📄 A Centering Approach to Pronouns

Summary

The paper addresses how to resolve personal pronouns across sentences using local discourse coherence. It builds on the centering framework of Grosz, Joshi and Weinstein [GJW86] and recasts its informal ideas as three explicit constraints and two rules. These operate over three notions: forward-looking centers (Cf), the preferred center (Cp) and the backward-looking center (Cb). The Cf list is ordered by grammatical obliqueness, so subjects come first.

The authors show that the original theory cannot choose between readings when an utterance contains several ambiguous pronouns and a shift is forced. They therefore split shifting into two kinds: "shifting-1", where the new Cb is also the Cp, and ordinary shifting, which is less coherent.

On this basis they define an algorithm that generates, filters and ranks candidate Cb–Cf pairings, then selects the most preferred transition. It runs inside the pragmatics processor of an HPSG natural language system that serves as a front end to a database query application. The paper evaluates the approach through worked examples, not a corpus. Its value lies in offering an explicit, implementable centering procedure for pronoun binding.

Key points

• Constraints: each utterance has exactly one Cb. Every Cf element must be realized in the utterance. The Cb is the highest-ranked element of the previous Cf that is realized in the current utterance.
• Rules: if any previous-Cf entity is pronominalized, the Cb must be pronominalized as well. The transition preference order is continuing > retaining > shifting-1 > shifting, where shifting-1 is the paper's extension.
• Cf ranking follows the main verb's subcategorization: subject, object, object2, then other subcategorized functions, then adjuncts.
• The algorithm has three phases:
  • It constructs proposed anchors (⟨Cb, Cf⟩ pairs), including a NIL Cb option.
  • It filters them by contraindexing, constraint 3 and rule 1.
  • It classifies and ranks the survivors.
  • In the "She often beats her" example, 16 anchors shrink to 2, and shifting-1 correctly binds "she" to Friedman.
• The authors argue that Kameyama's property-sharing (SUBJ/IDENT) parallelism constraint is largely unnecessary. In their view, Cf ordering combined with the continuing/retaining distinction already produces parallel readings.

Interesting ideas

• I1. Generate, filter, then rank. Enumerate all candidate interpretations, prune them with independent hard filters (which can run in parallel), and order the rest by soft preferences. This keeps the design modular, so new constraints can be added easily in any disambiguation pipeline.
• I2. Rank whole transitions, not single antecedents. Scoring how coherently the discourse moves from one utterance to the next outperforms simple recency or grammatical-parallelism heuristics. This view could inform modern coreference features or reranking.
• I3. Include a NIL candidate. Explicitly allowing an utterance to have no backward center lets the system represent "no link" without special-case logic. This is useful wherever an empty or null decision is legitimate.
• I4. Different preferences for generation and understanding. A generator could deliberately use a retention to signal an upcoming shift. An understander may instead need to flag genuine ambiguity rather than force a choice.
• I5. Test the ranking cross-linguistically. The authors suggest checking whether grammatical function or word order drives Cf ranking in freer-order languages such as German, and whether topic outranks subject in Japanese. The same check applies directly to morphologically rich, free-order languages like Ukrainian.