中文多方对话篇章结构语料库构建及分析模型  

Corpus construction and parsing model of Chinese multi-party dialoguediscourse structure

作　　者：何威恺 蒋玉茹[1,2] 陈杰 李宇 张仰森 HE Weikai;JIANG Yuru;CHEN Jie;LI Yu;ZHANG Yangsen(Institute of Intelligent Information Processing,Beijing Information Science&Technology University,Beijing 100101,China;Beijing Laboratory of National Economic Security Early-warning Engineering,Beijing 100044,China)

机构地区：[1]北京信息科技大学智能信息处理研究所,北京100101 [2]国家经济安全预警工程北京实验室,北京100044

出　　处：《北京信息科技大学学报（自然科学版）》2023年第5期10-18,共9页Journal of Beijing Information Science and Technology University

基　　金：国家自然科学基金资助项目(61602044)。

摘　　要：针对面向篇章结构分析的对话文本语料库缺乏的问题,构建了第一个中文的日常生活场景下多方对话篇章结构语料库。同时,提出了一种基于图神经网络的篇章结构分析模型,针对日常生活场景下多方对话的特点,使用图注意力网络进行图网络设计和基本篇章单元编码;并且模型融入了角色识别特征及少量手工特征,使得篇章结构分析性能进一步提升。该模型在所构建的语料库上的实验结果与基线模型相比,篇章结构分析中的链接识别和关系识别两个任务的F1值分别提升了7.8和3.6百分点。In view of the lack of dialogue text corpora for discourse structure parsing,the first Chinese multi-party dialogue discourse structure corpus in daily life scenarios was constructed.Meanwhile,a discourse structure parsing model based on graph neural networks was proposed,which designed graph networks and coded element discourse units using graph attention network according to the characteristics of multi-party dialogues in daily life scenarios.Moreover,the model incorporated character recognition features and a small amount of manual features,further improving the discourse structure parsing capability.The experimental results of this model on the constructed corpus show that the F 1 values of connection recognition and relation recognition tasks for discourse structure parsing increased by 7.8 and 3.6 percentage points compared with the baseline model respectively.

关 键 词：多方对话 篇章结构分析 图神经网络 角色识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]