Towards Robust Knowledge Graph Embedding via Multi-Task Reinforcement Learning

Nowadays, Knowledge graphs (KGs) have been playing a pivotal role in AI-related applications. Despite the large sizes, existing KGs are far from complete and comprehensive. In order to continuously enrich KGs, automatic knowledge construction update mechanisms usually utilized, which inevitably bring plenty of noise. However, most graph embedding (KGE) methods assume that all triple facts correct, project both entities relations into low-dimensional space without considering noise conflicts. This will lead low-quality unreliable representations KGs. To this end, paper, we propose general multi-task reinforcement learning framework, can greatly alleviate noisy data problem. our exploit for choosing high-quality triples while filtering out ones. Also, take full advantage correlations among semantically similar relations, selection processes trained collective way with learning. Moreover, extend popular KGE models TransE, DistMult, ConvE RotatE proposed framework. Finally, experimental validation shows approach is able enhance provide more robust scenarios.