Exploiting Linguistic Cues to Classify Rhetorical Relations

We propose a method for automatically identifying rhetorical relations. We use supervised machine learning but exploit cue phrases to automatically extract and label training data. Our models draw on a variety of linguistic cues to distinguish between the relations. We show that these feature-rich models outperform the previously suggested bigram models by more than 20%, at least for small training sets. Our approach is therefore better suited to deal with relations for which it is difficult to automatically label a lot of training data because they are rarely signalled by unambiguous cue phrases (e.g., "continuation").