Global Context-aware Representation for Implicit Discourse Relation Recognition

Meilin Zhou, Qi Liang, Peng Zhang, Lu Ma, Dan Luo, Bin Wang

Abstract: Identify implicit discourse relations requires an in-depth understanding of the discourse. We argue that discourse relation recognition should not be isolated from the discourse. However, previous works mostly treat two arguments as independent sentences or encode sentences within a synthetic paragraph context only containing annotated arguments. The overall discourse-level content and structure sequence are partly missed. In this paper, we propose a global context-aware model to leveraging discourse context to improve argument representation. Besides, we integrate a multi-view interaction representation between two arguments themselves to the global context-aware representation to boost implicit discourse recognition performance. Experiments show that our model achieves state-of-the-art results on the benchmark PDTB corpus.