Is Centering Theory a particular neural network?

No. It is a linguistic framework that can inform computational models and analysis.

Can the Cb differ from the subject?

Yes; U3 in the example demonstrates this.

Does it explain global document structure?

Its central focus is local coherence within discourse segments, so broader structure requires additional accounts.

AI Glossary / Natural language and speech / Advanced

Centering Theory

Definition

Centering Theory is a framework for explaining local discourse coherence through the prominence of entities and the referring expressions used for them. It tracks how utterances connect to entities made salient by nearby utterances. The theory relates continuity and changes of attention to how easily a reader or listener can follow a passage; it is not a complete theory of every aspect of coherence.

What it means

A paragraph can repeatedly mention the same people yet still be difficult to follow if its pronouns and grammatical organization suggest unexpected changes in focus. Centering provides concepts for describing those changes rather than simply counting repeated names.
The centers are entities in the discourse interpretation, not the words themselves. Mira and she can realize the same center when they refer to the same person. Conversely, repeating she does not guarantee a stable center if the intended person changes. The relevant analysis depends on the utterance in its context, not just an isolated sentence.

How it works

An utterance has forward-looking centers, written Cf: entities ranked by their prominence in that utterance. A backward-looking center, Cb, connects the current utterance to the previous one. In the basic account, it is the highest-ranked center from the previous utterance that is realized in the current utterance. Ties or difficult cases require further analysis.
The highest-ranked current forward-looking center is often called the preferred center, Cp. It need not be the same as the current backward-looking center. A simplified English analysis may rank subjects above objects, but grammatical role is not a universal, context-free solution to every ranking problem.
A continuation keeps the previous Cb and makes it highest-ranked now. Retaining keeps that Cb while another entity ranks higher now. Shifting changes the Cb. These are the three broad transition types used in the original account; later formulations can distinguish further subtypes.

Three practical examples

A writing assistant highlights a paragraph where successive pronouns unexpectedly switch between two people. Its suggestion is to clarify the intended referent, rather than ban pronouns or topic changes.
A summarizer considers whether a reordered sentence makes a previously secondary entity appear central. It checks reference clarity alongside factual accuracy before accepting the reordering.
A dialogue researcher annotates utterances with entity rankings and centers to compare how speakers maintain or change local focus. The study documents its ranking assumptions so another annotator can reproduce the analysis.

Think of it this way

Imagine a conversation illuminated by a movable spotlight. Several people or objects remain on stage, but one is especially central to the current exchange. Centering distinguishes the entity that connects this moment to the previous one from the entity positioned to become prominent next. The spotlight is an analogy for a discourse model, not a literal measurement of attention.

Worked example

Use this original sequence. Stipulate the intended pronoun readings and a simplified subject-before-object ranking. Track only the two people, Mira and Noor; take each sentence as one utterance within one segment.
U1: “Mira called Noor.” Its Cf list is [Mira, Noor]. Leave the segment-initial Cb unset for this exercise.
U2: “She thanked Noor.” Here she means Mira. Its Cf is again [Mira, Noor], and its preferred center is Mira. Both people from U1 occur in U2, so the highest-ranked surviving entity from U1 is Mira: Cb(U2)=Mira. We do not classify U1-to-U2 by inventing an initial Cb.
U3: “Noor answered her.” Here her means Mira. Its Cf is [Noor, Mira], so Cp(U3)=Noor. However, Cb(U3)=Mira because Mira ranked highest in U2 and is still realized in U3. The backward-looking center stays Mira while the preferred center becomes Noor. U2-to-U3 is therefore a retaining transition.
U4: “Noor waved.” Its Cf contains only Noor. Of the entities ranked in U3, Noor is the highest one still realized. Thus Cb(U4)=Noor, different from Cb(U3)=Mira. U3-to-U4 is a shift in this classification.
For comparison, replace U3 with “Mira thanked Noor.” Its Cb and preferred center are both Mira. Relative to U2, that alternative is a continuation. This comparison isolates the transition definitions; it does not assert that the alternative tells the same story or is always a better sentence.
Notice especially that the original U3 has Noor as grammatical subject but Mira as Cb. Equating backward-looking center with current subject would misclassify the retaining transition.

Illustrative case study

Fictional example: BriefFlow generates short project updates about two colleagues. Its first editing rule replaces repeated names with pronouns whenever possible. Reviewers find that some passages become harder to follow when responsibility moves from one colleague to the other.
The team adds a review of entity continuity and checks whether an explicit name would signal a change more clearly. It does not automatically rewrite every shift, because project updates legitimately move between people and tasks. The fictional case illustrates how centering can inform a clarity review; it does not report a validated score or prove that the framework alone resolves every pronoun.

Common misconceptions and mistakes

A center is not simply the most frequent word, the current subject, or the overall document topic. Local entity continuity is only one aspect of coherence: causal connections, goals, and broader discourse organization also matter.
A shift is not automatically a writing error, and continuing one entity forever is not a sensible universal objective. Ranking rules also need linguistic and annotation context. Finally, discourse attention in Centering Theory should not be equated with numerical attention weights inside a Transformer.

Business use

Centering concepts can support evaluations of generated summaries, handover notes, and conversational responses where readers must track who did what. Specify the utterance boundaries, entity links, and ranking conventions before computing transitions. Use resulting patterns to investigate clarity problems alongside reader judgments and factual checks, rather than treating a transition count as a complete quality metric.

Check your understanding

Why is U2-to-U3 retaining?

Answer: Mira remains the Cb, but Noor becomes highest-ranked in U3.

Why does U4 shift?

Answer: Its Cb is Noor, whereas U3's Cb is Mira.

What assumption made the rankings explicit?

Answer: Subjects were ranked above objects for this simplified example.

Related terms

Anaphora
Coreference resolution
Entity grid
Natural language generation
Discourse coherence
salience
referring expression
discourse segment

Learn next

Annotate a short passage's entity mentions before ranking its forward-looking centers. Then compute backward-looking centers from the preceding utterance and distinguish continuation, retaining, and shifting. Compare alternative annotation conventions before interpreting transition frequencies as evidence about readability.

Forward-looking centers

Sources

Grosz, Weinstein, and Joshi, Centering: A Framework for Modeling the Local Coherence of Discourse, especially sections 3–6: https://aclanthology.org/J95-2003.pdf

Previous: Centering Next: Centerpoint