基于机器阅读理解的行车故障诊断知识抽取

Knowledge extraction of crane fault diagnosis based on machine reading comprehension

行车故障调查单是对行车故障诊断过程的文本记录,基于这些历史记录构建知识图谱可以更好地支持行车故障诊断智能化。由于该语料具有实体嵌套、实体跨度大、关系重叠等特点,传统的命名实体识别和关系抽取模型难以对其进行有效的知识抽取。针对语料中存在的实体嵌套和长实体识别问题,本文提出了一种融合强化学习的机器阅读理解模型,以问答形式进行实体识别,以指针网络进行解码;对于语料中存在的关系重叠问题,将关系抽取分为先识别主体再识别客体的两阶段,将不同实体对的关系抽取进行隔离。实验结果表明,基于机器阅读理解的方法在行车故障诊断领域的知识抽取上具有较好的性能,可以有效支持领域知识图谱构建。

The crane fault investigation form is a text record of the crane fault diagnosis process.Constructing a knowledge graph based on these historical records can better support intelligent crane fault diagnosis.However,due to the characteristics of nested entities,large entity spans,and overlapping relations in this corpus,traditional named entity recognition and relationship extraction models are unable to perform effective knowledge extraction.To address the problems of nested entities and long entity recognition,this paper proposed a machine reading comprehension model fused with reinforcement learning.The model performed entity recognition in a question-answer format and decoded the output using a pointer network.For the problems of overlapping relations,relationship extraction was divided into two stages:first recognizing the subject and then recognizing the object,to isolate the relationship extraction of different entity pairs.Experiments show that the machine reading comprehension method has good performance in knowledge extraction for crane fault diagnosis,and can effectively support the construction of domain knowledge graphs.