Semantic Consistency: A Local Subspace Based Method for Distant Supervised Relation Extraction

One fundamental problem of distant supervision is the noisy training corpus problem. In this paper, we propose a new distant supervision method, called Semantic Consistency, which can identify reliable instances from noisy instances by inspecting whether an instance is located in a semantically consistent region. Specifically, we propose a semantic consistency model, which first models the local subspace around an instance as a sparse linear combination of training instances, then estimate the semantic consistency by exploiting the characteristics of the local subspace. Experimental results verified the effectiveness of our method.