Centering: A Framework for Modeling the Local Coherence of Discourse

The local component of attentional state: relationships among focus of attention, choice of referring expression (pronoun vs full noun phrase) and perceived coherence of the utterances within a discourse segment.

Centering was presented in 1983 and elaborated in a draft that circulated from 1986 and was widely cited and extended but never published; this paper gives the definitive statement of the original motivations, definitions and claims.

Linguistic structure (discourse segments, possibly embedded), intentional structure (the discourse purpose and discourse segment purposes, and relations among them) and attentional state (the participants’ focus of attention). Centering models the local level of attentional state.

Local coherence holds among utterances within a segment, modelled by centering. Global coherence holds between segments and depends on intentional structure; Grosz and Sidner model global attentional state with a stack of focus spaces pushed and popped according to intentional relations.

Grosz’s (1977) global vs immediate focusing; Sidner’s (1979) analysis of immediate focus (discourse focus vs potential foci) with rules for resolving pronouns; and Joshi, Kuhn and Weinstein’s work linking changes in immediate focus to the complexity of inference, which introduced the term “centering”.

(1)	(2)
John went to his favorite music store to buy a piano.	John went to his favorite music store to buy a piano.
He had frequented the store for many years.	It was a store John had frequented for many years.
He was excited that he could finally buy a piano.	He was excited that he could finally buy a piano.
He arrived just as the store was closing for the day.	It was closing just as John arrived.

Both convey the same content, but (1) stays centered on John while (2) flips back and forth between John and the store.

The resources a hearer needs to extract information from a discourse because of the particular linguistic forms chosen. Centering conjectures that discourses with the same propositional content but higher inference load are perceived as less coherent.

Terry has been the center, so “he” is first taken as Terry; only at “sick” does it become clear that Tony is meant, forcing the hearer to backtrack. Using “Tony” instead of “he” avoids the misdirection.

The entities that link the utterance to other utterances in its discourse segment. Utterances (not sentences in isolation) have centers, and centers are semantic objects, not words or phrases.

Each utterance \(U_n\) has a partially ordered set of forward-looking centers \(C_f(U_n)\), ranked by prominence. Each non-initial utterance has a single backward-looking center \(C_b(U_{n+1})\): the highest-ranked element of \(C_f(U_n)\) that is realized in \(U_{n+1}\).

A phrase directly realizes a center \(c\) if \(c\) is its semantic interpretation; “realizes” generalizes this (e.g. \(c\) is an element of the situation described). Realization combines syntactic, semantic, discourse and intentional factors.

transition	\(C_b(U_{n+1}) = C_b(U_n)\)?	\(C_b(U_{n+1})\) highest in \(C_f(U_{n+1})\)?
continuation	yes	yes
retaining	yes	no
shifting	no	—

(Later work split shifting into smooth and rough shifts.)

To the extent a discourse adheres to centering constraints, its coherence increases and the inference load on the hearer decreases.

A unique \(C_b\) per utterance; ranking of \(C_f\) by several factors; centering constrains realization (Rule 1); preferences among transition sequences (Rule 2); primacy of partial information (a semantic theory must allow partial interpretations); locality of \(C_b\) (chosen only from \(C_f(U_{n-1})\)); and center determination by a combination of discourse factors, not syntax or semantics alone.

\(C_b(U_n)\) must come from \(C_f(U_{n-1})\), not from earlier utterances; if it seems to come from \(C_f(U_{n-2})\), that’s only because it is also in \(C_f(U_{n-1})\).

After “Susan gave Betsy a pet hamster. She reminded her that such hamsters were quite shy.”, continuations get worse from “She asked Betsy whether she liked the gift” to “She told Susan that she really liked the gift” (unacceptable to many). If Susan and Betsy were equal candidates all would be equally good; Susan must be the \(C_b\).

By grammatical role: subject > object(s) > other. Other factors (word order, fronting, subordination, lexical semantics) were left to further research.

It supported a single \(C_b\) and a determining role of grammatical role in identifying it, but not of thematic role or surface position; grammatical role and surface position both affect the \(C_f\) ordering. Kameyama first argued for grammatical rather than thematic role.

The \(C_b\) can be realized in other roles (here, a pronominalized direct object, the tugboat); a discourse where pragmatics forces “the silly thing” to mean the bear instead of the higher-ranked boat is less coherent.

If any element of \(C_f(U_n)\) is realized by a pronoun in \(U_{n+1}\), then \(C_b(U_{n+1})\) must be realized by a pronoun also. A pronoun for the \(C_b\) signals that the speaker is continuing to talk about the same thing; other pronouns are allowed as long as the \(C_b\) is pronominalized.

Sequences of continuation are preferred over sequences of retaining, which are preferred over sequences of shifting. Frequent shifting lacks local coherence; a coherent segment has shifts followed by runs of continuations.

Speakers and generators should plan ahead to minimize shifts; for interpretation it doesn’t give a direct procedure but predicts that some sequences impose higher inference load (it was not yet validated psycholinguistically).

Preferentially by a pronoun in English, and by equivalent forms such as zero pronouns in languages like Japanese and Italian.

It violates Rule 1: the \(C_b\) (John) is realized by a name while another entity (Mike) is a pronoun, so the only reading is a second person named John, and even that is odd.

utterance	\(C_b\)	\(C_f\)	transition
He called up Mike…	John	John, Mike	continue
Mike has annoyed him…	John	Mike, John	retain
He called John…	Mike	Mike, John	shift

Rule 1 holds throughout.

When no pronouns are used and the description conveys extra information: “My dog is getting quite obstreperous. I took him to the vet. The mangy old beast always hates these visits.” The hearer infers the \(C_b\) is unchanged and that the description holds of it.

The house is the \(C_b\) of the later utterances though only realized, not directly realized: the door and the furniture express functional relations to it (an implicitly focused entity). The door ranks above the house in \(C_f\), yet the segment stays “about” the house.

The information needed for a unique interpretation (especially of definite noun phrases) may only arrive in later utterances; computing all interpretations or backtracking are costly, so the authors conjecture partial interpretations refined later are the right choice.

A definite NP can contribute both a value-free interpretation (the role: “Historically, he is the President’s key man…”, “He is required to be 35 years old”) and a value-loaded one (the current holder: “As Ambassador to China, he…”); a later pronoun can pick up either, unless an earlier utterance forced one.

Speaker’s referential intentions matter: one NP can contribute two entities (the actual husband and the man the speaker meant) to \(C_f\), and the next utterance’s \(C_b\) can be either.

Sidner has a discourse focus, an actor focus and potential foci per utterance; centering has one \(C_b\). In “I haven’t seen Jeff for several days. Carl thinks he’s studying for his exams, but I think he went to the Cape with Linda”, Sidner’s actor focus wrongly prefers Carl for the last “he”; centering makes Jeff the \(C_b\).

Cross-linguistic studies (German, Hebrew, Italian, Japanese, Korean, Turkish), centering-based pronoun resolution algorithms (e.g. Brennan, Friedman and Pollard 1987), psycholinguistic and corpus evaluation, and the interaction of centering with dialogue and global discourse structure.