Abstract:

The existing Chinese open domain Knowledge-Based Question Answering（KBQA） system is lack of support for Multi-Relational Question Answering（MRQA）. The KBQA process is divided into three steps：entity recognition, entity relation extraction and answer retrieval. The method of entity relationship extraction is discussed in detail. In the stage of entity relation extraction, a rule-based relational word extraction method is proposed to extract the relational words in the question. Then the similarity between relational words and predicates in knowledge base is calculated to get the relation set. Combining with the results of entity recognition, the question is transformed into triples with semantic information. Experimental results show that the method can support MRQA with higher average F1 score.

Key words: knowledge-based question answering, entity relation extraction, similarity calculation

摘要：

针对现有的中文开放领域知识库问答系统缺乏对多关系问答的支持，将知识库问答过程分为实体识别、实体关系抽取和答案检索三个步骤，重点讨论了实体关系抽取的实现方法。在实体关系抽取阶段，提出一种基于规则的关系词提取方法抽取问句中的关系词，然后将关系词与知识库中的谓词进行相似度计算，得到关系集合，结合实体识别的结果，将问句转换为具有语义信息的三元组形式。实验结果表明，该方法可以支持多关系问答，并且具有较高的平均F1分数。

关键词: 知识库问答, 实体关系抽取, 相似度计算