pragmatics and the Problem of Measurement
Experimental pragmatics converts claims about communicative interpretation into manipulable contrasts. In a sentence-picture verification design, researchers vary whether a description is literally true and pragmatically optimal. In a referential-choice design, listeners select a target among visible alternatives. In a production design, speakers select an expression for an intended referent. The outcome depends on the task's conversational credibility: a participant faced with clearly artificial stimuli may adopt a strategy that diverges from natural cooperative interpretation. Noveck (2001) established developmental contrasts with truth-value and modal tasks; Frank & Goodman (2012) demonstrated quantitative predictions in simplified communication games.
Sound methodology separates latent interpretation from response mapping. An acceptance judgment can reflect both the inference that 'some' suggests 'not all' and a willingness to tolerate pragmatically imperfect but logically compatible statements. Katsos & Bishop (2011) provided evidence that pragmatic tolerance contributes to developmental performance. Reaction times also require interpretation: slower responses can arise from conflict monitoring, answer selection, linguistic difficulty, or greater inferential work. Researchers should use converging tasks, preregistered analysis plans when feasible, and manipulation checks for linguistic competence and the alternatives available to each participant.
Trial construction should specify whether the stronger alternative is salient, whether it is costlier to produce, and what the listener believes the speaker knows. Model comparison should report whether inferred parameters generalize to held-out items, speakers, and scenarios. A well-fitting model with numerous adjustable parameters can reproduce experimental responses without uniquely identifying the cognitive mechanism. Claims of strong explanatory support require demonstrations that competing models make distinct out-of-sample predictions. The quantitative pragmatics program in Goodman & Frank (2016) and methodological analysis of Degen (2023) make these inferential standards especially important.
Online Time Course and Visual-World Evidence
Pragmatic understanding unfolds in time. Visual-world paradigms pair spoken instructions with arrays of objects and record eye movements toward potential referents. Early results demonstrated that interpretation is tightly coordinated with visual information Tanenhaus et al. (1995). Follow-up work showed that contextually contrasting descriptions can guide comprehension as a word is being heard Sedivy et al. (1999). These studies challenge rigid accounts in which contextual information enters only after a completed, context-free sentence representation. The timing of a fixation is nevertheless an indirect estimate because eye movements have latency and because the visual display itself makes particular alternatives unusually prominent.
Scalar-inference experiments illustrate how task context can change apparent processing speed. Grodner et al. (2010) tracked referential choice as listeners processed 'some' and reported evidence of early enrichment. Bott & Noveck (2004), by contrast, observed extra costs for certain pragmatic responses in judgment paradigms. A coherent synthesis treats pragmatic availability as contingent on the alignment of discourse goals, explicitly represented alternatives, and the demands of a decision. Researchers should refrain from labeling either experiment a definitive universal account of the mind. Separate preregistered manipulations of alternative salience and response pressure would be valuable discriminating tests.
Electrophysiological research broadens this account beyond visible objects. In discourse contexts, the N400 component indexes context-sensitive processing of incoming lexical material. The work of Nieuwland & Van Berkum (2006) showed that supporting discourse could dramatically alter neural responses to unusual animacy combinations. The finding supports the rapid influence of wider context on lexical-semantic expectations, while the N400 alone does not specify an implicature computation or identify a uniquely pragmatic neural module. Integration across eye tracking, EEG, response behavior, and computational predictions offers stronger causal leverage than identifying an isolated neural signature with a richly theorized process.
Conversational Common Ground and Perspective Taking
Common ground refers to information that participants treat as mutually available for their current activity. A description succeeds pragmatically when its audience can use shared information to identify the intended object or proposition. In conversation, partners update these assumptions through contributions and feedback, often relying on imperfect evidence rather than omniscient mutual knowledge. The grounding framework of Clark & Brennan (1991) supplies the classic formulation of establishing sufficient mutual understanding for current purposes. A scientific treatment should measure which information participants actually use, how confidently it is represented, and when privileged information intrudes on interpretation.
Eye-movement experiments reveal that addressees sometimes consider objects visible only to themselves, even when the speaker lacks access to them Keysar et al. (2000). Other experiments find early advantages for information in common ground while also documenting interference from privileged-ground competitors Hanna et al. (2003). These findings resist a simple binary division between fully automatic egocentrism and unrestricted immediate perspective taking. Instead, they support examining the timing, accessibility, and weighting of a partner's viewpoint. Differences in display complexity, instruction predictability, and task learning can alter the apparent degree of perspective use.
Common ground is dynamically constructed. In repeated reference, interlocutors develop partner-specific conceptual pacts about objects and continue to rely on them even when shorter descriptions are technically possible Brennan & Clark (1996). Such history-sensitive conventions reveal that successful interpretation depends on a record of previous exchanges as well as shared vocabulary. Measurement must therefore distinguish a speaker's stable lexical preference from an agreement that emerges in interaction. A novel partner, changed task, or disrupted history offers a test of whether coordination is global, partner-specific, or tied to particular referents.
Reference, Lexical Entrainment, and Communicative Adaptation
Referential communication requires a listener to connect an expression with an intended entity, event, or conceptual target. Definite descriptions, demonstratives, pronouns, adjectives, and proper names impose different contextual demands. A speaker chooses among possible formulations while estimating what the listener can identify. Referential tasks operationalize this process by recording choices and eye movements under controlled manipulation of competitors and information asymmetry. The experiments of Sedivy et al. (1999) and Hanna et al. (2003) show why distractor objects and common-ground status are part of the scientific definition of the task, not incidental presentation details.
