Topic-driven Distant Supervision Framework for Macro-level Discourse Parsing via Transferring Models

Anonymous

Abstract: Discourse parsing, the task of analyzing the internal rhetorical structure of texts, is a challenging problem in natural language processing due to the complex linguistic structure and lack of large-scale and high-quality corpora, especially at the macro level. Recent studies have attempted to overcome this limitation by utilizing results from other NLP tasks (source task) to distantly supervise the discourse parsing (target task). However, most of them only consider shallow connections across tasks using result-converting methods. It brings more cascading errors and makes it difficult to continue training due to the heterogeneity of the source and target task. To address these issues, we propose a topic-driven distant supervision framework via transferring models. The key recipe of this framework is to transfer the topic segmentation model into a discourse parser by additionally considering the global structural correlation instead of a simple converting result algorithm for transferring knowledge. The experimental results on two RST-style datasets, in both Chinese (MCDTB) and English (RST-DT), demonstrate that our method outperforms strong baselines not only in distant-supervised scenarios but also in fully supervised settings.

Paper Type: long

Research Area: Discourse and Pragmatics

Contribution Types: Publicly available software and/or pre-trained models

Languages Studied: Chinese