用于问题生成的知识增强双图交互网络  

Knowledge-enhanced Bi-graph Interaction Neural Nework for Question Generation

作　　者：李亚峰[1,2] 叶东毅 陈羽中 LI Yafeng;YE Dongyi;CHEN Yuzhong(College of Computer and Data Sciences,Fuzhou University,Fuzhou 350108,China;Fujian Provincial Key Laboratory of Network Computing and Intelligent Information Processing,Fuzhou 350108,China)

机构地区：[1]福州大学计算机与大数据学院,福州350108 [2]福建省网络计算与智能信息处理重点实验室,福州350108

出　　处：《小型微型计算机系统》2024年第5期1032-1038,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61672158,61972097,U21A20472)资助;福建省科技重大专项专题项目(科教联合)项目(2021HZ022007)资助;福建省高校产学研合作项目(2021H6022)资助;福建省自然科学基金项目(2020J01494)资助。

摘　　要：问题生成是一项具有挑战性的自然语言处理任务,旨在生成具有给定答案和上下文的问题,近年来受到了广泛关注.最近,由于神经网络的发展,问题生成任务取得了较大的进展.然而,现有模型仍然存在未有效利用外部知识以及在利用图神经网络捕获隐藏结构信息未捕获语法信息等问题.针对上述问题本文提出知识增强双图交互网络KE-BGINN(Knowledge-Enhanced Bi-Graph Interaction Neural Network).首先为了有效利用外部知识信息,KE-BGINN通过知识图谱本身的图结构信息构造知识增强图,并利用图卷积网络对文本以及答案上下文语义信息进行扩充.其次,KE-BGINN引入一种双图交互机制,利用两个图卷积网络学习上下文的隐藏结构信息以及语法信息,在图间信息融合时,构造一个虚拟图来充分融合不同图之间的异构信息.最后,KE-BGINN利用指针网络解码机制来解决问题生成时罕见和未知词的问题.在SQuAD数据集上的实验结果证明,与对比模型相比较,KE-BGINN模型的综合性能更加优异.Question generation,a challenging natural language processing task aimed at generating questions with a given answer and context,has received extensive attention in recent years.Recently,the task of question generation has made great progress due to the development of neural networks.However,existing models still have problems such as not effectively utilizing external knowledge and not capturing grammatical information when using graph neural networks to capture hidden structural information.Aiming at the above problems,this paper proposes a knowledge-enhanced bi-graph interaction neural network(KE-BGINN)for question generation.First,in order to effectively utilize the external knowledge information,KE-BGINN constructs the knowledge augmentation graph through the graph structure information of the knowledge graph itself,and uses the graph convolutional network to expand the contextual semantic information of text and answers.Secondly,KE-BGINN introduces a bi-graph interaction mechanism,which uses two graph convolutional networks to learn the hidden structure information and syntactic information of the context,and constructs a virtual graph to fuse the heterogeneous information of different graphs during inter-graph information fusion.Finally,KE-BGINN utilizes the pointer network decoding mechanism to solve the problem of rare and unknown words when question generation.The experimental results on the SQuAD dataset prove that the comprehensive performance of the KE-BGINN model is better than that of the comparison model.

关 键 词：问题生成 知识图谱 图卷积网络 双图交互 虚拟图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]