Improving a Pipeline Architecture for Shallow Discourse Parsing

We present a system that implements an end-to-end discourse parser. The system uses a pipeline architecture with seven stages: preprocessing, recognizing explicit connectives, identifying argument positions, identifying and labeling arguments, classifying explicit and implicit connectives, and identifying attribution structures. The discourse structure of a document is inferred based on these components. For NLP analysis, we use Illinois NLP software1 and the Stanford Parser. We use lexical and semantic features based on function words, sentiment lexicons, brown clusters, and polarity features. Our system achieves an F1 score of 0.2492 in overall performance on the development set and 0.1798 on the blind test set.