基于BERT的遥感应用领域知识图谱自动构建技术

Automatic Construction Technology of Remote Sensing Application Domain Knowledge Graph Based on BERT

随着对地观测技术的快速发展,卫星遥感影像应用场景也日益丰富。然而,由于遥感数据应用门槛相对过高,专业知识储备要求相对严格,遥感应用领域知识缺乏系统化、结构化表示,严重制约了研究者和从业人员在特定行业中快速开发和应用遥感数据资源。文中将高质量的遥感应用领域学术论文和学术专著作为高质量知识库来源,使用引入损失函数Lenloss的BERT-BiLSTMCRF模型作为命名实体识别模型,抽取出卫星实体、传感器实体、分辨率实体、任务实体以及方法实体,进一步将同一篇学术论文摘要中不同实体之间的共现行为用于定义实体之间的关系,抽取出不同实体之间的关系。最后,利用图数据库Neo4j实现遥感应用领域知识图谱的存储管理。

With the rapid development of earth observation technology,satellite remote sensing image application scenarios are increasingly rich. However,due to the threshold in the field of remote sensing application is very high,this domain knowledge is still short of a systematic and structural representation,which limits its application for researchers and practitioners. This paper selects academic thesis and monograph as the source of domain knowledge,uses BERT-BiLSTM-CRF with Lenloss as named entity recognition model,which extracts satellite entity,sensor entity,resolution entity,task entity and method entity. The co-occurrence of different entities in the same paper is defined as relationship,which is relation extraction between named entities. Finally,Neo4 j is used to realize the storage and management of remote sensing application domain knowledge graph.