📄 Providing a Unified Account of Definite Noun Phrases in Discourse

Summary

The paper addresses a long-standing confusion about whether the behavior of definite noun phrases and pronouns in discourse is a matter of syntax, semantics, or pragmatics. Grosz, Joshi and Weinstein argue that no single category suffices. They propose a framework built on two levels of discourse coherence. Global coherence links larger segments and is tied to global focusing. Local coherence binds neighboring utterances and is tied to centering. Each sentence gets one backward-looking center, Cb(S), which anchors it to prior discourse, and a set of forward-looking centers, Cf(S), which offers candidates for what follows. Through constructed examples they show three things. The center cannot be read off syntax alone. Centers may be value-free or value-loaded interpretations of descriptions, following Barwise and Perry. A "realization" relation, distinct from denotation, connects expressions to centers. Pronouns mainly serve local centering, while full definite descriptions mainly serve global focus. Mismatches between the two force extra inference by the hearer. This matters especially for language generation, where ignoring these links yields unnatural referring expressions.

Key points

• Two coherence levels map to two focusing processes: global focusing (whole-discourse entities, including implicitly focused ones) and centering (the single entity an utterance is mostly about).
• Identical sentences can have different centers depending on context. "Max saw Rosa" has Cb = Max after a question about whom Max saw, but Cb = Rosa after a question about who saw Rosa. Examples are drawn from Reinhart [1982].
• When a full NP such as "the vice president of the United States" appears, Cf receives two items from it: its value-free interpretation and its value-loading (George Bush). A later "he" can pick up either one.
• Their centering rule allows entities other than Cb(S) to be pronouns, provided Cb(S) is also pronominalized. This handles Sidner's example D4 (Jeff/Carl) without her extra "actor focus."
• Using a full name for the current center ("He was annoyed by John's call," where John is the center) sounds wrong unless intervening sentences have shifted the center.

Interesting ideas

• I1. Separate local and global salience. Track a local "current topic" for pronoun handling and a broader focus set for full descriptions. Coreference or dialogue systems could model these as two distinct structures rather than one salience score.
• I2. Use realization instead of plain denotation. Linking a mention to an entity through a looser relation lets one expression evoke a role, its current holder, or a related object. This could help entity-linking schemes deal with role-type mentions, such as officials or titles.
• I3. Treat pronominalization as a generation constraint. The rule that the center must be pronominalized whenever anything else is gives a simple, testable check. A text generator could use it to avoid awkward full names for an entity already in focus.
• I4. Read processing cost from form–level mismatches. A pronoun pointing to a globally focused entity, or a full NP pointing to the current center, signals extra inference for the hearer. This could serve as a coherence or readability diagnostic when evaluating generated or translated text.
• I5. Use minimal-pair discourses as probes. Keep one sentence fixed and vary only the preceding context to see how the center shifts. This is a cheap way to build diagnostic test sets for discourse models in any language, Ukrainian included.