基于全局信息语法增强的长文本机器阅读理解  

Global information with syntax enhancement for long-text machine reading comprehension

作　　者：丁美荣 周岑钰 曾碧卿 张鑫 DING Mei-rong;ZHOU Cen-yu;ZENG Bi-qing;ZHANG Xin(School of Software,South China Normal University,Foshan 528225,China)

机构地区：[1]华南师范大学软件学院,广东佛山528225

出　　处：《计算机工程与设计》2023年第6期1860-1866,共7页Computer Engineering and Design

基　　金：2021年度广东省基础与应用基础研究基金项目(2021A151501117)。

摘　　要：为更好捕获和理解长文本的上下文信息,提出一种基于全局信息语法增强的长文本机器阅读理解模型。在预训练语言模型基础上,引入语法依存关系树,将语法信息融合进上下文建模中,增强长文本表示;针对长文本存在的全局信息丢失问题,引入门控循环单元神经网络,负责处理和传递各分段信息。通过在大规模标准数据集上进行实验,验证了该模型较于其它方法有不同程度的提升,能够有效处理长文本机器阅读理解任务。To better capture and understand the context information of long text,a model of long text machine reading comprehension based on global information syntax enhancement was proposed.A syntactic dependency tree based on the pre-trained language model was introduced,syntactic information was integrated into context modeling to enhance long text representation.A gated recurrent unit neural network was introduced to solve the problem of global information loss in long text,responsible for processing and transferring information of each section.Results of experiments on large scale standard datasets show that the proposed model is better than other methods,and can effectively handle long text machine reading comprehension tasks.

关 键 词：机器阅读理解 预训练模型 循环神经网络 注意力机制 全局信息 语法关系 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]