Knowledge-Based Diverse Feature Transformation for Few-Shot Relation Classification

Yubao Tang, Zhezhou Li, Cong Cao, Fang Fang, Yanan Cao, Yanbing Liu, Jianhui Fu

Abstract: Few-shot relation classification is to classify novel relations having seen only a few training samples. We find it is unable to learn comprehensive relation features with information deficit caused by the scarcity of samples and lacking of significant distinguishing features. Existing methods ignore the latter problem. What’s worse, while there is a big difference between the source domain and the target domain, the generalization performance of existing methods is poor. And existing methods can not solve all these problems. In this paper, we propose a new model called Knowledge-based Diverse Feature Transformation Prototypical Network (KDFT-PN) for information deficit, lacking of significant distinguishing features and weak generalization ability. To increase semantic information, KDFT-PN introduces the information of knowledge base to fuse with the sample information. Meanwhile, we propose a novel Hierarchical Context Encoder based on prototypical network, which can enhance semantic interaction and improve cross-domain generalization ability. Moreover, this method has been evaluated on cross-domain and same-domain datasets. And experimental results are comparable with other state-of-the-art methods.