From Learning-to-Match to Learning-to-Discriminate:Global Prototype Learning for Few-shot Relation Classification

From Learning-to-Match to Learning-to-Discriminate: Global Prototype Learning for Few-shot Relation Classification Fangchao Liu1,4,∗, Xinyan Xiao3, Lingyong Yan1,4, Hongyu Lin1, Xianpei Han1,2,† , Dai Dai3, Hua Wu3, Le Sun1,2 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3Baidu Inc., Beijing, China 4University of Chinese Academy of Sciences, Beijing, China {fangchao2017,lingyong2014,hongyu,xianpei,sunle}@iscas.ac.cn Abstract Few-shot relation classification has attracted great attention recently, and is regarded as an ef- fective way to tackle the long-tail problem in relation classification. Most previous works on few-shot relation classification are based on learning-to-match paradigms, which focus on learn- ing an effective universal matcher between the query and one target class prototype based on inner-class support sets. However, the learning-to-match paradigm focuses on capturing the sim- ilarity knowledge between query and class prototype, while fails to consider discriminative infor- mation between different candidate classes. Such information is critical especially when target classes are highly confusing and domain shifting exists between training and testing phases. In this paper, we propose the Global Transformed Prototypical Networks(GTPN), which learns to build a few-shot model to directly discriminate between the query and all target classes with both inner-class local information and inter-class global information. Such learning-to-discriminate paradigm can make the model concentrate more on the discriminative knowledge between all candidate classes, and therefore leads to better classification performance. We conducted exper- iments on standard FewRel benchmarks. Experimental results show that GTPN achieves very competitive performance on few-shot relation classification and reached the best performance on the official leaderboard of FewRel 2.0 1. Introduction Few-shot relation classification aims to build relation extractors with only a few instances. Different from supervised learning that requires large scale training data of target classes, few-shot learning-based approaches can effectively build relation extractors with only a few examples (i.e., the support set) of each target class. This property makes it very appealing in the real application, where the training data of new target class can be scarce. Previous few-shot relation classification approaches can be summarized into a learning-to-match paradigm. Specifically, as shown in Figure 1 (a), these methods try to learn a universal matcher be- tween the query and each target relation type, and measure their similarity separately to infer the type of queries. Along this line, Gao (2019a) propose the hybrid attention to attach different importance for each relation feature, and match the query to each relation type with the reweighted features. Ye (2019) performs a local matching and aggregation between the query instance and each relation type. Besides, Gao (2019b) proposes an instance-pair matcher for scoring the similarity between query and each rela- tion instance. Generally, the goal of these methods is to effectively obtain a query-class matcher with the limited given instances in the support set. However, because such learning-to-match based approaches focus on capturing the similarity informa- tion between query and classes, they are unable to consider discriminative knowledge between different candidate classes. This can significantly undermine the model performance when target classes are …

Content truncated to 500 words