基于语义和主题模型的宏观篇章关系识别方法  

Macro Discourse Relation Identification on Semantic and Topic Model

作　　者：周俊颢 朱巧明[1] 蒋峰 褚晓敏[1] 李培峰[1] ZHOU Junhao;ZHU Qiaoming;JIANG Feng;CHU Xiaomin;LI Peifeng(School of Computer Science and Technology,Soochow University,Suzhou 215006,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《山西大学学报（自然科学版）》2021年第3期393-402,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61773276,61836007,61772354)。

摘　　要：篇章关系识别旨在分析篇章单元之间的逻辑关系。相较于微观层面,由于缺少明显关系连接词和更大的篇章单元粒度,宏观篇章关系识别更加困难。对此,文章提出一种将篇章单元语义信息建模和主题信息建模相结合的方法,能够压缩篇章单元的表示,更好发掘篇章单元之间隐含的语义关系。在基于中文宏观篇章树库(MCDTB)语料的实验中,有效识别了样例较少的篇章关系类别,并从整体上提高了篇章关系识别的性能。Discourse relation identification aims to analyze the logical relationship between discourse units. Compared with the micro level, due to the lack of obvious relational conjunctions and larger discourse unit granularity, it is more difficult to identify macro discourse relation. This paper combines semantic and topic modeling of discourse units to identify macro discourse relations. It can compress the representation of large discourse units and better explore the implicit semantic relationship among discourse units. In the experiment based on Macro Chinese Discourse Treebank(MCDTB) corpus, we effectively identify the relations with fewer samples, and improve the performance of discourse relation recognition as a whole.

关 键 词：宏观篇章关系识别 主题分布 语义模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]