In a relation classification task, few-shot learning is an effective method when the number of training instances decreases. The prototypical network is a few-shot classification model that generates a point to represent each class, and this point is called a prototype. The mean is used to select prototypes for each class from a support set in a prototypical network. This method is fixed and static, and will lose some information at the sentence level. Therefore, we treat the mean selection as a special attention mechanism, then we expand the mean selection to dynamic prototype selection by fusing a self-attention mechanism. We also propose a query-attention mechanism to more accurately select prototypes. Experimental results on the FewRel dataset show that our model achieves significant and consistent improvements to baselines on few-shot relation classification.
CITATION STYLE
Wu, L., Zhang, H. P., Yang, Y., Liu, X., & Gao, K. (2020). Dynamic Prototype Selection by Fusing Attention Mechanism for Few-Shot Relation Classification. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 12033 LNAI, pp. 431–441). Springer. https://doi.org/10.1007/978-3-030-41964-6_37
Mendeley helps you to discover research relevant for your work.