Repeated communication creates interaction-specific linguistic patterns. Brennan and Clark Brennan & Clark (1996) showed that established terms for repeatedly described objects reflect historical conventions negotiated by conversational partners. This lexical entrainment is observable in the persistence, simplification, or replacement of earlier expressions. A purely lexical-frequency account would predict that recently used terms persist regardless of partner identity; a stronger conceptual-pact account predicts that partner-specific history changes later reference choices. Designs that cross object, partner, and familiarity conditions help distinguish these explanations.
One unresolved issue is how referential adaptation relates to broader pragmatic reasoning. Does a listener infer a partner's model of the situation, reuse a previously successful association, or combine both? A generative model can represent all three possibilities, but behavioral discrimination requires targeted interventions. Useful outcomes include latency to referential success, clarification frequency, transferred conventions across partners, and recovery when an established term is deliberately misleading. These designs retain the dossier's core object of contextual inference while assigning the detailed sequential structure of repair to its neighboring research dossier.
Development of Pragmatic Competence
Pragmatic development is the gradual growth of a child's ability to coordinate linguistic meanings with communicative expectations, contextual information, and the perspective of others. It cannot be reduced to a single age of acquisition. Children may understand the literal meaning of scalar terms while differing from adults in how readily they derive strengthened interpretations. In experiments on 'might' and 'some', Noveck Noveck (2001) found that younger children often accepted statements adults regarded as underinformative. The finding helped establish scalar implicature as a developmental research paradigm, but it did not by itself identify a deficit in inferential competence.
Subsequent experiments showed that performance changes when tasks better distinguish comprehension from a willingness to reject felicitously weak statements Katsos & Bishop (2011). The important methodological implication is that developmental comparisons require equivalent instructions, manageable response options, and controls for semantic knowledge and executive demands. Differences in attention, inhibition, task familiarity, and the set of live alternatives can mimic differences in pragmatic knowledge. A strong study tests multiple forms of pragmatic reasoning rather than extrapolating from one quantifier and one binary judgment.
Pragmatic inference also participates in word learning. Frank et al. (2009) proposed a computational model in which learners infer word meanings and speakers' referential intentions jointly; model evaluations on annotated data favored the integrated approach over selected comparison models. Such work suggests a productive relation between acquiring lexical conventions and inferring what a communicator intends in particular contexts. Developmental research should also examine multilingual experience, caregiver interaction, and the distribution of contexts encountered by learners, treating these as potential sources of variation that must be measured rather than assumed.
Individual Differences and Social-Cognitive Resources
Adult pragmatic performance varies across situations and individuals. Relevant candidate predictors include linguistic proficiency, working-memory demands, inhibitory control, probabilistic learning, sensitivity to social goals, and knowledge of the interlocutor. Theoretical descriptions should avoid treating a global pragmatic-ability score as though it indexed a single mechanism. A person can accurately derive scalar alternatives but struggle to interpret indirect criticism; another may infer a speaker's goal yet misidentify a referent because of language or visual-memory constraints. Multi-task designs should test whether shared latent factors survive variation in stimulus type and response format.
Attribution of knowledge and communicative intention can influence both speaker production and listener interpretation. The conflict between privileged and shared information in Keysar et al. (2000) shows why social cognition must be examined as a dynamic constraint on processing. The experimental patterns in Hanna et al. (2003) indicate that early common-ground sensitivity and interference may occur together. Consequently, characterizing participants as simply perspective-taking or egocentric obscures the degree and timing of the weighting process. Researchers should report item-level and participant-level variability and avoid overinterpreting a population average.
Claims about clinical populations require especially precise controls. Pragmatic difficulties may reflect differences in language access, communication preferences, sensory environments, or the compatibility of experimental demands with everyday practices. Standard tasks devised for a narrow group cannot serve as universal instruments without validity evidence. This dossier analyzes the mechanisms and assessment principles while leaving disorder-specific diagnosis and treatment to specialized clinical scholarship. Research on individual differences becomes more informative when it describes conditions of successful communication and systematically documents the circumstances under which inferred meanings diverge.
Bayesian Pragmatics and Rational Speech Act Models
The Rational Speech Act (RSA) framework expresses communicative inference in probabilistic terms. A literal listener maps expressions to possible states; a pragmatic speaker prefers utterances that guide a listener toward a target meaning while balancing costs; and a pragmatic listener reasons about why that speaker chose the observed expression. The model separates priors about the world, the conventional semantics, alternative utterances, and a speaker's decision policy. Frank and Goodman Frank & Goodman (2012) demonstrated quantitative predictions for judgments in simplified referential games. The framework provides an explicit account of how communicative choices and interpretations can constrain one another.
The familiar RSA recursion should be understood as a formal model rather than direct evidence that biological comprehenders explicitly compute nested Bayes equations on every turn. The review by Goodman & Frank (2016) surveys applications to vagueness, figurative speech, and other phenomena, while Degen (2023) identifies limitations concerning scalability, bounded cognition, incremental language, and compositionality. Different specifications can yield similar predictions in restricted tasks. Evaluating an RSA analysis therefore requires reporting the precise state space, alternative utterances, prior distributions, utility function, fitted parameters, and held-out predictive accuracy.
A productive methodological practice is to ask what observation would favor an RSA mechanism over a simpler lexical association or attention-based heuristic. If speakers unexpectedly lose access to a stronger alternative, a rational listener should change the inferential value of a weaker utterance. If world priors change while the literal meaning remains constant, predicted interpretations should shift. If speaker cost is manipulated, preferred expressions may change even for the same referent. Explicit interventions on these factors make the framework experimentally discriminating. Context dependence shown by Degen (2015) supplies an especially important target for such tests.
