A DYNAMIC REPRESENTATION FRAMEWORK FOR UNCERTAIN KNOWLEDGE GRAPH

For now there is no unified standard for how to apply embedding models on deterministic knowledge graphs to uncertain knowledge inference. The computing mechanism for fact confidence is also to be improved. A framework named BFUKG (better framework for uncertain knowledge graph) is proposed to extend mainstream deterministic embedding models to uncertain graphs. Value intervals of deterministic model score functions were adjusted as well as the loss function design combining confidence and rules with the help of a slack co-efficient. In the meantime, the generation process of negative triples was improved while a computing and updating mechanism for confidence scores was put forward. Framework performance was verified on tasks including tail entity prediction and confidence evaluation on mainstream datasets. The results show that the performance of the proposed framework has a certain improvement compared with mainstream models.