A Recursive Information Flow Gated Model for RST-Style Text-Level Discourse Parsing

Text-level discourse parsing is notoriously difficult due to the long-distance dependency over the document and the deep hierarchical structure of the discourse. In this paper, we attempt to model the representation of a document recursively via shift-reduce operations. Intuitively, humans tend to understand macro and micro texts from different perspectives, so we propose a recursive model to fuse multiple information flows and strengthen the representation of text spans. During parsing, the proposed model can synthetically grade each information flow according to the granularity of the text. Experimentation on the RST-DT corpus shows that our parser can outperform the state-of-the-art in nuclearity detection under stringent discourse parsing evaluations.