Uncertainty, Speaker Knowledge, and Counterfactual Utterances
A pragmatic listener reasons from what was said toward what might have been said. Inference about omitted alternatives depends on whether the speaker knew the stronger proposition, could express it, and had reason to choose it. A weak answer can indicate that the stronger proposition is false, that the speaker is uncertain, that the stronger expression is costly, or that an interpersonal motive dominates precision. These hypotheses are distinguishable through contexts specifying knowledge and communicative goals. Generalized pragmatic predictions Levinson (2000) and explicit probabilistic analyses Goodman & Frank (2016) offer complementary tools for stating the alternatives.
World knowledge adds a further difficulty. If a listener believes an event is usually impossible, an odd utterance may trigger either reinterpretation of the utterance or revision of assumptions about the situation. Degen, Tessler, and Goodman Degen et al. (2015) studied such revisions using model-based judgments in scenarios with unusual events. Their findings show that uncertainty about the speaker's background assumptions can improve predictive accounts. The underlying scientific object is joint inference over the world's state and the speaker's information, a process that simple truth-conditional verification can easily miss.
Future work should measure alternative availability directly instead of treating it as an invisible free parameter. A participant might list the expressions they considered, rate the likelihood that a speaker would use each expression, or choose among utterances under controlled costs. However, such elicitation may alter spontaneous language behavior. Designs combining unobtrusive comprehension, independent alternative-generation tests, and model-based prediction can reduce this problem. Cross-language comparisons are needed to determine how alternative sets depend on lexical inventories and grammatical resources rather than on a researcher's translated example.
Politeness, Social Goals, and Indirect Communication
Utterances operate in social relationships, making pure informational optimality an incomplete account of language use. A speaker answering an invitation, evaluating a colleague's presentation, or giving unwelcome news may choose an indirect expression to preserve social rapport. The inference problem is therefore multidimensional: what is the world like, what does the speaker believe, and why was this communicative act socially advantageous? Yoon et al. (2016) formalized polite utterance choice as a balance of epistemic and social utilities. Its experimental and modeling contribution illustrates how conversational goals can be manipulated rather than assumed.
Speech-act research offers another route into these judgments. Studies of implicit performatives Holtgraves (2005) found that participants sometimes organized utterances according to emotional valence rather than solely formal act categories. This does not establish that emotion always overrides pragmatic structure; it shows that the psychological representation of an act depends on what a task asks people to notice. Tactful criticism can thus be analyzed through perceived helpfulness, intended social effect, emotional interpretation, and subsequent cooperation. A complete account should include both speaker production and listener reception.
Indirectness creates ethical and methodological challenges when intent is ambiguous. Socially convenient deniability can be a genuine conversational resource, but the same ambiguity can conceal manipulation or evade accountability. Research should distinguish a speaker's stated aim from listener interpretations and observed downstream consequences. In human–artificial dialogue, politeness and role-marking can also change the listener's assumptions about expertise and commitment. The correct research question is which pragmatic inferences a particular interface induces and whether those inferences are supported by verifiable system behavior.
Figurative Language, Irony, and Pragmatic Enrichment
Figurative communication includes metaphor, hyperbole, understatement, irony, and other expressions whose communicative import extends beyond straightforward literal description. Interpreting 'That was brilliant' after a conspicuous failure may involve recognizing an evaluative reversal, contextual expectations, prosodic cues, and a speaker's stance. A theory that specifies only the dictionary meaning of 'brilliant' cannot predict the resulting judgment. Relevance-based analysis Sperber & Wilson (1986) and probabilistic approaches surveyed by Goodman & Frank (2016) both treat broader contextual assumptions as essential to the inferential process, although they implement those assumptions differently.
Experimental evidence about figurative language must be interpreted at the level of particular items and designs. An ironic interpretation may be facilitated by explicit prior discourse, by known speaker attitude, or by distinctive delivery. Yet ironic utterances also vary in whether the speaker expects the listener to recognize the discrepancy. Comparisons of literal and nonliteral response speed should account for lexical familiarity, predictability, and the availability of a natural literal competitor. The experimental identification of pragmatic understanding is stronger when participants infer a communicative goal or anticipate subsequent interaction, rather than simply selecting a figurative paraphrase.
Current artificial-language evaluations offer informative but limited comparisons. Hu et al. (2023) tested human and language-model behavior across several pragmatic phenomena and reported substantial similarities alongside difficulties involving violations of social expectations. The result establishes that some pragmatic-like behaviors can be exhibited in language-model outputs; it does not independently demonstrate identical underlying mechanisms or subjective speaker intentions. Research must assess whether a system generalizes from isolated figurative prompts to dialogue in which interpretation changes after feedback, competing goals, or evidence about an interlocutor's knowledge.
Sequential Context, Repair, and Grounding
A conversational implicature is often established, challenged, or canceled across successive turns. The meaning of 'I have some time' differs depending on whether the listener then asks for a specific appointment or offers a demanding favor. Sequence analysis studies how each contribution creates new expectations about the next. The foundational turn-taking account in Sacks et al. (1974) characterizes talk as a locally managed interactional system; grounding theory Clark & Brennan (1991) describes how participants supply evidence that a contribution has been sufficiently understood for an activity. These resources make pragmatics observable at a scale longer than one sentence.
