Learning Connective-based Word Representations for Implicit Discourse Relation Identification

International audience; We introduce a simple semi-supervised approach to improve implicitdiscourse relation identification. This approach harnesses largeamounts of automatically extracted discourse connectives along withtheir arguments to construct new distributional wordrepresentations. Specifically, we represent words in the space ofdiscourse connectives as a way to directly encode their rhetoricalfunction. Experiments on the Penn Discourse Treebank demonstrate theeffectiveness of these task-tailored representations in predictingimplicit discourse relations. Our results indeed show that, despitetheir simplicity, these connective-based representations outperformvarious off-the-shelf word embeddings, and achieve state-of-the-artperformance on this problem.