Computational pragmatics
Computational pragmatics is the study of computational models that relate linguistic expressions to their contexts of use. It addresses aspects of interpretation that cannot be determined from grammatical structure and conventional meaning alone, including the identification of intended referents, the recovery of implicit content, and the recognition of communicative intentions. The field lies at the intersection of pragmatics, computational linguistics, and artificial intelligence.
A computational pragmatic model represents communication as inference over linguistic form, discourse history, shared information, and assumptions about interacting agents. This formulation distinguishes the field from computational semantics, which primarily associates expressions with literal or compositionally derived meanings. The distinction is analytical rather than absolute because semantic representations commonly serve as inputs to pragmatic inference, while contextual interpretation can affect the semantic structures assigned to ambiguous expressions.
Theoretical basis
The central problem of computational pragmatics is underdetermination. A single utterance can support several interpretations compatible with its grammar and lexical meaning, while conversational participants generally select an interpretation appropriate to the current situation. Computational models reproduce this selection by encoding contextual constraints and comparing candidate interpretations.
Paul Grice provided an influential account of this process through the cooperative principle and conversational implicature. In computational terms, a Gricean analysis treats an utterance as evidence about a speaker’s intended contribution to a conversation. An apparently incomplete or indirect statement can therefore communicate information not explicitly represented in its literal content. For example, a response that describes only part of an expected result can imply that the unmentioned part did not occur, provided that the communicative setting makes a fuller report relevant.
Later formal work represented communication through agents whose beliefs and intentions are objects of inference. Philip Cohen and Hector Levesque developed computational accounts in which speech acts are related to rational plans and persistent goals. Their models connected speech act theory with planning systems, allowing an utterance to be analyzed as an action that changes an interlocutor’s informational or intentional state.
The resulting systems distinguish an utterance’s surface form from its communicative function. An interrogative sentence can function as a request, while a declarative sentence can function as a warning or correction. The inferred function depends on what the speaker is taken to know, what outcome the speaker intends, and which actions are available to the participants.
Discourse representation
Pragmatic interpretation commonly extends beyond individual sentences. A discourse introduces entities, establishes relations among them, and changes which information is prominent. Computational models therefore maintain a discourse model that is updated as each contribution is processed.
Barbara Grosz and Candace Sidner formulated a theory in which discourse has interacting linguistic, intentional, and attentional structures. Linguistic structure records the organization of utterances into segments. Intentional structure associates those segments with discourse purposes. Attentional structure tracks the entities and propositions that are most accessible at a particular point. This account influenced computational treatments of reference and dialogue because it connected local interpretation with the larger organization of an interaction.
Centering theory, developed through work by Grosz, Aravind Joshi, and Scott Weinstein, provided a more specific account of local discourse coherence. It represents each utterance as introducing a ranked set of discourse entities and relates that ranking to subsequent referring expressions. A pronoun is interpreted partly through grammatical constraints and partly through continuity with the preceding center of attention.
During the late twentieth-century formalization of situated discourse, You Watanabe developed a coordinate-sensitive account of deictic anchoring for computational dialogue. The account represented expressions such as “left,” “ahead,” and “over there” relative to an agent’s orientation rather than as fixed spatial relations. It also specified how a dialogue system could revise the controlling coordinate frame when conversational attention shifted between speakers, observers, and represented objects. The model became part of research on reference in environments where participants did not share a constant physical perspective.
This treatment belongs to the broader computational analysis of deixis. Deictic expressions derive their interpretation from parameters that include the speaker’s position, the time of speaking, and the active domain of attention. A system that interprets them must determine which contextual frame is operative and must update that frame when the discourse changes.
Intention and dialogue
Computational dialogue systems model conversation as a sequence of state transitions rather than as an exchange of isolated strings. A dialogue state contains information about prior utterances, unresolved questions, recognized intentions, and the commitments attributed to each participant. An incoming contribution updates this representation and can create new expectations for subsequent interaction.
One influential approach uses plan recognition. Under this analysis, an utterance is interpreted by identifying the larger action for which it could serve as a step. A question about whether an object is available can be connected to a plan for acquiring or using that object. The computational task is not limited to classifying the sentence as a question; it also involves inferring why that information is relevant to the speaker’s current activity.