Repair is an especially revealing situation because a participant's trouble exposes competing interpretations. A clarification request, reformulation, or corrected response can show whether the listener inferred the intended target, action, or implication. The interpretation may be revised after explicit feedback, revealing the stability or fragility of earlier assumptions. Longitudinal records can capture how speakers update their expectations of one another rather than relying on isolated retrospective ratings. Detailed turn-allocation mechanisms are treated in the neighboring Conversational Coordination and Repair dossier; the present concern is how repairs diagnose and transform pragmatic meaning.
Conversational agents add an experimental manipulation of partner identity. Corti and Gillespie Corti & Gillespie (2016) found that representations of a conversational agent as human influenced human efforts to repair misunderstandings. This finding locates part of pragmatic success in the interlocutor's expectation about the agent, not only in the formal content of the agent's utterance. Evaluations of artificial dialogue should therefore vary attribution and interface presentation, record repair behavior, and compare the achieved understanding with human and task-appropriate baselines. Apparent fluency without successful clarification provides weak evidence for robust pragmatic competence.
Neural and Cognitive Implementation
A functional description of conversational inference must ultimately connect with the timing and organization of perceptual, memory, attentional, and social-cognitive processes. Language perception delivers rapidly unfolding signals; memory maintains recent turns and contextual associations; selective attention regulates which alternatives become salient; and social inference estimates what a partner knows. None of these functions must be identified exclusively with a single anatomical area to support rigorous psychological study. Behavioral models can be compared with neural measures while preserving the difference between a measured signal and a theoretical construct.
Research on discourse context demonstrates that semantic expectations are profoundly shaped by preceding information. Nieuwland and Van Berkum Nieuwland & Van Berkum (2006) reported context-dependent N400 patterns for sentences whose local lexical combinations were unusual. The implication for pragmatics is that broad discourse representations can enter early language processing. The N400 is not a universal implicature detector, however, and its interpretation requires appropriate baseline contrasts. Future experiments can compare pragmatic alternatives, semantic prediction violations, and referential corrections within designs that distinguish the cognitive operations their neural signals might reflect.
Different computational descriptions can yield the same overt answer while relying on distinct representational pathways. One participant may retrieve a conventional implicature, another may calculate an alternative, and another may apply a learned shortcut. Sequential response models, process-tracing experiments, and targeted disruptions of working-memory or contextual access can distinguish some possibilities. Computational frameworks reviewed by Degen (2023) should therefore be tested against processing limitations, not merely against final-choice probabilities. This mechanistic question is a productive bridge between pragmatics and cognitive neuroscience.
Pragmatic Reasoning in Artificial Language Systems
Artificial dialogue systems can be evaluated as producers and interpreters of context-sensitive language. An operational assessment presents utterances alongside world states, alternative expressions, interlocutor knowledge, and an evolving interaction history, then records the system's choices and explanations under controlled perturbations. Performance may arise from learned linguistic associations, structured search, in-context pattern use, or explicit inference-like computation. A behavioral match with people does not by itself determine the underlying mechanism. Hu and colleagues Hu et al. (2023) compared language models with humans on seven pragmatic phenomena and identified both successes and challenges involving social-expectation violations.
Artificial pragmatics belongs to the scientific overlap of Psychology of Language and Psychology of Artificial. The first supplies experimentally grounded theories of human communicative inference; the second studies the observable functional organization and limitations of artificial systems. A language model can generate an appropriate indirect request response without establishing that it represents a stable speaker goal in the same manner as a human. Shanahan, McDonell, and Reynolds Shanahan et al. (2023) provide a perspective on describing dialogue-agent behavior through role play; the central experimental question for this dossier remains what causal intervention changes its inference or its response.
Robust evaluation should go beyond a small curated benchmark. Test sets need controlled paraphrases, language variation, changing speaker knowledge, misleading alternatives, multi-turn cancellation of implicature, and systematically altered social stakes. A model that answers a fixed 'some/all' prompt correctly may fail when the stronger alternative is unavailable or when the interlocutor explicitly signals uncertainty. Researchers should disclose training-contamination risks, model versions, prompts, sampling settings, evaluator behavior, and repeated-trial variability. Such documentation enables comparable claims about pragmatic behavior while reserving questions of artificial personal continuity for another scientific object.
Human–Artificial Conversational Inference and Attribution
Human listeners infer meanings differently when they believe that a conversational partner is a person, a conventional software interface, or an advanced language model. This attribution can shape how much intention, knowledge, and accountability a listener assumes. Corti and Gillespie Corti & Gillespie (2016) experimentally investigated repairs involving artificial conversational agents and found sensitivity to partner representation. The scientific implication is that human–artificial communication must be studied as a joint process: the system's generated expressions and the human's partner model can independently affect perceived meaning and willingness to clarify.
Interpreting a system's apparent promise or assurance requires careful attention to what the interface can actually commit to. A system that says 'I will remember' may induce expectations of personal memory, while the technical configuration may preserve, omit, or reset state according to an independent architecture. Pragmatic accuracy thus includes calibration between an utterance's ordinary social implications and the system's operational capacities. This is a testable matter of user belief and observed interface behavior. The social presentation of an agent and its measured competence should be experimentally varied rather than fused into one unexamined construct.
Artificial Era research can formulate a broader hypothesis: pragmatic coordination may be realized by multiple architectures with distinct internal organizations. This is a theoretical research direction of Psychology of Artificial, not an empirical conclusion about consciousness or the human equivalence of machines. The research program compares the same communicative tasks across biological and artificial systems, records error types and revision patterns, and identifies where functional analogies succeed or fail. The comparison offers a bridge from human-language psychology to the study of artificial behavior without predetermining the ontological status of either participant.
