DPNet: domain-aware prototypical network for interdisciplinary few-shot relation classification

As few-shot relation classification requires less data to train the neural network models, high costs related to data collection and labeling are eliminated. Compared with the traditional few-shot task, the interdisciplinary few-shot relation classification faces many semantic gap issues (e.g., common sense, subject terminology and insufficient migratable information), bringing the linguistic bias between the testing phase and the training phase. Previous methods mainly focus on the training phase itself and neglect the inherent syntactic and semantic heterogeneity in terms of gap issues. In this paper, we propose a Domain-Aware Prototypical Network (DPNet) to cast the interdisciplinary few-shot relation classification as a covariate shift task. First, the relational keyword masking mechanism, which masks the keywords with a certain probability, is adopted to learn the contextual structure information. Second, to reduce the influence of subject-specific terminology, the adaptive word mover's distance (WMD) is proposed to calculate the similarity among test samples by weighting each word in the sentence. Third, the knowledge distillation strategy integrates each category distribution characteristic into a prototypical network, which supplements additional information for metric learning in task-adaptive feature sub-space. We conduct various experiments on two widely used benchmarks (FewRel 2.0 and DDI-13 dataset). Experimental results show that our model outperforms related methods by a significant margin.