摘要： 传统自动问答方法通常依赖谓词等先验信息实现知识库问答，需要耗费较多的人力且泛化能力不佳。提出一种针对弱依赖信息的知识库问答方法，结合BERT与BiLSTM-CRF网络提取问句中的命名实体，定位知识库中与该实体相关的三元组信息，通过答案匹配网络为三元组集合中的答案标上相似度分数，使用阈值选择策略选取符合要求的答案集合，并按照相似度分数由高到纸排序后呈现给用户。实验结果表明，该方法弱化了对先验信息的依赖，在减少人工干预的同时保证了问答质量，并且在NLPCC-ICCPOL-2016KBQA数据集上取得了87.05%的F1分数。

关键词: 弱依赖信息, 知识库问答, 命名实体识别, 答案匹配, 阈值选择

Abstract: Traditional automatic question answering methods mostly rely on priori information such as predicate to realize knowledge base question answering, which is labor-intensive and leads to poor generalization performance.To address the problem, this paper proposes a KBQA method for weak dependency information.The method uses BERT in combination with the BiLSTM-CRF network to extract the named entity in a question.Then, it locates the triple information related to the entity in the knowledge base, and uses the answer matching network to give similarity scores to the answers in the triple set.Finally, it uses the threshold selection strategy to select the answer set that meets the requirements, and the answers are sorted according to the similarity before being presented to the user.Experimental results show that the method weakens the dependence on priori information, and reduces the manual intervention while ensuring the quality of KBQA.It achieves an F1 score of 87.05% on the NLPCC-ICCPOL-2016KBQA data set.

Key words: weak dependency information, knowledge base question answering, named entity recognition, answer matching, threshold selection

中图分类号: 

• TP391