Creating Local Coherence: An Empirical Assessment

Annie Louis, Ani NenkovaNAACL 2010 · OpenReview

Abstract

Two of the mechanisms for creating natural transitions between adjacent sentences in a text, resulting in local coherence, involve discourse relations and switches of focus of attention between discourse entities. These two aspects of local coherence have been traditionally discussed and studied separately. But some empirical studies have given strong evidence for the necessity of understanding how the two types of coherence-creating devices interact. Here we present a joint corpus study of discourse relations and entity coherence exhibited in news texts from the Wall Street Journal and test several hypotheses expressed in earlier work about their interaction.

Introduction

Local coherence is the linguistic property that transforms a sequence of sentences into a unified text. Traditionally, research has bifurcated this phenomenon into two structures: Attentional Structure, which focuses on entity coherence (tracking salient objects or concepts across sentences), and Relational Structure, which focuses on discourse relations (inferring logical connections like Cause or Contrast).

A long-standing theoretical assumption suggests these systems are complementary: if a sentence pair lacks a strong discourse relation, entity sharing must provide coherence; conversely, if no entities are shared, a "core" discourse relation must bridge the gap. This study empirically tests these assumptions using a large-scale, jointly annotated corpus.

Methods

The researchers utilized the intersection of the Penn Discourse Treebank (PDTB) and OntoNotes, analyzing adjacent sentences in 590 Wall Street Journal articles. Discourse relations were categorized into:

  • Core Relations: Strong logical connections (Temporal, Contingency, Comparison).
  • Weak Relations: Expansion and Entity Relations (EntRel).

These were cross-referenced with coreference data to determine the frequency of entity sharing across different relation types.

Results: Challenging Theoretical Assumptions

The empirical data largely contradicted established hypotheses regarding the interaction between discourse and entity coherence.

The Complementarity Myth

Contrary to the idea that core relations and entity sharing are mutually exclusive, the study found that core relations often exhibit higher rates of entity sharing than weak relations. For instance, Temporal relations showed the highest rate of entity overlap at nearly 67%.

Entity sharing percentages across different discourse relations.

The 30% Gap

The most significant finding is the existence of a "30% gap." Approximately 29.84% of all sentence pairs in the corpus neither share an entity nor are joined by a "core" discourse relation. This suggests that nearly a third of coherent transitions in professional news text are not explained by current primary theories of local coherence.

Distribution of core and weak relations across entity-sharing and non-sharing sentence pairs.

Furthermore, the hypothesis that sentences without shared entities are primarily held together by core relations was disproven; only 38.65% of non-entity-sharing pairs utilized core relations, with the majority falling under "weak" Expansion relations.

Discussion and Future Work

The results indicate that current models of local coherence are incomplete. The "Expansion" category, often treated as a catch-all, is highly diverse: Restatements show high entity overlap (56%), while List and Instantiation relations show very low overlap (under 35%).

The authors suggest that the missing piece of the coherence puzzle may be bridging anaphora—semantic relationships between non-identical but related entities (e.g., "the house" followed by "the door"). While difficult to annotate, incorporating these indirect links is essential for a more comprehensive understanding of how text achieves local coherence. Future work must move beyond treating entity and relational structures as simple substitutes and instead model their complex, overlapping interactions.

Summary updated

Abstract

Two of the mechanisms for creating natural transitions between adjacent sentences in a text, resulting in local coherence, involve discourse relations and switches of focus of attention between discourse entities. These two aspects of local coherence have been traditionally discussed and studied separately. But some empirical studies have given strong evidence for the necessity of understanding how the two types of coherence-creating devices interact. Here we present a joint corpus study of discourse relations and entity coherence exhibited in news texts from the Wall Street Journal and test several hypotheses expressed in earlier work about their interaction.
BibTeX
@inproceedings{louis2010creating,
  author = {Annie Louis and Ani Nenkova},
  title = {{Creating Local Coherence: An Empirical Assessment}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2010},
  url = {https://openreview.net/forum?id=Hybohz-uWH},
}
Figure