AI Glossary / Natural language and speech / Intermediate

Discourse

Definition

Discourse is connected language interpreted as a structured whole, such as a conversation, report, story, or article. In natural language processing, discourse analysis examines how parts of that whole relate to one another and to their context. Understanding individual sentences is often insufficient to understand the complete message.

What it means

A report can introduce a problem, explain its cause, compare solutions, and recommend an action. Each sentence contributes to a larger organization. Pronouns, repeated entities, topic development, and relationships such as explanation or contrast help readers connect the pieces.
The same principle applies to dialogue. An answer may depend on a question several turns earlier, especially when clarification interrupts the exchange.
Discourse therefore concerns more than grammatical correctness. A sequence of well-formed sentences can still be confusing if references are unclear, ideas jump unpredictably, or the relationships between claims are missing. Conversely, a short conversational fragment can make sense when its shared context supplies the missing information.

How it works

A computational system can represent several aspects of discourse. Reference resolution links expressions to the people, objects, or events they refer to. Relation analysis identifies connections between spans, while topic and entity tracking help model what the text is about.
The analysis may operate locally, linking neighboring clauses, or globally, considering a document's larger organization. Different frameworks represent these connections differently; some use trees, while others annotate relations between span pairs.
Practical systems need not explicitly produce every possible discourse annotation. However, evaluations should still test whether their outputs preserve important connections. A summary that lists facts but reverses which event caused which can fail despite containing many correct words.

Three practical examples

A summarizer preserves the link between a service interruption, its stated cause, and the action taken to restore service.
A conversational assistant connects “the second option” to an earlier list, even when a clarification turn intervenes.
A document-review tool checks whether a recommendation follows from the evidence presented and whether repeated references remain understandable.

Think of it this way

Think of individual sentences as scenes in a film. Each scene may be understandable alone, but their order and connections shape the overall story. The analogy captures structure beyond isolated units; discourse is not limited to storytelling and can include debate, instructions, explanations, or everyday coordination.

Worked example

Consider this invented three-sentence incident note: “The upload service stopped accepting files. Its storage volume was full. After the team expanded the volume, uploads resumed.”
At the sentence level, a system can identify a service, storage, an intervention, and resumed activity. At the discourse level, it must connect “its” to the service and interpret the second sentence as explaining the failure described first.
The final sentence establishes an intervention followed by recovery. A careful summary might say, “Uploads resumed after the team expanded storage that had become full.” The source supports that sequence and its stated explanation, but the summary should not invent a root-cause investigation or a quantified performance improvement.
Now reorder the text as “Uploads resumed. Its storage volume was full. The upload service stopped accepting files.” Each sentence remains grammatical, but the timeline and the reference become harder to interpret. This comparison shows why preserving sentence-level facts does not guarantee a coherent overall account.

Illustrative case study

Fictional case: BriefBridge creates short summaries of internal project updates. An early prototype extracts one sentence from each paragraph and presents them in order of keyword importance. Reviewers find that a proposed solution appears before the problem and that “this decision” has no clear antecedent.
The team adds review criteria for reference clarity, explanatory links, and the relationship between conclusions and supporting details. It compares complete summaries with their source passages rather than checking isolated extracted sentences alone.
The revised process aims to preserve the document's meaning as a whole. This fictional case describes an evaluation approach; it does not claim that a particular architecture or score guarantees coherent summaries.

Common misconceptions and mistakes

Discourse is not simply a longer sentence or a bag of related keywords. Relationships and context matter.
Do not equate coherence with truth. A passage can tell a clear, internally connected story while making false claims.
Avoid assuming that adding connective words automatically repairs an argument. “Therefore” signals an intended relationship, but the evidence may not justify the conclusion. Likewise, a text can express a relationship without an explicit connective.

Business use

For business-facing AI, discourse matters whenever an output must communicate more than isolated facts: reports, meeting summaries, explanations, support conversations, and instructions. It helps teams specify failures such as losing a qualification, attaching a reason to the wrong outcome, or leaving a pronoun without a clear referent.
A useful review combines factual checks with questions about organization: Can a reader follow the entities, timeline, reasons, and decisions? These dimensions are related but should not be collapsed into one vague judgment of writing quality.

Check your understanding

A summary contains individually accurate sentences but makes an action appear to happen before the problem it addressed. Has preserving the sentences preserved the complete meaning?

Answer: No. The discourse-level temporal relationship has changed, so the overall account can mislead even when individual statements resemble the source.

Related terms

Learn next

Annotate a short report by marking its main entities and connecting each explanation, contrast, and consequence to the relevant span. Then compare a faithful summary with a reordered version to see which relationships disappear.

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, Discourse Coherence: https://web.stanford.edu/~jurafsky/slp3/25.pdf ; Conversation and its Structure: https://web.stanford.edu/~jurafsky/slp3/26.pdf

Previous: Discounted state-action occupancy measure Next: Discourse connectives