一种应用于KBQA关系检测的多视角层次匹配网络

Multi-view Hierarchical Matching Network for KBQA Relation Detection

知识库问答(KBQA)是指利用知识库中的一个或多个知识三元组回答一个自然语言问题,需要检测问题中提及的知识库实体和关系.关系检测是知识库问答的核心.为了解决现有关系检测方法存在的匹配视角单一和信息瓶颈问题,本文提出了一种多视角层次匹配网络(M-HMN,Multi-view Hierarchical Matching Network),M-HMN利用双向注意力机制对齐问题与候选关系的不同特征,强化两者匹配部分的观察精细度,将匹配信息封装成向量,再由自注意力机制有效聚合多个向量以进行正确关系检测.对于KBQA最终任务的评估,本文提出一种简易的实体重排序算法,利用M-HMN网络优化候选实体集.实验结果表明,M-HMN能有效缓解关系检测的信息瓶颈问题,而提出的实体重排序算法能够进行实体消歧,获得更小更为精准的候选实体集,对KBQA最终任务性能有显著的提升.

Knowledge Base Question Answering(KBQA)systems answer a natural language question by one or more knowledge triples in KB which need to detect the KB entities and relations mentioned in the question.Relation detection is the core of KBQA.In order to solve the problem of single matching perspective and information bottleneck existing in the existing methods,this paper proposes a multi-perspective hierarchical matching network M-HMN.M-HMN makes use of bi-directional attention mechanism alignment question and different characteristics of candidate relation to enhance the observation precision of their matching part,encapsulate the matching information into vectors,and effectively aggregate multiple vectors by self-attention mechanism to make correct relation prediction.For the evaluation of KBQA final task,this paper proposes a simple entity ranking algorithm to optimize candidate entity sets by using M-HMN network.Experimental results showthat M-HMN can effectively alleviate the information bottleneck of relation detection,and the proposed entity ranking algorithm can perform entity disambiguating and obtain a smaller and more accurate candidate entity set,which significantly improves the final task performance of KBQA.