Research And Construction Of Temporal Knowledge Graph For The Equity Information

Finance is the core of modern economy,in the modern financial system,the links between companies are more and more diverse and change frequently,resulting in rapid and hidden financial risk transmission.Faced with a large number of financial groups and the diverse links between them,how to effectively sort them out has become a challenge for the financial regulatory system.As the latest generation of knowledge engineering technology,knowledge graphs are semantic networks that model objective world knowledge and are used in several specialized fields.Building a knowledge graph for equity information can effectively sort out the equity relationships between companies and provide a working point for practitioners in the financial regulatory industry.The construction of knowledge graphs for equity information faces two main challenges.First,the text data of equity information is complex,and there are a large number of overlapping relationships and entities,so it is difficult to extract them;second,the knowledge of equity information is frequently updated and highly correlated with time,so it is very necessary to add time attributes and support incremental updates in the process of knowledge graph construction.At present,none of the financial domain knowledge graphs consider temporal attributes in the construction process,so the relevant construction methods are not applicable to equity information.In summary,this paper researches and designs a temporal knowledge graph construction solution for equity information,and the main work includes the following aspects.1.We propose a knowledge extraction method for equity-oriented information,which has two steps:Step 1:We extracted triples using the joint entity and relationship extraction model based on the multi-headed self-attentive mechanism proposed in this paper.This model can identify overlapping entities and relationships in the text,and avoid the problems of exposure bias and too rigid thresholds commonly found in mainstream joint extraction models.This model also introduces a word-level adaptive threshold to replace the global threshold for decision making.Step 2:We used a rule-based approach to expand the temporal attributes into quadruples by adding them to the effective triples,laying the foundation for the subsequent mapping work.2.Design and construction of a temporal knowledge graph for equity information.Abstracting companies and people as nodes,abstracting the equity association and employment relationship between them as edges,and adding time information to the attributes of nodes and edges.The construction scheme includes several aspects such as schema layer design,knowledge extraction,and storing knowledge and visualization with the help of graph database neo4j.The total storage currently contains more than 250,000 nodes and 340,000 relationships.3.An incremental update method for temporal knowledge graphs is proposed,specifically including a TransE-TAE model,a coincidence calculation model and an incremental update algorithm for the incremental update of the built knowledge graphs.