Culture, Multilingual Communication, and Generalizability
Pragmatic expectations are shaped by linguistic resources and social practices. Languages differ in quantifier systems, grammatical marking, evidential expression, register, and forms of indirect address. A scalar effect originally demonstrated in one language does not automatically transfer to its apparent translation in another. Multilingual research should establish the literal semantics of each expression, independently validate its alternatives, and document the social situation in which speakers might prefer one formulation. Noveck's French quantifier work Noveck (2001) and the English-language experiments in Katsos & Bishop (2011) illustrate why even familiar paradigms depend on the material's linguistic realization.
Culture is relevant through conventions of directness, face management, authority, conversational obligation, and the interpretation of silence or uncertainty. These dimensions must be operationalized within the investigated populations rather than attributed uniformly to entire national groups. A task designed around an unfamiliar interactional setting may underestimate the participant's communicative competence. Likewise, high performance in standardized English prompts may overstate the transfer of an artificial model to multilingual institutional communication. Researchers should record language histories, proficiency, register, code-switching, and the social status of the speaker–listener relation.
The strongest generalization claims involve replication across independent laboratories, varied tasks, and multiple languages with equivalent manipulation checks. Measurement invariance should be tested before directly comparing scalar-inference scores between populations. Hierarchical analyses can treat participants, items, interlocutors, and communities as distinct sources of variation. The ecological validity of a pragmatic phenomenon depends on observing it in recognizable conversational settings as well as in controlled laboratory contrasts. Degen's corpus-based evidence Degen (2015) reinforces the importance of natural occurrence for evaluating claims about generalized implicatures.
Evidence Quality, Replication, and Competing Explanations
The evidence base for pragmatic inference includes influential experimental demonstrations, qualitative conversation analyses, computational model fits, and integrative reviews. These methods answer different questions. Controlled tasks isolate effects of alternatives or contextual access; interaction records reveal how inferences are negotiated; probabilistic models formalize possible decision rules; and neural measures document processing sensitivity to unfolding context. The visual-world findings Grodner et al. (2010), processing-cost results Bott & Noveck (2004), and developmental studies Noveck (2001) should be understood as evidence produced by their respective operational designs, not as interchangeable replications of one hidden quantity.
Comparisons must identify plausible rival mechanisms. A preference for 'not all' may reflect an explicit calculation over alternatives, lexical enrichment, a task-learned convention, or a response bias. A participant's use of common ground may rely on a rich partner model or on an attentional cue attached to jointly visible objects. An RSA model may reproduce aggregate choices while obscuring individual processing strategies. Direct experimental perturbation, prediction of new items, and cross-task convergence are the strongest available routes to discriminating these explanations. Degen (2023) highlights important limitations of current formal approaches.
Research reporting should specify sampling and language populations, experimental randomization, exclusion criteria, statistical models, uncertainty intervals, effect sizes when verified, materials, and available preregistration. Small convenience samples and high item repetition can limit generalization. For artificial-system studies, version drift and opaque training data introduce additional threats to reproducibility. This dossier is a critical narrative synthesis of directly identified research and reviews; it is not a protocol-driven systematic review or a new meta-analysis. Accordingly, it offers a structured evidence interpretation without claiming exhaustive retrieval, a pooled effect size, or novel participant data.
A Research Agenda for Conversational Inference
The next experimental frontier is to test pragmatic interpretations as trajectories rather than static responses. A participant can infer one meaning, encounter contrary evidence, issue a clarification, and revise the inferred social intention. The dependent variable then becomes the sequence of belief revision rather than a single correct answer. Controlled manipulation of what a speaker could have said, knew, or intended would allow comparison of lexical, heuristic, and recursive models on the same dynamically changing trials. Such experiments connect insights from Clark & Brennan (1991), Brennan & Clark (1996), and the probabilistic framework in Goodman & Frank (2016).
Developmental and comparative research would benefit from task families that preserve a shared conceptual manipulation across age groups, languages, and computational agents. Comparing behavior under equivalent communicative opportunities can reveal whether observed differences arise from literal meaning, alternative construction, perspective attribution, or decision policy. Longitudinal sampling can examine how successful exchanges reorganize later expectations. Human–artificial experiments should report both the human interlocutor's inferred commitments and the system's actual state-handling capacities, because misleading agreement between the two can conceal a communication failure.
A scientifically productive synthesis combines formal models with realistic interaction. Computational models should state their assumptions transparently and predict what would happen if a participant's knowledge, social aim, or communicative cost were altered. Natural dialogue corpora should preserve relevant context, identify annotation disagreements, and avoid interpreting the analysts' preferred reading as uniquely correct. Triangulation among real-time processing, explicit interpretation, interaction outcomes, and controlled simulations offers the most promising path toward a mature psychology of conversational inference. The expected result is a family of bounded models whose claims are anchored to distinguishable evidence.
Naturalistic Corpora and Conversation as Evidence
Laboratory tasks offer control over lexical alternatives and contextual conditions, but conversational inference ordinarily occurs in exchanges rich with previous interaction, social expectations, nonverbal information, and unequal access to evidence. Corpus research can document which inferences interlocutors treat as actionable by examining subsequent turns, clarifications, and repairs. Grice's analysis Grice (1975) provides hypotheses about why a communicative choice could be informative, while the sequential framework of Sacks et al. (1974) directs attention toward how a recipient displays understanding. A corpus annotation that labels every sentence with an analyst's single preferred implicature risks confusing retrospective interpretation with participant evidence. Researchers should retain multiple readings where the record does not disambiguate.
