Disambiguating Explicit Discourse Connectives without Oracles

Deciding whether a word serves a discourse function in context is a prerequisite for discourse processing, and the performance of this subtask bounds performance on subsequent tasks. Pitler and Nenkova (2009) report 96.29% accuracy (F1 94.19%) relying on features extracted from gold-standard parse trees. This figure is an average over several connectives, some of which are extremely hard to classify. More importantly, performance drops considerably in the absence of an oracle providing gold-standard features. We show that a very simple model using only lexical and predicted part-of-speech features actually performs slightly better than Pitler and Nenkova (2009) and not significantly different from a state-of-the-art model, which combines lexical, part-ofspeech, and parse features.