桥梁检测领域知识图谱构建与知识问答方法

Knowledge graph construction and knowledge question answering method for bridge inspection domain

针对桥梁检测报告中数据抽取融合不充分以及管养决策过程中知识问答服务不足的问题,提出一种桥梁检测领域知识图谱构建与知识问答方法。首先,采用Web本体语言(OWL)对桥梁检测领域知识进行形式化语义建模,定义了该领域的概念、属性及公理约束;然后,采用基于Transformer编码器、双向长短时记忆(BiLSTM)网络和条件随机场(CRF)的联合模型以及网格化长短时记忆(Lattice-LSTM)网络对细粒度息进行抽取,并将融合后的实例数据存储在Neo4j图数据库中,实现知识图谱化表示;最后,采用朴素贝叶斯分类算法进行问题模板匹配,根据匹配结果生成结构化查询,并以自然语言形式返回问题答案,实现细粒度领域信息的交互式问答。在与卷积神经网络(CNN)、BiLSTM的对比实验中,该方法在命名实体识别、关系抽取任务中的F1值分别为93.28%、74.00%,优于上述神经网络模型。实验结果表明,所提方法能较好地适应桥梁检测领域交互式问答实际需求。

Focused on the issue that deficient data extraction and fusion in the bridge inspection reports and insufficient knowledge question answering services in the management and maintenance process,a method of knowledge graph construction and knowledge question answering for the field of bridge inspection was proposed.Firstly,a formal semantic model of the bridge detection including concepts,properties and axioms was established by the Web Ontology Language(OWL).Secondly,the fine-grained information in the bridge inspection reports was extracted by the joint model based on Transformer,Bidirectional Long Short-Term Memory(BiLSTM)network,Conditional Random Field(CRF)and the Lattice Long Short-Trem Memory(Lattice-LSTM)network,and the fused data was stored in Neo4j graph database,so that knowledge graph representation was realized.Finally,the Naive Bayes algorithm was employed to match question templates,the structured querying was generated by matching result,and the answer was returned in the form of natural language to realize interactive question answering of fine-grained domain information.In the comparative experiments with Convolutional Neural Network(CNN)and BiLSTM,the F1 values of the proposed method in named entity recognition and relation extraction tasks are respectively 93.28%and 74.00%,which outperform the above neural network models.Experimental results show that the actual demands of bridge inspection information interaction are met by the proposed method.