Designing a Discourse Parser for the Evaluative Text Genre

We propose designing a discourse parser specifically for the evaluative text genre. We aim to see whether focusing on a certain genre and relations specific to that genre offers performance gain beyond more generic discourse parsers. In this extended abstract we describe the approach we intend to take, and how this differs from what has been done previously. The problem of discourse parsing It has often been pointed out that in any text sentences are not understood in isolation but in relation to each other. Several theories postulate that all discourse has a hierarchical structure of coherence relations which reflect the author's intended effect on the reader. Discourse structure is often based on semantic and pragmatic relationships, and results in a tree structure whose leaves are the elementary discourse units (EDUs) of the text. The goal of discourse parsing is to build this tree structure and to identify the rhetorical relations labeling the internal nodes of the tree. Discourse parsing consists of three co-dependent subtasks: i) identifying the EDUs, ii) determining which discourse segments are related to each other, and iii) identifying the rhetorical relation between these segments. Discourse parsing has applications in other NLP tasks such as sentence compression, summarization and natural language generation (NLG). For instance, knowing that two EDUs are in an Elaboration relation, and one of them is the nucleus, would enable us to prune the satellite EDU when generating a summary. In NLG, knowing what features usually signal a Concession relation can enable us to generate such instances ourselves.