Abstract
This paper presents a systematic, large-scale evaluation of Centering Theory (CT) using real-world corpora. By testing thousands of parameter configurations, the researchers identify a fundamental trade-off between the theory's ability to model discourse coherence and its ability to predict pronominalization. While Rule 1 (salience) remains robust across various settings, the study highlights significant challenges in modeling "NULL" transitions and demonstrates the theory's high sensitivity to the definition of basic linguistic units.
Introduction
In the fields of Natural Language Processing (NLP) and linguistics, Centering Theory (CT) is a prominent framework for understanding discourse coherence. It posits that human discourse is coherent because we track a "center" (a primary entity or topic) from one utterance to the next. The theory makes specific claims: every sentence has exactly one center, and if a pronoun is used, it should refer to that center.
However, Centering Theory is notoriously "underspecified." It does not strictly define what counts as an "utterance" (e.g., a full sentence vs. a clause) or how to "rank" entities to identify the center. Historically, this has led to small-scale studies with conflicting results. This research seeks to provide a definitive, corpus-based evaluation across a wide range of these parameter settings.
Methodology
The researchers moved away from hand-picked examples to a rigorous, automated approach:
- Multi-Parameter Testing: The team developed a script to test thousands of parameter configurations, varying definitions of utterances (sentences vs. finite clauses), realization (direct mention vs. indirect "bridging" references), and ranking (grammatical function vs. linear order).
- Reliable Annotation: They utilized the GNOME corpus, applying a detailed manual to ensure that human annotations were replicable and consistent.
- Statistical Validation: Sign tests and chi-square tests were used to determine if the theory's constraints and rules were statistically significant.
- Rhetorical Integration: The study also used Relational Discourse Analysis (RDA) to examine how CT interacts with the broader rhetorical structure of a text.
Datasets
The study utilized three primary sources to provide a diverse set of technical and instructional texts:
- GNOME Corpus (Museum Subcorpus): Descriptions of museum objects and artists.
- GNOME Corpus (Pharmaceutical Subcorpus): Patient information leaflets for medicines.
- Sherlock Corpus: Tutorial dialogues between students and tutors, used specifically for rhetorical structure experiments.
Key Insights: The Coherence-Salience Trade-off
The authors discovered a significant "Coherence-Salience Trade-off." When parameter definitions were adjusted to make the text appear more coherent (satisfying Constraint 1), the theory's ability to predict pronoun usage (Rule 1) often decreased.
Other key findings include:
- The "Vanilla" Failure: The standard textbook version of CT often fails to support the idea that every sentence has a center. In real-world text, many sentences simply do not refer back to the previous one.
- The "Null" Transition: In some texts, up to 47% of transitions were "NULL," meaning adjacent sentences shared no entities. Standard CT largely ignores these, yet they are fundamental to instructional writing.
- Genre Sensitivity: The theory performed significantly better on museum descriptions than on pharmaceutical leaflets, which frequently use the second-person "you," creating noise the model struggled to handle.
Results
- Rule 1 (Salience): This claim—that the "Backward-Looking Center" is the most likely candidate for pronominalization—is highly robust, with satisfaction rates as high as 97.1%.
- Constraint 1 (Continuity): The "Strong" version of this constraint only holds if utterances are defined as full sentences or if "indirect realization" (bridging) is permitted.
- Ranking Methods: While grammatical function (Subject > Object) is the standard, linear order (first mentioned) performed nearly as well and is computationally simpler to implement.
Conclusion and Future Work
The study concludes that while Centering Theory is a powerful tool for tracking entity salience, it is insufficient as a standalone model for discourse coherence. Future work must better integrate entity-based tracking with relational coherence (how actions and ideas connect) and address the high frequency of NULL transitions in non-narrative genres. Additionally, the difficulty of reliably annotating "bridging" references remains a hurdle for automated implementations.
Summary updated