AI Glossary / Natural language and speech / Intermediate

Discourse-new

Definition

Discourse-new describes a referent that is being introduced into the current discourse rather than already being established in it. The distinction concerns information status within a text or conversation. Something can be discourse-new even when the listener already knows about it outside that conversation, so discourse-new does not necessarily mean unfamiliar.

What it means

If a conversation first mentions a particular museum, that museum is new to the discourse even if both participants have visited it. By contrast, a newly introduced unnamed visitor might also be new to the listener.
These are different dimensions: discourse status concerns what has been introduced in the current exchange; hearer status concerns what the listener is presumed to know. Keeping them separate prevents a system from equating first mention with real-world novelty.
The relevant unit is the referent, not merely the spelling of a word. A new phrase may describe an already established person, while a repeated common noun may introduce a different person or object.

How it works

A discourse model tracks the entities introduced as the text unfolds. When a referring expression appears, analysis asks whether it accesses an existing referent or introduces a new one.
Coreference systems can use anaphoricity decisions to help distinguish mentions that link backward from mentions that begin a new chain. The exact relationship between these labels depends on the dataset and model; a missing antecedent is not always equivalent to a straightforward new entity.
Inference complicates the picture. Mentioning a building can make its entrance inferable without explicitly introducing the entrance earlier. Information-status schemes often distinguish such bridging or inferable cases rather than reducing every reference to a simple binary choice.

Three practical examples

  • A meeting assistant notices the first mention of a project and introduces its name clearly in the notes, even if the participants already know it.
  • A coreference annotation task distinguishes a newly mentioned customer from another customer described earlier in the document.
  • A writing exercise checks whether a first reference supplies enough context for the intended audience to identify what is being discussed.

Think of it this way

Think of adding a new card to a conversation's working index. The subject on the card may already be familiar to everyone, but it has just entered this particular index. The analogy explains discourse-relative newness; real understanding also uses shared knowledge, inference, and context beyond an explicit list.

Worked example

Consider this invented conversation. A: “I visited the Riverside Museum yesterday.” B: “The museum has a new exhibition.” A: “A guide explained it to me. The guide also recommended another gallery.”
At the first sentence, Riverside Museum is introduced into this discourse. Assume both speakers already know the museum: it is discourse-new but hearer-old. B's “the museum” then accesses the same referent and is discourse-old.
In A's next turn, “a guide” introduces a person not previously mentioned. Under the stated context, this person is discourse-new and may also be hearer-new. “The guide” in the following sentence refers back to that person.
The phrase “another gallery” introduces an additional referent even though the conversation already concerns a museum. A system should not merge it with the museum simply because both are cultural venues. The analysis tracks identity and introduction, not just topical similarity or repeated vocabulary.

Illustrative case study

Fictional case: ContextLedger creates entity lists from customer-service conversations. Its first rule treats every familiar organization name as already established, using a background directory to decide information status. Reviewers find that first mentions disappear from the conversation's introduction record.
The team separates two questions: whether an entity is known in the directory and whether it has already been introduced in this conversation. It also adds examples where a new description refers to an existing person and where a repeated noun introduces a different person.
The revised specification makes the intended distinction explicit. This fictional case illustrates modeling choices, not a measured improvement or a claim that a simple entity list captures all discourse interpretation.

Common misconceptions and mistakes

Discourse-new does not mean recently created, recently discovered, or unknown to the world. It is relative to the discourse being analyzed.
Do not infer newness solely from an indefinite article, or oldness solely from a definite expression. Linguistic form is evidence, not an infallible identity rule.
Avoid confusing a new mention with a new referent. Different expressions can identify the same entity, and similar expressions can identify different entities. Inferable referents may require their own annotation category.

Business use

This distinction helps teams design clearer summaries, introductions, and entity-tracking systems. A summary intended for readers outside a meeting may need to explain a referent that was already familiar to participants, while still accurately reflecting when it entered the conversation.
For evaluation, specify the discourse boundary: a single message, a complete conversation, or a collection of documents. The same mention can have different status under different boundaries. Record assumptions about audience knowledge separately so they do not silently distort the entity model.

People also ask

Can a famous place be discourse-new?

Yes. Its first introduction into the current discourse can be new even when the listener recognizes it immediately.

Does every new noun phrase introduce a new entity?

No. A new expression can refer to an entity already established in the discourse.

Is a referent inferred from an earlier mention simply discourse-old?

Not necessarily. Many annotation schemes distinguish inferable or bridging references from explicitly evoked old referents.

Check your understanding

Two colleagues know the Atlas project well, but one mentions it for the first time in today's conversation. Is Atlas necessarily discourse-old because both colleagues know it?

Answer: No. It can be discourse-new in this conversation while being hearer-old for both colleagues. The two kinds of status answer different questions.

Related terms

  • Discourse-old
  • Discourse model
  • information status
  • hearer-new
  • hearer-old
  • coreference
  • anaphoricity
  • bridging reference

Learn next

Annotate first mentions and later references in a short conversation. Add a separate column for assumed audience familiarity, then inspect cases where the two columns differ. Finally, identify referents that are inferred rather than explicitly introduced.

More in Natural language and speech

  • Discourse deixis - Discourse deixis is reference to some aspect of a stretch of discourse, often using expressions such as “this” or “that.
  • Discourse parsing - Discourse parsing is the computational analysis of relationships and structure across parts of a text.
  • Discourse relation - Definition: A discourse relation is a meaning-based connection between stretches of text, such as explanation, contrast, elaboration or temporal sequence.
  • Discovery procedure - In the linguistic context of this glossary, a discovery procedure is a systematic method intended to derive a description of a language's structure from...
  • Disfluency - A disfluency is an interruption, hesitation, repetition, or revision in the flow of speech.
  • Disjunction - Disjunction is the logical operation commonly expressed as “or.

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, Chapter 24, Section 24.1.2 and anaphoricity passages consulted directly: https://web.stanford.edu/~jurafsky/slp3/24.pdf. Discourse versus hearer status, inferables and missing-antecedent distinctions were checked. The museum-and-guide example is original.

Previous: Discourse model Next: Discourse-old

  • Home
  • Daily AI Updates
  • Blog
  • AIFBL Academy
  • AI Glossary
  • Products and Tools
  • Store
  • Free CFO AI Prompt Pack
  • About Shihan Sheriff
  • Privacy Policy
  • Terms of Use
  • Refund Policy
  • Cookie Policy
  • Earnings Disclaimer