Automatic Discourse Segmentation using Neural Networks

In example (1), a sentence from a Wall Street Journal article taken from the Penn TreeBank corpus is further segmented into four EDUs, (1a), (1b), (1c) and (1d) (RST, 2002). Discourse segmentation, clearly, is not as easy as sentence boundary detection. The lack of consensus with regards to what constitutes an elementary discourse unit adds to the difficulty. Building a rule based discourse segmenter can be a tedious task since these rules would have to be based on the underlying grammar of the particular parser that is to be used. Therefore, we adopted a neural network model for automatically building a discourse segmenter from an underlying corpus of segmented text. We chose to use part-of-speech tags, syntactic information, discourse cues and punctuation. Our ultimate goal is to build a discourse parser that uses this discourse segmenter. The data that we used to train and test our discourse segmenter is the RST-DT (RST, 2002) corpus. The corpus contains 385 Wall Street Journal articles from the Penn Treebank. The training set consists of 347 articles for a total of 6132 sentences, whilst the test set contains 38 articles for a total of 991 sentences. The RST-DT corpus provides us with pairs of sentences and EDUs. For the syntactic structure of the sentences, we have used both the gold standard Penn Treebank data and syntactic parse trees generated by (Charniak, 2000). As regards the discourse cues, we used a list of 168 possible discourse markers.