Is a discourse model the same as a language model?

No. A language model may support contextual interpretation, while a discourse model describes the representation of the ongoing discourse.

Must every pronoun have one obvious referent?

No. Context can leave several plausible interpretations, and some expressions do not refer to entities in the usual way.

Does an entity need a proper name?

No. Descriptions, pronouns and other expressions can introduce or access entities.

AI Glossary / Statistics and probability / Intermediate

Discourse model

Definition

A discourse model is a representation of the entities and relevant relationships established during an unfolding conversation or text. It helps a language-processing system interpret later references in light of earlier statements. For example, a pronoun may refer back to an entity already represented in the model. The term describes a contextual representation, not a single universally specified algorithm or necessarily a separate neural network.

What it means

Sentences rarely stand alone. Readers connect names, descriptions and pronouns across a passage, distinguish newly introduced entities from familiar ones, and update their understanding when later information arrives. A discourse model provides a way to describe or implement that evolving context. In the source's entity-focused sense, it contains representations of what has been mentioned and the relationships in which those entities participate. Broader systems may also represent events, speaker perspectives or discourse structure, but those additions depend on the framework rather than being one mandatory data format.

How it works

A system identifies referring expressions, introduces representations for new entities and links later mentions to existing representations when the evidence supports that interpretation. It tracks relevant attributes and relationships, potentially including which entities are currently salient. Reference resolution draws on syntax, semantics, previous context and sometimes world knowledge. Ambiguous references should remain uncertain or prompt clarification rather than forcing an unsupported link. A discourse model can be explicitly structured or represented through other computational mechanisms. Updating it requires care with negation, hypothetical statements, quotations and corrections: a statement in the conversation is not automatically a verified fact about the external world.

Three practical examples

Customer assistance: connect a later phrase such as the replacement to the particular item already discussed, while distinguishing it from the original order.
Document analysis: group repeated descriptions of the same organization across sentences to support a coherent summary.
Dialogue systems: maintain references to selected objects so that a follow-up instruction can be interpreted in context.

Think of it this way

Think of keeping a conversation notebook with separate entries for the people and objects under discussion. New mentions add entries; later references point back to them. The notebook helps you follow the discussion, but an incorrect link or an unverified statement can still make the resulting interpretation wrong.

Worked example

Consider this invented passage: Mira collected a parcel from a locker. She carried it home. The parcel was damaged. Under the intended reading, six referring expressions map to three entities: Mira and She refer to one person; a parcel, it and The parcel refer to one object; and a locker introduces a second object. The model can represent that Mira collected the parcel from the locker, carried the parcel home, and that the parcel was described as damaged. Linking it to the locker would produce the wrong carrying relation despite the word being grammatically compatible with an inanimate object. The passage also does not establish who damaged the parcel or when that happened. Those would be additional inferences requiring evidence, not facts supplied by the entity links.

Illustrative case study

Fictional case: Pine Helpdesk prototypes an assistant that handles conversations about multiple deliveries. A customer mentions an original package and a replacement, then asks whether it can be redirected. The system identifies two plausible referents and asks which package the customer means before initiating a workflow. Engineers review the entity links and ensure that a later correction updates the intended package. This invented case illustrates the value of tracking references and ambiguity. It does not claim that any particular implementation resolves every pronoun correctly or that a discourse representation alone authorizes an action.

Common misconceptions and mistakes

Do not equate identical words with identical entities: two mentions of a manager may concern different people. Different words can refer to the same entity. A discourse model is not simply a transcript, although it is derived from discourse, and it is not automatically a permanent memory of a user. Coreference links are one component rather than the entirety of contextual understanding. Inferring a related entity, such as a building's entrance, differs from saying that the building and entrance are the same entity. Avoid turning speaker claims, hypothetical objects or quoted statements into unqualified real-world facts.

Business use

Businesses encounter discourse models in conversational assistance, document extraction and summarization. Better reference handling can support more coherent interactions, but errors can propagate into summaries or actions if unresolved links are treated as certain. Evaluation should include conversations with multiple similar objects, corrections, long gaps and ambiguous pronouns. Keep source mentions available for review, and separate interpreting a user's reference from verifying an external fact or obtaining permission for an action. The representation should serve the application's actual context needs rather than collecting unnecessary personal information.

Check your understanding

Why should the worked example not record that the locker damaged the parcel?

Answer: The text introduces the locker and states that the parcel was damaged, but supplies no causal link between them. A discourse model should preserve what the passage supports and distinguish any further conjecture from an established relation.

Related terms

Coreference resolution
Anaphora
Bridging inference
Entity linking
Natural language understanding
referring expression
discourse entity
salience
dialogue context
discourse coherence
dialogue state

Learn next

mark referring expressions in a short passage and group those that refer to the same entity. Introduce a second similar object and examine how ambiguity changes. Then distinguish coreference from relations between different entities, and study how negation, reported speech and corrections affect contextual representations.

Sources

Jurafsky and Martin, Speech and Language Processing, reference and discourse model discussion: https://web.stanford.edu/~jurafsky/18beg.pdf Jurafsky and Martin, Coreference Resolution, bridging inference: https://web.stanford.edu/~jurafsky/slp3/old_sep21/21.pdf Scoring Coreference Chains with Split-Antecedent Anaphors, ACL Anthology: https://aclanthology.org/2023.dnd-14.3/

Previous: Discourse deixis Next: Discourse-new