Evaluating Extreme Hierarchical Multi-label Classification

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 517 - 527 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Learn to Adapt for Generalized Zero-Shot Text Classification Yiwen Zhang, Caixia Yuan∗, Xiaojie Wang, Ziwei Bai, Yongbin Liu Beijing University of Posts and Telecommunications {garyzhang, yuancx, xjwang, bestbzw, liuyb} @bupt.edu.cn Abstract Generalized zero-shot text classification aims to classify textual instances from both previ- ously seen classes and incrementally emerg- ing unseen classes. Most existing methods generalize poorly since the learned parame- ters are only optimal for seen classes rather than for both classes, and the parameters keep stationary in predicting procedures. To ad- dress these challenges, we propose a novel Learn to Adapt (LTA) network using a variant meta-learning framework. Specifically, LTA trains an adaptive classifier by using both seen and virtual unseen classes to simulate a gen- eralized zero-shot learning (GZSL) scenario in accordance with the test time, and simul- taneously learns to calibrate the class proto- types and sample representations to make the learned parameters adaptive to incoming un- seen classes. We claim that the proposed model is capable of representing all prototypes and samples from both classes to a more consistent distribution in a global space. Extensive exper- iments on five text classification datasets show that our model outperforms several competi- tive previous approaches by large margins. The code and the whole datasets are available at https://github.com/Quareia/LTA. 1 Introduction Text classification plays an important role in many natural language processing (NLP) applications, such as question classification, news categorization, user intent classification and so on (Minaee et al., 2021). Although a wide variety of methods have been proved successful in supervised text classifica- tion, they often break down when applied to make predictions for incrementally emerging classes without labeled training data (Pourpanah et al., 2020). Unlike zero-shot learning (ZSL) that aims to classify unseen class instances at test time (Romera- Paredes and Torr, 2015; Wang et al., 2019), gener- alized zero-shot learning (GZSL), which we focus ∗Corresponding author. on in this work, aims to classify text samples from both previous seen and emerging novel classes. Since there is a strong bias towards seen classes (Xian et al., 2019a), GZSL is a more challenging yet critical problem. Previously methods mainly focus on transduc- tive approaches for generalized zero-shot text clas- sification. Rios and Kavuluru (2018) use a graph convolution network to enhance the unseen class label embeddings. Zhang et al. (2019) and Song et al. (2020) generate illusion feature embeddings for unseen classes based on side information, i.e., class-level attributes or text description. More re- cently, Ye et al. (2020) use reinforced self-training methods to leverage unlabeled data during training. With the assumption that no knowledge about unseen categories is available during the model learning phase, researchers resort to inductive ap- proaches to handle generalized zero-shot text clas- sification. ReCapsNet (Liu et al., 2019) uses a dimensional attention-based intent capsule network and constructs zero-shot class prototypes by simi- larity matrix transformation. …

Content truncated to 500 words