Subgraph-based Classification of Explicit and Implicit Discourse Relations

Current approaches to recognizing discourse relations rely on a combination of shallow, surfacebased features (e.g., bigrams, word pairs), and rather specialized hand-crafted features. As a way to avoid both the shallowness of word-based representations and the lack of coverage of specialized linguistic features, we use a graph-based representation of discourse segments, which allows for a more abstract (and hence generalizable) notion of syntactic (and partially of semantic) structure. Empirical evaluation on a hand-annotated corpus of German discourse relations shows that our graphbased approach not only provides a suitable representation for the linguistic factors that are needed in disambiguating discourse relations, but also improves results over a strong state-of-the-art baseline by more accurately identifying Temporal, Comparison and Reporting discourse relations.