Chinese Macro Discourse Parsing on Generative Fusion and Distant Supervision

Longwang He, Feng Jiang, Xiaoyi Bao, Yaxin Fan, Peifeng Li, Xiaomin Chu

Abstract: Most previous studies on discourse parsing have utilized discriminative models to construct tree structures. However, these models tend to overlook the global perspective of the tree structure as a whole during the step-by-step top-down or bottom-up parsing process. To address this issue, we propose DP-GF, a macro Discourse Parser based on Generative Fusion, which considers discourse parsing from both process-oriented and result-oriented perspectives. Additionally, due to the small size of existing corpora and the difficulty in annotating macro discourse structures, DP-GF addresses the small-sample problems by proposing a distant supervision training method that transforms a relatively large-scale topic structure corpus into a high-quality silver-standard discourse structure corpus. Our experimental results on MCDTB 2.0 demonstrate that our proposed model outperforms the state-of-the-art baselines on discourse tree construction.