Distantly supervised relation extraction is a powerful learning method to recognize relations of entity pairs. However, wrong label problem is inevitable among large-scale training data. In this work we propose a hierarchical attention neural network to effectively alleviate the impact of noise instances. Moreover under distantly supervised scenario, connections and dependencies widely appear among relation classes, which we call class interactions. Previous end-to-end methods that considered the relations as independent failed to make use of these interactions. To better utilize these important interactions, we propose a soft target as training objective to learn class relationships jointly. Experiments show that our model outperforms state-of-the-art methods.
CITATION STYLE
Huang, K., Li, S., & Chen, G. (2017). Aggregating Class Interactions for Hierarchical Attention Relation Extraction. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 10635 LNCS, pp. 551–561). Springer Verlag. https://doi.org/10.1007/978-3-319-70096-0_57
Mendeley helps you to discover research relevant for your work.