Improving distant supervision relation extraction with entity-guided enhancement feature

Selective attention in distant supervision extraction relation is advantageous to deal with incorrectly labeled sentences in a bag, but it does not help in cases where many sentence bags consist of only one sentence. To resolve the deficiencies, we propose an entity-guided enhancement feature neural network for distant supervision relation extraction. We discover that key relation features are typically found in both significant words and phrases, which can be captured by entity guidance. We first develop an entity-directed attention that measures the relevance between entities and two levels of semantic units from word and phrase to capture reliable relation features, which are used to enhance the entity representations. Furthermore, two multi-level augmented entity representations are transformed to a relation representation via a linear layer. Then we adopt a semantic fusion layer to fuse multiple semantic representations such as the sentence representation encoded by piecewise convolutional neural network, two multi-level augmented entity representations, and the relation representation to get final enhanced sentence representation. Finally, with the guidance of the relation representations, we introduce a gate pooling strategy to generate a bag-level representation and address the one-sentence bag problem occurring in selective attention. Extensive experiments demonstrate that our method outperforms the state-of-the-art methods.