Zero-shot Text Classification via Reinforced Self-training

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3014–3024 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3014 Zero-shot Text Classification via Reinforced Self-training Zhiquan Ye1,2, Yuxia Geng1,2, Jiaoyan Chen4, Xiaoxiao Xu3, Suhang Zheng3, Feng Wang3, Jingmin Chen3, Jun Zhang3, Huajun Chen†1,2 1College of Computer Science, Zhejiang University 2AZFT Joint Lab of Knowledge Engine 3Alibaba Group 4Department of Computer Science, Oxford University {yezq,gengyx,huajunsir}@zju.edu.cn {jiaoyan.chen}@cs.ox.ac.uk, {xiaoxiao.xuxx}@alibaba-inc.com {suhang.zhengsh,wf135777,jingmin.cjm,zj157077}@alibaba-inc.com Abstract Zero-shot learning has been a tough problem since no labeled data is available for unseen classes during training, especially for classes with low similarity. In this situation, transfer- ring from seen classes to unseen classes is ex- tremely hard. To tackle this problem, in this paper we propose a self-training based method to efficiently leverage unlabeled data. Tradi- tional self-training methods use fixed heuris- tics to select instances from unlabeled data, whose performance varies among different datasets. We propose a reinforcement learn- ing framework to learn data selection strategy automatically and provide more reliable selec- tion. Experimental results on both benchmarks and a real-world e-commerce dataset show that our approach significantly outperforms previ- ous methods in zero-shot text classification. 1 Introduction Zero-shot learning (ZSL) is a challenging task as no labeled data is available for unseen classes dur- ing training. There are extensive works proposed in zero-shot image classification task. The main focus of these works is how to transfer knowl- edge from seen classes to unseen classes. To as- sociate unseen classes with seen classes, they usu- ally resort to semantic information such as visual attributes (Lampert et al., 2009), word embeddings of class names (Norouzi et al., 2013) and class hierarchy (Socher et al., 2013). For example, if the model has not seen any instances of “hump- back whale” in the training stage, it could still make predictions at testing stage since “humpback whale” is semantically close to “killer whale” and “blue whale” in the seen class set ∗, so the model is capable of transferring knowledge from seen †Corresponding Author. ∗This example is taken from awa2 dataset, https:// cvml.ist.ac.at/AwA2/. classes to unseen classes. These methods assume that semantically similar classes share similar im- age features, however, they may fail in the cases where classes share low similarities. This problem becomes even more salient in typ- ical NLP tasks such as text classification. For ex- ample, let us consider a 10-class emotion clas- sification task (Yin et al., 2019), in which the model is trained on class “sadness” while makes predictions on instances from class “joy”. No- tice that most emotions are relatively independent, which means the way we express certain emotion is pretty different from other emotions. As a result, for an unseen class we can hardly find a similar class in the seen class set. Transferring from seen classes to unseen classes can be extremely hard as matching patterns that can be shared among classes are rare. Essentially, ZSL methods aim to learn a match- ing model between feature space and semantic …

Content truncated to 500 words