基于多特征实体消歧的中文知识图谱问答

Chinese Knowledge Based Question Answering Based on Multi-feature Entity Disambiguation

问答系统应用于人工智能、自然语言处理和信息检索领域获得了较好的效果,知识图谱问答(KBQA)作为其中的重要组成部分,是一项极具挑战性的自然语言处理任务。然而,目前常见的中文KBQA系统对于实体链接的实体消歧部分并没有给出很好的解决方法。提出一种基于多特征实体消歧的中文KBQA系统,通过结合实体自身的知名度特征、问句与实体关系的语义相似度特征、问句与实体的字符相似度特征和语义相似度特征,构建多特征实体消歧模型,提高实体链接准确率,为系统的问句分类和最优路径选取部分提供更准确的主题实体,从而提升系统性能。实验结果表明,该系统在CCKS2019-CKBQA评测数据的验证集上平均F1值为72.08%,其中采用多特征消歧模型的实体链接准确率达到90.84%,较使用知名度消歧模型和评测大赛第1名分别提升6.35和0.11个百分点。

The application of question answering system to the fields of artificial intelligence,natural language processing and information retrieval has got excellent results.Knowledge Based Question Answering(KBQA)is an important part of question answering,and is a challenging natural language processing task.The commonly used Chinese KBQA systems do not provide a satisfying entity disambiguation solution for entity linking.To address the problem,this paper proposes a Chinese KBQA system based on multi-feature entity disambiguation.It jointly utilizes the entity’s own popularity features,semantic similarity features of question and entity relations,character similarity features of question and entity,and semantic similarity features of question and entity,so as to implement entity disambiguation and improve entity linking.On this basis,the proposed system can provide more accurate subject entities for the question classification part and the optimal path selection part of the system to improve system performance.The experimental results show that the average F1 value of the proposed system on the verification set of CCKS2019-CKBQA evaluation data reaches 72.08%.Its entity linking module based on the multi-feature disambiguation model displays an accuracy of 90.84%,which is 6.35 percentage points higher than the module based on the popularity disambiguation model and 0.11 percentage points higher than the top 1 in CCKS2019-CKBQA evaluation competition.