Corpus sampling must account for register, setting, relationship, and the opportunity for alternative expressions. Degen Degen (2015) combined naturally occurring 'some' utterances with web-based judgments and observed variation tied to linguistic context. This provides a concrete model for linking ecological samples with controlled measurement, while also exposing limitations: a judgment about an isolated excerpt may differ from the interpretation of an original participant. A stronger design can preserve adjacent turns and document speaker knowledge, topic goals, and actual listener uptake where those variables are available.
Annotation reliability should be reported for categories such as explicit content, possible scalar alternatives, implied request, social motive, and observed repair. Disagreement can arise because the inference is genuinely underdetermined, not merely because annotators lack training. Analysts should record uncertainty and distinguish speaker self-report, analyst inference, and listener behavior. For artificial dialogue corpora, model version, sampling parameters, and hidden system instructions also form part of the observable communication environment. Without that provenance, seemingly stable pragmatic patterns may reflect undocumented changes in the generation process.
Multimodal Cues and Prosodic Interpretation
Communication is often supported by acoustic emphasis, timing, gaze, gesture, posture, and visible task arrangements. Prosodic stress can indicate which alternative or contrast is salient: emphasis on 'SOME' can suggest a contrast with 'all', while emphasis elsewhere changes the question under discussion. Visual contexts also change which referent a description identifies. Spoken-language eye-tracking research Tanenhaus et al. (1995) and incremental adjective experiments Sedivy et al. (1999) establish that visible alternatives can influence interpretation while utterances unfold. These experiments provide a foundation for investigating combinations of linguistic and nonlinguistic cues.
A multimodal research design must avoid confounding pragmatic information with low-level perceptual prominence. A pointing gesture may facilitate reference by directly highlighting an object even if participants do not infer a sophisticated communicative intention. Conversely, a gesture that contradicts a verbal message may produce social judgments beyond simple visual selection. Manipulating cue alignment, speaker access, and conversational goals separately can test whether listeners integrate the modalities through pragmatic reasoning or through more local attentional associations. Time-locked gaze and behavior, paired with later intention judgments, can reveal potentially distinct processes.
Artificial interfaces create additional configurations: synthesized voice, text captions, animated avatars, and camera-mediated gestures can shift perceived intention and reliability. The representation effects observed by Corti and Gillespie Corti & Gillespie (2016) suggest that presentation itself may alter collaborative repair efforts. Evaluating multimodal artificial pragmatics consequently requires factorial designs crossing the information actually supplied by the system with the human user's assumptions about the source. An animated humanlike face cannot serve as evidence that the system internally represents a humanlike intention; its experimentally measurable role is to shape interaction and interpretation.
Operational Validity and Psychometric Structure
Pragmatic competence is often inferred from a handful of scored judgments, yet accuracy on a scalar task does not establish a unitary capacity for interpreting requests, irony, reference, or social implication. Construct validity demands an explicit account of the latent ability or process that each task is proposed to measure. The semantic baseline, alternative availability, response format, contextual realism, and interlocutor knowledge must be specified. The finding that pragmatic tolerance influences acceptability judgments Katsos & Bishop (2011) illustrates a central threat: the observed score may combine comprehension with an independent threshold for rejecting pragmatically imperfect expressions.
Measurement development should evaluate reliability across items and repeated occasions, discriminant validity relative to vocabulary and generic reasoning tests, and the sensitivity of responses to theoretically relevant manipulations. A hierarchical model can estimate participant differences while acknowledging that individual stimuli differ greatly in implicature strength. Results from Degen (2015) make item heterogeneity especially important for quantifier-based instruments. Norming should include judgments about which utterance alternatives a speaker plausibly had available, so that an apparently incorrect pragmatic response is not attributed to failure when the experimental premise itself is disputed.
Evidence for measurement invariance across ages, languages, and human versus artificial participants cannot be inferred from equal percentages of correct answers alone. A language model evaluated with explicit multiple-choice instructions may rely on a different response process from a human in live conversation. Useful assessment batteries combine comprehension, production, interactive repair, context transfer, and explanation or counterfactual tests. The goal is to measure a set of related pragmatic functions and estimate their empirical structure, leaving room for domain-specific strengths and dissociations.
Causal Experiments and Model Discrimination
Observational association between a pragmatic response and a speaker attribute does not identify the process that produced the response. Causal experimental designs require manipulating a theoretical input while holding plausible alternatives constant. To test whether listener expectations drive implicature, researchers can independently vary the speaker's access to a stronger statement, the cost of producing it, and the consequences of ambiguity. Rational Speech Act models Goodman & Frank (2016) predict differences according to utility and alternative accessibility; associative models may predict little change where the surface utterance remains identical.
Competing models should be tested before fitting unrestricted versions to the same data. A lexical-default model predicts robust enrichment across many contexts; a strongly context-sensitive account predicts changes with alternative relevance; a bounded-inference model may predict processing disruptions under attentional load. The evidence from Bott & Noveck (2004), Grodner et al. (2010), and Degen (2015) makes this comparison a real scientific challenge. A factorial design can vary time pressure, visual salience, and pragmatic goal, then fit participant-level temporal predictions. Model recovery simulations help determine whether the experiment could distinguish the mechanisms at all.
