Interactive optimization of relation extraction via knowledge graph representation learning

Relation extraction is a vital task in constructing large-scale knowledge graphs, aiming to identify factual relations between entities from plain texts and generate triples. However, it is inevitable that a large amount of noise will be generated and should be given special attention; otherwise, they will seriously downgrade the performance of knowledge reasoning. In this paper, we propose a visual analytics system that facilitates automatic extraction and interactive optimization of relations between entities, enabling users to refine these extraction results with low confidence. First, a triple-based embedding method is designed to provide an overview of the triples by capturing the semantic similarity between entities and relations. Then, the contextual information in the embedding space is utilized to evaluate the correctness of triples and infer more probable relations for correction. Finally, a visual analysis system integrating the above method and multiple coordinated views is developed, enabling the higher-quality data corrected by users to assist in achieving iterative optimization of the relation extraction model in an interpretable way. Case studies based on real-world datasets and expert interviews further demonstrate the effectiveness of the system for effective analysis and exploration of the knowledge graph relation extraction.