Abstract
To construct a legal knowledge graph to support the intelligent computing in the domain of law, a conceptual framework of this domain needs to be set up first. This framework usually comprises domain-specific inter-related concepts, their relations, and their attributes, to achieve a comprehensive understanding of the subject domain. While some of the topics in the task of concept framework construction had been studied, such as the identification of hypernymy/hyponymy relations, other key parts in this task have not been fully investigated. In this work, we propose a system TraConcept, in which we use PLM to encode traffic concept pairs with context to alleviate the "lexical memorization" problem. In addition, we model this task as a multi-relation identification problem and use a Siamese Network with a tensor layer to solve this problem. Compared with state-of-the-art methods, our method is more effective in detecting multi-relations between Chinese traffic concepts and finding attributes of these concepts in a large corpus of traffic legal texts.