Improved Induction Network with Hierarchical Self-attention Scheme for Few-Shot Relation Classification

The vast majority of present techniques for relation classification rely heavily on large-scale labelled datasets, but when the training data is insufficient, these methods are not available. Intuitively, we can regard relation classification as a few-shot learning task when the data is hard to collect. In this paper, we present an advanced induction network with a hierarchical self-attention scheme to effectively tackle the issue of few-shot relation classification. We utilize attention-based dynamic routing algorithm to learn a query-related class representation. Besides, our hierarchical self-attention scheme, including instance-level and task-level attention, can adjust support set and capture high-level information between support set and query set. The results of our extensive experiment on FewRel dataset show that our model is adaptable to task and outperforms the state-of-the-art models.