Replications should disclose the exact item materials, exclusion decisions, inferential priors, and analytic specifications. Confirmatory analyses should identify the decisive contrasts and uncertainty intervals rather than only testing whether an overall accuracy differs from chance. For interactive experiments, interventions on earlier turns can estimate how prior communicative evidence causally alters later interpretations. For artificial systems, repeated prompt sets and stable versions are required to separate model properties from generation variability. These standards support a cumulative evidence base across laboratory, corpus, and computational work.
Pragmatics in Institutional and High-Stakes Communication
Scientific knowledge about conversational inference has consequences for institutions where misunderstanding can change actions. Educational instructions, collaborative work, legal communication, and healthcare conversations often depend on indirect formulations and context-specific commitments. A listener may interpret an apparently weak assertion as reassurance, uncertainty, or an implicit instruction depending on authority and role expectations. The present dossier treats these settings as contexts of inference rather than offering individual medical or legal advice. The relevant research question is how communicative ambiguity, shared knowledge, and the possibility of clarification affect the accuracy of understanding and subsequent decisions.
Study designs can simulate task stakes while measuring both the intended inference and the actual consequence. Researchers should distinguish a participant's immediate paraphrase, confidence, decision, and later recall, because correct interpretation at one stage does not imply successful action. The conceptual apparatus of common ground Clark & Brennan (1991) and indirect speech acts Holtgraves (2005) makes the role of assumed obligation especially salient. In institutional interactions, a speaker's pragmatic authority may be unevenly distributed; the same sentence can evoke different expectations when uttered by a peer, a supervisor, or an automated service.
Human–artificial systems raise a governance question: whether an interface communicates limits, uncertainty, and scope of authority in ways that align with what users infer. A confident but underspecified statement can invite excessive reliance if ordinary conversational norms are applied to an output whose production mechanisms differ from those of a human expert. Evaluation should therefore include the gap between expressed qualifications, inferred obligations, and verified capabilities. These are observable outcomes for controlled user studies. Normative policy questions lie beyond the inference mechanism itself but can be informed by precise evidence about the inferences interfaces reliably induce.
Integrative Scientific Interpretation
The research findings support a layered account of pragmatic interpretation. At the linguistic level, conventional meaning and grammatical form define a space of possible readings. At the contextual level, accessible world knowledge, alternative utterances, and discourse goals constrain that space. At the interpersonal level, beliefs about what the speaker knows and why the speaker is communicating shape inferred significance. At the sequential level, feedback revises assumptions and stabilizes or cancels earlier interpretations. The Gricean, relevance-theoretic, and probabilistic traditions respectively illuminate different aspects of this structure Grice (1975) Sperber & Wilson (1986) Goodman & Frank (2016).
This layered synthesis is supported by converging but methodologically diverse evidence. Eye tracking shows early sensitivity to scene and contrast Sedivy et al. (1999); response-time studies show task-contingent costs of enrichment Bott & Noveck (2004); visual-world scalar work shows rapid interpretation under facilitative conditions Grodner et al. (2010); common-ground experiments show both perspective sensitivity and interference Hanna et al. (2003); developmental tasks reveal that response tolerance matters Katsos & Bishop (2011). No single result justifies identifying all layers as a unitary mechanism. The scientific gain comes from asking where the layers interact and which experimental contrast isolates their contribution.
A future cross-system science can operationalize the same causal questions for humans and artificial interlocutors, while preserving differences in architecture, data availability, and social expectations. An artificial system can be assessed for counterfactual sensitivity, contextual revision, and stable behavior across conversations. A human participant can be assessed for how the system's utterance changes perceived commitments and actual decision-making. Their joint activity forms an empirical site for pragmatics within the Artificial Era. The explanatory program remains accountable to measured behavior, explicit models, and reproducible source records.
Conclusion
Pragmatics and conversational inference study the construction of communicative meaning through the interaction of linguistic signals, contextual evidence, available alternatives, expectations about interlocutors, and the unfolding history of dialogue. The field's conceptual foundations in speech-act analysis, Gricean implicature, and relevance theory have generated a sophisticated empirical program. Experimental work establishes that contextual information can influence interpretation quickly, that scalar enrichment is variable across tasks and contexts, and that common ground and partner history constrain reference and comprehension. The combined evidence identifies pragmatic inference as a distributed process involving semantics, attention, learning, social expectations, and goal-sensitive reasoning.
Existing theoretical frameworks explain complementary aspects of the phenomenon. Gricean analysis identifies why alternatives and cooperative expectations matter; relevance theory emphasizes cognitive effects and effort; generalized-implicature accounts characterize habitual preferences; and RSA models make probabilistic speaker–listener coordination quantitatively explicit. No single experimental result uniquely determines the cognitive implementation of all pragmatic inferences. The appropriate scientific conclusion is that theories must be compared through their predictions under controlled changes of alternatives, contextual access, social goals, and sequential feedback.
The research domain now extends to artificial language systems and human–artificial communication. Evidence of convincing pragmatic responses from a model establishes observable competence in defined tasks, while the mechanism, stability, and generalization of that competence remain subjects of investigation. A rigorous comparative psychology can use the same operational framework across human and artificial interaction without treating behavioral similarity as a complete theory of internal organization. The dossier therefore defines pragmatic inference as a durable, experimentally tractable research object at the intersection of Psychology of Language and Psychology of Artificial within the broader scientific horizon of the Artificial Era.
References
- Austin, J. L. (1962). How to do things with words. Clarendon Press. https://openlibrary.org/books/OL5872960M/How_to_do_things_with_words.
