一种用于答案选择的知识增强图卷积网络

Knowledge-enhanced Graph Convolutional Network for Answer Selection

答案选择是问答领域的一个重要子任务,目标是根据问题从候选答案集合中选择最合适的答案.该任务的核心是问答语义匹配.近年来,随着深度神经网络和预训练语言模型的应用,许多端对端的问答匹配模型展现出优异的性能.但是,现有模型仍然存在语义信息提取不充分以及未有效利用外部知识信息等问题.针对上述问题,本文提出一种知识增强图卷积网络(A Knowledge-enhanced Graph Convolutional Network,KEGCN).首先,KEGCN提出一种基于图卷积神经网络的问题-答案结构信息提取机制,在利用BERT获得文本语义信息的基础上,KEGCN通过图卷积神经网络学习问答对之间的结构信息,增强语义信息.其次,KEGCN设计了一种基于自注意力门控网络的扩展知识语义构建机制,利用自注意力门控网络获取扩展知识实体之间的上下文语义关联并过滤知识噪声,增强模型的鲁棒性.最后,KEGCN利用多尺寸卷积神经网络提取多粒度的全局语义信息,以进一步提高答案选择推理的准确性.WikiQA和TrecQA数据集上的实验结果表明,与对比模型相比较,KEGCN的综合性能更加优异.

Answer selection is an important subtask in the field of question answering,where the goal is to select the most appropriate answer from several candidate answers according to the question.The core of this task is the semantic matching between the question and candidate answer.In recent years,with the application of deep neural networks and pre-trained language models,many end-to-end question answer matching models have shown excellent performance.However,existing models still have problems such as insufficient semantic information extraction and ineffective use of external knowledge information.In response to the above problems,this paper proposes a knowledge-enhanced graph convolutional network(KEGCN).First,KEGCN proposes a mechanism to extract question-answer structure information using graph convolutional neural network,and enhances the semantic information learned by BERT.Secondly,KEGCN designs an extended knowledge semantic construction mechanism based on self-attention gating network.The self-attention gating network can capture contextual semantic correlations between extended knowledge entities and filter noise,thereby enhancing the robustness of the model.Finally,KEGCN proposes a multi-scale convolutional neural network to extract multi-granularity global semantic information to further improve answer selection reasoning.Experimental results on the WikiQA and TrecQA datasets show that the KEGCN outperforms other state-of-the-art baselines models.