WMRE: Enhancing Distant Supervised Relation Extraction with Word-level Multi-instance Learning and Multi-hierarchical Feature

Distant supervised relation extraction (DSRE) obtains large amounts of data cost-effectively by aligning knowledge base with natural texts but also brings noisy data. Existing methods deal with noise through multi-instance learning (MIL) with attention. However, these approaches typically use attention at sentence-level and above, while ignoring word-level information. Intuitively, words in the same sentence are also of different importance. However, effective methods for distinguishing word-level importance differences are still lacking. Because multiplying with weighted attention leads word embeddings to be shifted in vector space. Therefore, we proposed WMRE. Specifically, WMRE concatenates the embeddings of multiple sentences. Then it uses filtering attention to remove low-importance words at different levels to extract multi-hierarchical features. In WMRE, words are the smallest units of mining information, and filtering attention distinguishes word-level importance differences while avoiding the embedding shifts. Extensive experiments on two widely used datasets show that WMRE fully utilizes word-level information and achieves state-of-the-art (SOTA) on both datasets.