A Temporal Knowledge Graph Completion Method with Text Descriptions Enhancement for Bank Compliance Management

With technologies of knowledge graphs, the compliance management of banking business based on documents issued by financial regulatory agencies can help relevant personnel realize intelligent queries. During our research, it is found that there are still existing some undiscovered relationships between entities in the current KG. Meanwhile, the needs of bank compliance management for KG are constantly changing. Besides queries such as keyword retrieval, higher requirements are put forward, hoping to obtain more abundant queries with relevance inference. These require us to supplement the KG. For these, this paper designs a temporal knowledge graph completion method with text descriptions enhancement (TD-TKGC). TD-TKGC builds a temporal completion model with LSTM to learn time-aware structural embeddings and uses ALBERT to learn textual embeddings. In addition, TD-TKGC adopts the clustering-based negative samples construction strategy to provide more effective negative triples. Experiments on multiple real-world temporal KG datasets show that for MRR, Hits@l and Hits@3, TD-TKGC performs better on completion tasks.