作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (2): 41-47. doi: 10.19678/j.issn.1000-3428.0053810

• 人工智能与模式识别 • 上一篇    下一篇

基于BiLSTM-CRF的细粒度知识图谱问答

张楚婷a, 常亮a, 王文凯a, 陈红亮a, 宾辰忠b   

  1. 桂林电子科技大学 a. 广西可信软件重点实验室;b. 卫星导航定位与位置服务国家地方联合工程研究中心, 广西 桂林 541004
  • 收稿日期:2019-01-25 修回日期:2019-03-14 发布日期:2019-03-22
  • 作者简介:张楚婷(1993-),女,硕士研究生,主研方向为自然语言处理、智能问答;常亮,教授、博士;王文凯、陈红亮,硕士研究生;宾辰忠,讲师、博士。
  • 基金资助:
    国家自然科学基金(U1711263,U1811264);广西创新驱动发展专项(科技重大专项)(AA17202024)。

Fine-grained Question Answering over Knowledge Graph Based on BiLSTM-CRF

ZHANG Chutinga, CHANG Lianga, WANG Wenkaia, CHEN Honglianga, BIN Chenzhongb   

  1. a. Guangxi Key Laboratory of Trusted Software;b. National-Local Joint Engineering Research Center for Satellite Navigation Positioning and Location Service, Guilin University of Electronic Technology, Guilin, Guangxi 541004, China
  • Received:2019-01-25 Revised:2019-03-14 Published:2019-03-22

摘要: 基于知识图谱的问答中问句侯选主实体筛选步骤繁琐,且现有多数模型忽略了问句与关系的细粒度相关性。针对该问题,构建基于BiLSTM-CRF的细粒度知识图谱问答模型,其中包括实体识别和关系预测2个部分。在实体识别部分,利用BiLSTM-CRF模型提高准确性,并将N-Gram算法与Levenshtein距离算法相结合用于候选主实体的筛选,简化候选主实体筛选过程。在关系预测部分,分别应用注意力机制和卷积神经网络从语义层次和词层次捕获问句与关系之间的相互联系。使用FreeBase中的FB2M和FB5M评估数据集进行实验,结果表明,与针对单一关系的问答方法相比,该模型对于实体关系对的预测准确率更高。

关键词: 实体识别, 关系预测, 知识图谱, 卷积神经网络, 问答模型, N-Gram算法

Abstract: Question answering over knowledge graph is complex in the filtering of candidate master entities of questions,and most existing models ignore the fine-grained correlation between questions and relationships.To address the problem,this paper proposes a fine-grained question answering model over knowledge graph based on BiLSTM-CRF.The model is divided into two parts: entity recognition and relationship prediction.In the entity recognition part,the model uses the BiLSTM-CRF algorithm to improve accuracy,and the N-Gram algorithm is combined with the Levenshtein Distance algorithm to simplify the filtering process of candidate master entities.In the relationship prediction part,attention mechanism and Convolutional Neural Network(CNN) are used to capture the correlation between questions and relationships at the semantic level and the word level.Experimental results on the FB2M and FB5M evaluation datasets in FreeBase show that the proposed model has higher accuracy of entity relationship pair prediction compared with existing question answering methods for a single relationship.

Key words: entity recognition, relation prediction, knowledge graph, Convolutional Neural Network(CNN), question answering model, N-Gram algorithm

中图分类号: