船舶舾装设计经验知识图谱的自动构建方法

A Novel Automatic Construction Approach for Ship Outfitting Design Experiential Knowledge Graph

为解决船舶舾装设计经验知识图谱构建中标注语料较少、知识抽取工作量巨大的问题,提出一种知识图谱自动构建方法。采用自顶向下的构建逻辑,结合领域特点和专家知识,制定船舶舾装设计概念实体和实体间关系的数据模式。以自然语言处理预训练模型为基础,研究船舶舾装设计知识的实体识别、关系抽取模型设计和训练方法,提出了实体边界嵌入、实体遮罩、两阶段训练结合的改进措施。在设计质量案例集上的实验表明,和基准相比,实体识别的精确率提升了9.95%,关系抽取的精确率提升了14.4%,证明了文中自动抽取船舶舾装设计知识方法的可行性和有效性。以此为基础提出船舶舾装设计经验知识图谱的构建流程,可灵活地构建服务于语义关系计算等下游知识重用活动的知识图谱。

In order to solve the problems of tagging corpus shortage and the large workload of knowledge extraction in the construction of the ship outfitting design experiential knowledge graph,an automatic knowledge graph construction approach is proposed.Using top-down construction logic,combining domain text characteristics and expert knowledge,a scheme for ship outfitting design entities and entity relationships is formulated.Natural language processing is used to pre-train the model;designing and training methods of the entity recognition model and relation extraction model in ship outfitting design knowledge are developed.Improved measures combining entity boundary embedding,entity masking and two-stage training are proposed.Experiments on the ship design quality case set show that,comparing to the benchmark,the precision of entity recognition is increased by 9.95%and the precision of relation extraction is increased by 14.4%,which proves the feasibility and effectiveness of the proposed ship outfitting design knowledge extracting method.Based on this,the automatic construction process of the ship outfitting design knowledge graph is proposed,which can flexibly construct a knowledge graph that serves downstream knowledge reuse activities such as semantic relation computing.