A Model for the Recognition of Discourse Relations

Discourse parsing is often realized as a process of identifying elementary discourse units (EDUs) and adding rhetorical structure on top of them, by specifying which (spans of) EDUs interact with each other and what kind of rhetorical relations hold between them. In this paper we describe our model for the recognition of discourse relations and we report the results of experiments performed in this model. The experiments are run on the RST-DT corpus, using a fine-grained taxonomy of relations consisting in 167 distinct labels, compared to 41 distinct labels used in previous research. We also show how a combination of two classifiers, which discriminate between explicit and implicit relations, achieves better performance than a single feature-rich classifier.