计算机工程

• •    

基于改进图节点的图神经多跳阅读理解研究

  

  • 发布日期:2021-01-08

Research on multi-hop reading comprehension based on graph neural network with improved graph nodes

  • Published:2021-01-08

摘要: :多跳阅读理解需要基于问题在多个支撑文档中寻找相关信息并进行跳跃式推理来回答问题。针对当 前多跳阅读理解模型中所存在的实体图内缺乏关键问题信息以及信息量冗余的问题,提出了一种基于改进 图节点的图神经多跳阅读理解模型。首先,采用基于语法的实体提取方法提取实体,将提取到的实体基于 问题关联实体构建实体图。然后,对实体图中的节点进行编码预处理,并通过门机制的图卷积网络模拟推 理得到推理信息。最后,计算推理信息与问题信息的双向注意力并进行结果预测。通过在 WikiHop 数据集 中测试实验,结果表明相对于其他模型在测试集上效果有所提升,证明了该方法的有效性。

Abstract: Multi-hop reading comprehension requires searching for relevant information in supporting documents to answer the question according to the question understanding and leaping reasoning. In the current multi-hop reading comprehension models, the entity graph lacks pivotal answering information but contains superabundant useless information. To address this problem, we propose a multi-hop reading comprehension model based on graph neural network with improved graph nodes. Firstly, we build the entity graph based on the question-associated entities which are extracted based on syntax. After the nodes in the entity graph are preprocessed by encoding, we use the gated graph convolution network to obtain the inference information. Finally, we calculate the bidirectional attention between reasoning and question information and predict the answer. The experiment in the WikiHop dataset shows that the results are better than those of other models, which proves the effectiveness of the method.