Centering: A Framework for Modeling the Local Coherence of Discourse
What it is about
The paper gives the definitive statement of centering, a model of the local attentional state inside one discourse segment. Each utterance has a ranked set of forward-looking centers (Cf), mostly ordered by grammatical role, and exactly one backward-looking center (Cb), which is the highest-ranked center of the previous utterance that the current one realizes. Two rules follow: if any center is a pronoun, the Cb must be a pronoun too (Rule 1), and transitions are preferred in the order continue, retain, shift (Rule 2). The paper is a theory, not an evaluation. Its examples (John and Mike, Susan and Betsy, the Vice-President) show that breaking Rule 1 or running many shifts lowers coherence and raises the hearer's inference load, and the authors say the rules had no psycholinguistic test yet (section 6).
Ideas
- I1. Use the Cb as a tie-breaker in the coref sieve. Per sentence, rank the mentions by UD role (nsubj above obj above other) and prefer the previous sentence's Cb as the antecedent of a pronoun. Each link gets the reason "Cb continued", so
explainshows it. Measure with the coref eval on GUM, two-foldcv, and two identical runs before any claim. - I2. Rule 1 as a warning, not a rule. When a pronoun points to a non-Cb entity while the Cb is a name, emit a flag in the explain output. Count how often the flag fires on correct links and on wrong ones. It is a precision check on the sieve, and it changes no link.
- I3. Transition labels (continue, retain, shift) per paragraph as a coherence diagnostic in the world's tale readers. Negative control: the same paragraphs with shuffled sentence order must show more shifts. If they do not, the labels measure nothing.
- I4. Value-free and value-loaded readings of role nouns ("the Vice-President") in
onto: a role concept and its current holder are separate nodes, and a pronoun links to one of them with a reason that says which. Write our own test sentences; do not copy the paper's examples.
For Mova
The chapter is languages/en/35-reference-and-coreference.md. The code is coref/src/sieve.rs and coref/src/mention.rs, the tale readers in world, and the role nouns in onto. The measure is the coref eval on GUM; check that corpus's licence in licence/DATA.md before any release use. The paper is NC-SA: keep only these ideas in our words, and no text in the repo.
Library summary
Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/J95-2003.pdf
Summary
This paper is the archival version of centering theory. The authors first outlined it in 1983, and an unpublished 1986 manuscript of it circulated widely for years. The theory deals with local coherence: why some sequences of utterances within one discourse segment read more smoothly than others that convey exactly the same content. Centering is placed inside the Grosz and Sidner (1986) model of discourse structure, where it models the local part of attentional state. The work is purely theoretical. There is no corpus and no experiment, only constructed English examples. Each utterance gets a ranked set of forward-looking centers (Cf) and exactly one backward-looking center (Cb), which is the entity linking it to the previous utterance. Two rules connect these centers to the choice of referring expression and to how attention moves between utterances. The main claim is that differences in coherence come partly from the inference effort that a given type of referring expression demands in the current attentional state. The framework later became a basis for anaphora resolution algorithms, for measuring coherence, and for discourse research across many languages.
Key points
- Core constructs: Cf elements are ranked mainly by grammatical role, with the subject above the object and the object above other roles. The top-ranked element is the preferred center (Cp). The Cb is the highest-ranked element of the previous utterance's Cf that also appears in the current utterance.
- Two rules:
- Rule 1: if any entity from the previous Cf is pronominalized, then the Cb must be pronominalized as well.
- Rule 2: transitions are ranked by preference: continuing is preferred over retaining, and retaining over shifting.
- Motivating example: two versions of the same story about John and a music store. The version that keeps John in focus is judged more coherent than the one that switches back and forth between the store and John.
- Comparison with Sidner (1979): centering replaces her separate actor focus and discourse focus with a single Cb (the Jeff/Carl example). The authors admit the rules do not say which pronoun refers to the Cb when an utterance has several pronouns.
- Follow-on work surveyed:
- cross-linguistic studies of German, Hebrew, Italian, Japanese, Korean and Turkish
- algorithms such as Brennan, Friedman and Pollard (1987)
- psycholinguistic tests such as Gordon, Grosz and Gilliom (1993)
Cite
@article{grosz1995centering,
title = {{Centering: A Framework for Modeling the Local Coherence of Discourse}},
author = {Grosz, Barbara J. and Joshi, Aravind K. and Weinstein, Scott},
journal = {Computational Linguistics},
year = {1995},
doi = {10.21236/ada324949},
url = {https://doi.org/10.21236/ada324949}
}