融合全局和局部信息的汉语宏观篇章结构识别

Identification of Chinese Macro Discourse Structure with Global and Local Information

作为宏观篇章分析中的基础任务,篇章结构识别的目的是识别相邻篇章单元之间的结构,并层次化构建篇章结构树。已有的工作只考虑局部的结构和语义信息或只考虑全局信息。该文提出了一种融合全局和局部信息的指针网络模型,该模型在考虑全局语义信息的同时,又考虑局部段落间的语义关系密切程度,从而有效提高宏观篇章结构识别的能力。在汉语宏观篇章树库(MCDTB)的实验结果表明,该文所提出的模型性能优于目前性能最好的模型。

The discourse structure recognition task aims to identify the structure between adjacent discourse units for a hierarchical discourse structure tree.This paper proposes a pointer network model that integrates global and local information.It can effectively improve the ability of macro text structure recognition by considering the global semantic information and the closeness of the semantic relationship between paragraphs.The experimental results in the Chinese macro discourse Treebank(MCDTB)show that the proposed model outperforms the state-of-the-art model.