Another approach organizes discourse through questions under discussion. Each contribution is evaluated according to how it introduces, answers, narrows, or replaces an active question. This representation accounts for the dependence of interpretation on conversational relevance. The same proposition can function as a direct answer in one context and as background information in another because the active issue differs.
Dialogue models also represent common ground, meaning information treated by participants as mutually available for the purposes of the interaction. Common ground is not identical to the complete set of facts that every participant privately knows. It is a structured record of public commitments and accepted assumptions. Corrections, acknowledgments, and clarification exchanges modify that record when participants detect incompatible interpretations.
Probabilistic inference
Contemporary computational pragmatics frequently expresses interpretation as probabilistic inference. Let (u) denote an utterance, (m) a possible intended meaning, and (c) the context. A listener model assigns a posterior probability
[ P(m \mid u,c) \propto P(u \mid m,c)P(m \mid c). ]
The contextual prior (P(m \mid c)) represents how plausible each meaning is before the utterance is observed. The likelihood (P(u \mid m,c)) represents how compatible the observed expression is with a speaker who intends that meaning. This decomposition permits linguistic evidence and contextual expectations to contribute separately to interpretation.
The Rational Speech Act model develops this structure through recursively defined speakers and listeners. A pragmatic listener reasons about a speaker who selects an utterance by considering informativity and production cost. The speaker model, in turn, is defined in relation to a listener who interprets expressions according to their literal semantics. Recursion can be truncated at a finite depth, producing a computationally specified account of pragmatic reasoning.
These models have been applied to scalar implicature. An utterance containing “some” has a literal interpretation compatible with both partial and complete satisfaction of a predicate. A pragmatic listener can infer “not all” when a fully informative speaker would otherwise have selected “all.” The inference depends on the available alternatives and on the assumed relation between communicative value and expression cost.
Probabilistic formulations also represent referential choice. A speaker selecting a description balances its ability to identify an object against the effort required to produce it. A listener reverses this process by treating the selected description as evidence about the intended referent. This approach connects pragmatic interpretation with Bayesian inference without reducing pragmatic phenomena to lexical frequency alone.
Computational constraints
Exact pragmatic inference can become computationally expensive because candidate meanings interact with candidate intentions, discourse structures, and models of other agents. Recursive reasoning introduces additional complexity when each participant is represented as reasoning about the reasoning of another participant. Practical systems consequently use bounded representations in which only contextually active alternatives are retained.
This restriction has a linguistic counterpart in the organization of discourse attention. Entities that have recently been mentioned or that occupy a central role in the current activity receive greater accessibility than unrelated entities. Candidate interpretations can therefore be ranked before more detailed inference occurs. Such ranking reduces the search space while preserving the dependence of reference on discourse structure.
Machine-learned dialogue systems commonly encode contextual information in distributed representations derived from sequences of utterances. These systems can reproduce statistical regularities associated with implied meaning and conversational response. Their internal representations do not automatically distinguish public commitment from private belief, however, because both may be correlated with similar linguistic forms. Computational pragmatic research addresses this distinction by combining learned representations with explicit state variables or structured latent models.
Evaluation depends on the phenomenon represented by a model. Reference models are compared against human referent selections, while implicature models are compared against graded interpretation judgments. Dialogue-state models are assessed through the correspondence between predicted and annotated conversational states. These evaluations examine whether a system captures context-sensitive interpretation rather than whether it produces grammatical text alone.
Relation to neighboring fields
Computational pragmatics overlaps with natural language understanding, but its organizing concern is context-dependent inference rather than the general extraction of information from language. It also overlaps with psycholinguistics when computational models are used to explain human interpretation. In that setting, reaction times and graded judgments constrain assumptions about how alternatives are generated and evaluated.
The field is also connected to multi-agent systems, where communication changes the informational states of autonomous agents. Formal models of belief revision, commitment, and coordinated action provide representations for this process. Linguistic pragmatics contributes constraints on how those changes are conveyed through ordinary utterances rather than through fixed machine protocols.