Using Compositional Embeddings for Fact Checking

Unsupervised fact checking approaches for knowledge graphs commonly combine path search and scoring to predict the likelihood of assertions being true. Current approaches search for said metapaths in the discrete search space spanned by the input knowledge graph and make no use of continuous representations of knowledge graphs. We hypothesize that augmenting existing approaches with information from continuous knowledge graph representations has the potential to improve their performance. Our approach ESTHER searches for metapaths in compositional embedding spaces instead of the graph itself. By being able to explore longer metapaths, it can detect supplementary evidence for assertions being true that can be exploited by existing fact checking approaches. We evaluate ESTHER by combining it with 10 other approaches in an ensemble learning setting. Our results agree with our hypothesis and suggest that all other approaches can benefit from being combined with ESTHER by 20.65% AUC-ROC on average. Our code is open-source and can be found at https://github.com/dice-group/esther.