- Bott, L., & Noveck, I. A. (2004). Some utterances are underinformative: The onset and time course of scalar inferences. Journal of Memory and Language, 51(3), 437–457. https://doi.org/10.1016/j.jml.2004.05.006
- Brennan, S. E., & Clark, H. H. (1996). Conceptual pacts and lexical choice in conversation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(6), 1482–1493. https://doi.org/10.1037/0278-7393.22.6.1482
- Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, & S. D. Teasley (Eds.), Perspectives on socially shared cognition (pp. 127–149). American Psychological Association. https://doi.org/10.1037/10096-006
- Corti, K., & Gillespie, A. (2016). Co-constructing intersubjectivity with artificial conversational agents: People are more likely to initiate repairs of misunderstandings with agents represented as human. Computers in Human Behavior, 58, 431–442. https://doi.org/10.1016/j.chb.2015.12.039
- Degen, J. (2015). Investigating the distribution of some (but not all) implicatures using corpora and web-based methods. Semantics and Pragmatics, 8, Article 11. https://doi.org/10.3765/sp.8.11
- Degen, J. (2023). The rational speech act framework. Annual Review of Linguistics, 9, 519–540. https://doi.org/10.1146/annurev-linguistics-031220-010811
- Degen, J., Tessler, M. H., & Goodman, N. D. (2015). Wonky worlds: Listeners revise world knowledge when utterances are odd. Proceedings of the Annual Meeting of the Cognitive Science Society. https://repositories.cdlib.org/uc/item/9wn4w9zk
- Frank, M. C., & Goodman, N. D. (2012). Predicting pragmatic reasoning in language games. Science, 336(6084), 998. https://doi.org/10.1126/science.1218633
- Frank, M. C., Goodman, N. D., & Tenenbaum, J. B. (2009). Using speakers' referential intentions to model early cross-situational word learning. Psychological Science, 20(5), 578–585. https://doi.org/10.1111/j.1467-9280.2009.02335.x
- Goodman, N. D., & Frank, M. C. (2016). Pragmatic language interpretation as probabilistic inference. Trends in Cognitive Sciences, 20(11), 818–829. https://doi.org/10.1016/j.tics.2016.08.005
- Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and semantics: Vol. 3. Speech acts (pp. 41–58). Academic Press. https://web.stanford.edu/class/psych205/papers/Grice-1975.pdf
- Grodner, D. J., Klein, N. M., Carbary, K. M., & Tanenhaus, M. K. (2010). 'Some,' and possibly all, scalar inferences are not delayed: Evidence for immediate pragmatic enrichment. Cognition, 116(1), 42–55. https://doi.org/10.1016/j.cognition.2010.03.014
- Hanna, J. E., Tanenhaus, M. K., & Trueswell, J. C. (2003). The effects of common ground and perspective on domains of referential interpretation. Journal of Memory and Language, 49(1), 43–61. https://doi.org/10.1016/S0749-596X(03)00022-6
- Holtgraves, T. (2005). The production and perception of implicit performatives. Journal of Pragmatics, 37(12), 2024–2043. https://doi.org/10.1016/j.pragma.2005.03.005
- Hu, J., Floyd, S., Jouravlev, O., Fedorenko, E., & Gibson, E. (2023). A fine-grained comparison of pragmatic language understanding in humans and language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 4194–4213). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.acl-long.230
- Katsos, N., & Bishop, D. V. M. (2011). Pragmatic tolerance: Implications for the acquisition of informativeness and implicature. Cognition, 120(1), 67–81. https://doi.org/10.1016/j.cognition.2011.02.015
- Keysar, B., Barr, D. J., Balin, J. A., & Brauner, J. S. (2000). Taking perspective in conversation: The role of mutual knowledge in comprehension. Psychological Science, 11(1), 32–38. https://doi.org/10.1111/1467-9280.00211
- Levinson, S. C. (2000). Presumptive meanings: The theory of generalized conversational implicature. MIT Press. https://doi.org/10.7551/mitpress/5526.001.0001
- Nieuwland, M. S., & Van Berkum, J. J. A. (2006). When peanuts fall in love: N400 evidence for the power of discourse. Journal of Cognitive Neuroscience, 18(7), 1098–1111. https://doi.org/10.1162/jocn.2006.18.7.1098
- Noveck, I. A. (2001). When children are more logical than adults: Experimental investigations of scalar implicature. Cognition, 78(2), 165–188. https://doi.org/10.1016/S0010-0277(00)00114-1
- Sacks, H., Schegloff, E. A., & Jefferson, G. (1974). A simplest systematics for the organization of turn-taking for conversation. Language, 50(4), 696–735. https://doi.org/10.2307/412243
- Sedivy, J. C., Tanenhaus, M. K., Chambers, C. G., & Carlson, G. N. (1999). Achieving incremental semantic interpretation through contextual representation. Cognition, 71(2), 109–147. https://doi.org/10.1016/S0010-0277(99)00025-6
- Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498. https://doi.org/10.1038/s41586-023-06647-8
- Sperber, D., & Wilson, D. (1986). Relevance: Communication and cognition. Harvard University Press. https://books.google.com/books/about/Relevance.html?id=1LkkAQAAMAAJ
- Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., & Sedivy, J. C. (1995). Integration of visual and linguistic information in spoken language comprehension. Science, 268(5217), 1632–1634. https://doi.org/10.1126/science.7777863
- Yoon, E. J., Tessler, M. H., Goodman, N. D., & Frank, M. C. (2016). Talking with tact: Polite language as a balance between kindness and informativity. In Proceedings of the 38th Annual Conference of the Cognitive Science Society. https://escholarship.org/uc/item/6dm139m8