How does pragmatics differ from semantics?

Semantics concerns linguistic meaning, while pragmatics emphasizes language use and contextual interpretation. They interact, and theories differ on exactly where to draw the boundary.

Is an implicature always true?

No. An implicature is an inferred communicated meaning, and conversational implicatures can depend on assumptions that turn out to be wrong.

Does a fluent AI response demonstrate perfect pragmatic understanding?

No. Fluency does not establish that a model has correctly identified the speaker's goal, reference or shared context in a particular exchange.

AI Glossary / Retrieval and knowledge systems / Beginner

Pragmatics

Definition

Pragmatics is the study of how people use language in context and how listeners infer what speakers intend to communicate. It concerns meanings that depend on the situation, shared knowledge and conversational goals, including indirect requests, implied messages and the interpretation of expressions such as "here" or "that one".

What it means

Words and grammar provide essential information, but they do not always specify the whole message. If someone says "the room is cold" beside an open window, they may be describing the temperature, requesting that the window be closed, or explaining why they want to leave.

Pragmatics examines how context supports these interpretations. It interacts closely with semantics, which studies linguistic meaning, and the boundary between the two is debated. The practical point for AI is that a plausible interpretation is an inference from evidence, not direct access to a person's private intention.

How it works

An interpretation begins with the utterance and its linguistic content. The listener considers who is speaking, what has already been discussed, what the participants can observe and what goal the conversation appears to serve. This can help resolve references, identify a communicative action and infer an unstated implication.

For example, a question about ability may function as a request in an appropriate setting. A statement can also imply something stronger than its literal content when the speaker appears knowledgeable and has chosen a less informative expression.

These inferences can change when context changes. Computational approaches range from dialogue-state and intent models to formal models of speakers and listeners. No single method captures every pragmatic phenomenon. Where several interpretations remain plausible, a useful system can preserve uncertainty or ask a focused clarification.

Three practical examples

1. A support assistant interprets "I still cannot sign in" as a continuation of an unresolved access problem, using the preceding exchange rather than treating it as an isolated statement.

2. A scheduling assistant resolves "move that meeting" by checking the meeting currently under discussion, instead of selecting an arbitrary calendar event.

3. A writing assistant helps distinguish a polite indirect request from a factual question, while allowing for different cultural and conversational conventions.

Think of it this way

Think of language as a message accompanied by a situation. The same note saying "the door is open" can be helpful directions, a warning or a request, depending on where it appears and what the people involved are trying to do. Pragmatics studies how that surrounding evidence shapes interpretation.

Worked example

Consider an original workplace exchange. A coordinator asks, "Have all eight draft pages been reviewed?" An editor replies, "Some have been reviewed."

For this example, use an inclusive literal interpretation of "some": at least one. On that interpretation, the literal statement is compatible with any reviewed count from one through eight. If the editor knows the status of every page and is trying to provide an informative answer, the coordinator may infer that fewer than all eight have been reviewed.

Now add the context that the editor has only checked the first two pages and cannot see the rest. The same reply no longer strongly supports the inference that some pages remain unreviewed; the editor may simply lack complete information. The arithmetic range distinguishes literal compatibility from the stronger contextual inference. Neither reply identifies an exact reviewed count, so an automated tracker should not invent one.

Illustrative case study

Fictional case study: Harbour Desk deploys an internal assistant for meeting-room support. A colleague writes, "Can you make this room less bright?" The assistant initially responds only that adjustable lighting is available.

The team reviews the interaction as an indirect request, but also notices that the assistant lacks a reliable room reference. It changes the response design to identify the intended room and the available action before proceeding under the user's permissions. Evaluation cases include factual questions about lighting as well as requests to change it.

This fictional case illustrates two distinct requirements: interpreting the likely conversational purpose and resolving the information needed for action. Recognizing an indirect request does not remove the need to identify its target correctly.

Common misconceptions and mistakes

Pragmatics is not simply sentiment analysis or detecting sarcasm. It includes many ordinary acts of reference and inference in neutral conversation. Nor is an inferred implication always a logical consequence of the words; additional context can weaken or cancel it.

Avoid treating one communication style as universal. Politeness, indirectness and shared assumptions vary across people and situations. Another mistake is using a plausible implied intention to invent facts or authorization. A system should distinguish what was explicitly stated, what was inferred and what still needs clarification.

Business use

Pragmatic competence can make business interfaces more useful by helping them handle follow-up questions, indirect requests and references to shared context. It can also improve evaluation: a response may be grammatically correct yet fail to address what the user was trying to accomplish.

Test conversations rather than only isolated prompts. Include changes of topic, ambiguous references and cases where the same sentence serves different purposes. Review the consequences of over-interpreting as well as under-interpreting a request. For consequential actions, the system should maintain a clear link between the user's expressed intent, the selected target and the action actually taken.

Check your understanding

A speaker says "some reports are finished" but has only inspected two reports. Can a listener conclude with certainty that the remaining reports are unfinished?

Answer: No. Limited knowledge weakens that inference. The statement supplies no complete account of the uninspected reports, so their status remains unresolved.

Related terms

Semantics
Speech act
Common ground
conversational implicature
reference resolution
deixis
dialogue state and intent recognition

Learn next

write two contexts in which the same sentence would reasonably communicate different intentions. Mark which information is explicitly stated and which is inferred. Then design a clarification that resolves the uncertainty without presuming the answer.

Sources

Primary sources: Potts, Semantics-Pragmatics Interaction: https://web.stanford.edu/~cgpotts/entries/potts-cup-semantics-pragmatics.html

Goodman and Stuhlmüller, Knowledge and implicature: Modeling language understanding as social cognition: https://web.stanford.edu/~ngoodman/papers/KnowledgeImplicature-v1.pdf

Jurafsky and Martin, Speech and Language Processing, author-hosted manuscript: https://web.stanford.edu/~jurafsky/slp3/

Previous: Pragmatic mapping Next: Pre-processing algorithm