Providing a Unified Account of Definite Noun Phrases in Discourse

Barbara J. Grosz, Aravind K. Joshi, Scott WeinsteinACL 1983 · DOI

Introduction

In the early 1980s, linguistic theory was largely fragmented into distinct silos: syntax (the structure of sentences), semantics (literal meaning), and pragmatics (contextual intent). In their foundational 1983 paper, Barbara Grosz, Aravind Joshi, and Scott Weinstein proposed a framework to bridge these gaps. This framework, known as Centering Theory, provides a unified account of how discourse remains coherent and how listeners track entities as a conversation unfolds. By focusing on the relationship between the structure of an individual sentence and the broader discourse context, the authors established a model for understanding the cognitive load associated with different referring expressions.

The Problem: Local Coherence and Referring Expressions

The paper addresses a fundamental challenge in computational linguistics: why do speakers choose specific linguistic tools—such as a pronoun ("he") versus a full definite description ("the vice president")—at different points in a discourse? Existing theories struggled to explain the mechanics of local coherence, or how individual sentences bind together to form a cohesive segment. For natural language generation systems, this presented a practical hurdle: how should a computer decide which expression to use to refer to an object without confusing the human listener?

The Centering Framework

The authors introduce "centering" as a mechanism for tracking local focus. The theory posits that every utterance in a discourse has two types of centers:

  • Backward-looking center ($Cb$): This is the single entity that the current sentence focuses on to link it back to the previous discourse. It serves as the "topic" of the sentence.
  • Forward-looking centers ($Cf$): This is a set of entities introduced in the current sentence that are candidates for becoming the $Cb$ of the subsequent sentence.

By tracking these centers, the theory explains how the focus of a conversation shifts or persists from one sentence to the next. This process is not purely syntactic; the authors demonstrate that the center of a sentence can change based on the preceding context, even if the sentence structure remains identical (e.g., the difference in focus between "Max saw Rosa" as an answer to "Who did Max see?" versus "Who saw Rosa?").

Integrating Semantics and Pragmatics

A key contribution of the paper is the integration of Situation Semantics (drawing on the work of Barwise and Perry). This allows the authors to distinguish between:

  1. Value-free interpretations: Referring to a role or description (e.g., "the president" as a functional role).
  2. Value-loaded interpretations: Referring to the specific individual currently occupying that role.

Furthermore, the authors incorporate Keith Donnellan’s distinction between referential and attributive uses of descriptions. They argue that a "center" is often determined by the speaker's intention rather than literal denotation. For example, if a speaker mistakenly refers to a person as "her husband," that expression still realizes the intended person as the center of the discourse, provided the listener understands the intended referent.

The Pronoun Signal and Cognitive Constraints

One of the paper's most influential insights is the treatment of pronouns as cognitive signals rather than mere linguistic shortcuts. The authors propose a core rule: If the $Cb$ of the current sentence is the same as the $Cb$ of the previous sentence, the speaker should use a pronoun.

When a speaker follows this rule, they signal continuity, making the discourse easier to process. If a speaker uses a full noun phrase to refer to a continuing center, it forces the listener to perform additional "inferencing" to determine if the speaker is still discussing the same entity or shifting focus. This formalizes the relationship between linguistic form and the listener's mental effort.

Impact and Theoretical Integration

The centering model successfully simplified prior approaches, such as Candace Sidner’s "actor focus" model, by providing a more streamlined account of how multiple pronouns are resolved. It also demonstrated that the interpretation of definite noun phrases—whether they are seen as value-loaded or value-free—is directly influenced by whether those phrases are currently "centered" in the local discourse.

Finally, the paper distinguishes between Centering (local, sentence-to-sentence focus) and Global Focusing (the overarching theme of a multi-segment conversation), providing a roadmap for how these two levels of discourse structure interact.

Limitations and Future Research

As a theoretical foundation, the paper acknowledges several areas for further development:

  • The Nature of the Rule: The pronoun constraint is described as a "Gricean maxim"—a principle of cooperative conversation that can be violated. The specific conditions under which speakers choose to violate this rule remain an open question.
  • Implicit Focusing: The current model does not fully account for functional relations, such as when mentioning a "house" allows a subsequent reference to "the door" to be easily understood.
  • Global/Local Interaction: While the distinction between local and global focus is defined, the exact rules governing how local shifts eventually trigger a change in the global discourse state require further investigation.

Summary updated

BibTeX
@inproceedings{grosz1983providing,
  author = {Barbara J. Grosz and Aravind K. Joshi and Scott Weinstein},
  title = {{Providing a Unified Account of Definite Noun Phrases in Discourse}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1983},
  doi = {10.3115/981311.981320},
}
Figure