Sentence-aware Adversarial Meta-Learning for Few-Shot Text Classification

Proceedings of the 29th International Conference on Computational Linguistics, pages 4844–4852 October 12–17, 2022. 4844 Sentence-aware Adversarial Meta-Learning for Few-Shot Text Classification Suhe Wang , Xiaoyuan Liu , Bo Liu , Diwen Dong College of Computer National University of Defense Technology Changsha, China wangsuhe16,xyliu,KyLe.liu,ddw_bak@nudt.edu.cn Abstract Meta-learning has emerged as an effective ap- proach for few-shot text classification. How- ever, current studies fail to realize the impor- tance of the semantic interaction between sen- tence features and neglect to enhance the gen- eralization ability of the model to new tasks. In this paper, we integrate an adversarial net- work architecture into the meta-learning system and leverage cost-effective modules to build a novel few-shot classification framework named SaAML. Significantly, our approach can ex- ploit the temporal convolutional network to encourage more discriminative representation learning and explore the attention mechanism to promote more comprehensive feature expres- sion, thus resulting in better adaptation for new classes. Through a series of experiments on four benchmark datasets, we demonstrate that our new framework acquires considerable su- periority over state-of-the-art methods in all datasets, increasing the performance of 1-shot classification and 5-shot classification by 7.15% and 2.89%, respectively. 1 Introduction Deep learning usually relies on a vast amount of labeled examples to accomplish tasks. However, this requirement is problematic for few-shot text classification when only a few examples are avail- able in novel classes, which leads to poor model generalization for new tasks. Motivated by the fact that humans can quickly recognize new knowledge after learning a few examples, few-shot learning is becoming a hot research topic. Early studies exploit data augmentation and reg- ularization procedures (Salamon and Bello, 2017) to deal with the overfitting due to data sparse- ness. More recent research mainly falls into two approaches: (1) transfer-learning based methods (Pan et al., 2019; Gupta et al., 2020), which trans- fer and propagate knowledge attained from the source domain to classify unseen examples in the target domain. (2) meta-learning based methods (Tong, 2019; Sun et al., 2019), which learn knowl- edge by repeating lots of meta-tasks in a training episode manner and leverage knowledge extracted to briskly predict new samples during meta-testing. Specifically, Bao et al. (Bao et al., 2019) incor- porated distributional signatures of words into the meta-learning framework to make impressive per- formance. Han et al. (Han et al., 2021) first pro- posed to introduce an adversarial domain adapta- tion network to strengthen the generalization ability of the meta-learning system. Despite the remarkable progress of few-shot classification approaches (Geng et al., 2019; Bao et al., 2019; Han et al., 2021), most existing meta- learning models still suffer tough challenges: ex- ample diversity. Even examples in the same class have various representations, bringing about the difficulty of extracting generic features based on a few training examples and the urgent demand for strong adaptability of the model to new tasks. In this paper, we propose a straightforward but remarkably powerful framework to cope with the above challenges. We deploy an adversarial net- work architecture to …

Content truncated to 500 words