The Overall Markedness of Discourse Relations

Lifeng Jin, Marie-Catherine de MarneffeEMNLP 2015 · OpenReview

Abstract

Discourse relations can be categorized as continuous or discontinuous in the hypothesis of continuity (Murray, 1997), with continuous relations expressing normal succession of events in discourse such as temporal, spatial or causal. Asr and Demberg (2013) propose a markedness measure to test the prediction that discontinuous relations may have more unambiguous connectives, but restrict the markedness calculation to relations with explicit connectives only. This paper extends their measure to explicit and implicit relations and shows that results from this extension better fit the continuity hypothesis predictions both for the English Penn Discourse (Prasad et al., 2008) and the Chinese Discourse (Zhou and Xue, 2015) Treebanks.

Introduction

In human communication, discourse relations link ideas to create coherent narratives. These links are often categorized as "continuous"—maintaining a steady frame of reference (e.g., “I was tired, so I drank coffee”)—or "discontinuous," which shift perspective or create conflict (e.g., “I drank coffee, but I was still tired”).

The Continuity Hypothesis posits that continuous relations are the cognitive "default" and are easier to process. Previous attempts to prove this used "markedness"—an information-theoretic measure of how strongly a relation (like Contrast) is tied to specific signaling words (like "but"). However, earlier models (M-exp) only considered explicit connectives. This led to a theoretical contradiction: "Temporal" relations, which should be cognitively simple, appeared highly "marked" because they use very specific words when they are explicit. This paper addresses this gap by accounting for implicit relations—cases where no connective is used.

Methods

The authors propose a new, inclusive model, M-all, which extends the markedness measure beyond explicit connectives.

  1. Information Theory: The model uses Normalized Pointwise Mutual Information (NPMI) to quantify the association between a discourse relation and a connective. A high score indicates a "marked" relation that rarely occurs without a specific, unambiguous signal.
  2. The "Null" Connective: The core innovation is treating implicit relations as having a "null" or "zero" connective. In the M-all model, if a relation frequently occurs implicitly, its markedness score decreases because it does not strictly "need" a word to be understood.
  3. Surprisal Theory: From a psycholinguistic perspective, a highly marked relation is more "expensive" to process because it represents a significant shift from the expected discourse path, exerting a strong influence on the distribution of upcoming words.

Datasets

The study evaluates the M-all measure across two major corpora:

  • Penn Discourse Treebank (PDTB): The standard for English discourse, containing hierarchical annotations of relations.
  • Chinese Discourse Treebank (CDTB): A corpus for Chinese discourse using a flatter structure of ten relations.

Key Results and Insights

The inclusion of implicit relations significantly improves the alignment between empirical data and the Continuity Hypothesis.

Fixing the Temporal Anomaly

In the PDTB, the M-all measure successfully lowered the markedness of Temporal and Contingency relations compared to the previous M-exp measure. By accounting for the high frequency of implicit Temporal relations, the model correctly identifies them as "unmarked" or expected.

Markedness of top-level PDTB relations comparing the explicit-only model (M-exp) with the inclusive model (M-all).

Expansion as the Cognitive Default

Across both English and Chinese, "Expansion" (adding information) emerges as the least marked relation. This suggests that in the absence of a connective, the human brain's default assumption is that the next sentence simply expands on the previous one. Conversely, discontinuous relations like "Exception" or "Comparison" consistently show higher markedness, validating that they pose a conceptual difficulty requiring explicit signaling.

Cross-Linguistic Validation

Despite the different distributions of explicit connectives in English and Chinese (Chinese uses them much less frequently), the relative markedness of relations remains remarkably similar across both languages.

Markedness of CDTB (Chinese) relations, demonstrating similar trends across languages.

Conclusion

By incorporating implicit relations into the calculation of markedness, this research provides strong evidence for the Continuity Hypothesis. The results suggest that "Expansion" may be a universal default discourse relation. Future work could further integrate measures of "implicitness" to create a more robust model of how humans navigate the conceptual map of conversation across different languages.

Summary updated

Abstract

Discourse relations can be categorized as continuous or discontinuous in the hypothesis of continuity (Murray, 1997), with continuous relations expressing normal succession of events in discourse such as temporal, spatial or causal. Asr and Demberg (2013) propose a markedness measure to test the prediction that discontinuous relations may have more unambiguous connectives, but restrict the markedness calculation to relations with explicit connectives only. This paper extends their measure to explicit and implicit relations and shows that results from this extension better fit the continuity hypothesis predictions both for the English Penn Discourse (Prasad et al., 2008) and the Chinese Discourse (Zhou and Xue, 2015) Treebanks.
BibTeX
@inproceedings{jin2015overall,
  author = {Lifeng Jin and Marie-Catherine de Marneffe},
  title = {{The Overall Markedness of Discourse Relations}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2015},
  url = {https://openreview.net/forum?id=ry-LdzGuZH},
}
Figure