DaMSTF: Domain Adversarial Learning Enhanced Meta Self-Training for Domain Adaptation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1650–1668 July 9-14, 2023 ©2023 Association for Computational Linguistics DaMSTF: Domain Adversarial Learning Enhanced Meta Self-Training for Domain Adaptation Menglong Lu1†, Zhen Huang1†, Yunxiang Zhao2∗, Zhiliang Tian1∗, Yang Liu1 and Dongsheng Li1 1National Key Laboratory of Parallel and Distributed Computing, National University of Defense Technology, China 2 Beijing Institute of Biotechnology, China {lumenglong, huangzhen, tianzhiliang, liuyang12a, dsli}@nudt.edu.cn, zhaoyx1993@163.com Abstract Self-training emerges as an important research line on domain adaptation. By taking the model’s prediction as the pseudo labels of the unlabeled data, self-training bootstraps the model with pseudo instances in the target domain. However, the prediction errors of pseudo labels (label noise) challenge the per- formance of self-training. To address this problem, previous approaches only use reli- able pseudo instances, i.e., pseudo instances with high prediction confidence, to retrain the model. Although these strategies effectively reduce the label noise, they are prone to miss the hard examples. In this paper, we propose a new self-training framework for domain adap- tation, namely Domain adversarial learning en- hanced Self-Training Framework (DaMSTF). Firstly, DaMSTF involves meta-learning to es- timate the importance of each pseudo instance, so as to simultaneously reduce the label noise and preserve hard examples. Secondly, we de- sign a meta constructor for constructing the meta validation set, which guarantees the ef- fectiveness of the meta-learning module by im- proving the quality of the meta validation set. Thirdly, we find that the meta-learning mod- ule suffers from the training guidance vanish- ment and tends to converge to an inferior op- timal. To this end, we employ domain ad- versarial learning as a heuristic neural net- work initialization method, which can help the meta-learning module converge to a better op- timal. Theoretically and experimentally, we demonstrate the effectiveness of the proposed DaMSTF. On the cross-domain sentiment clas- sification task, DaMSTF improves the perfor- mance of BERT with an average of nearly 4%. 1 Introduction Domain adaptation, which aims to adapt the model trained on the source domain to the target domain, † contributed equally to this work ∗ corresponding author attracts much attention in Natural Language Pro- cessing (NLP) applications(Du et al., 2020; Chen et al., 2021; Lu et al., 2022). Since domain adapta- tion involves labeled data from the source domain and unlabeled data from the target domain, it can be regarded as a semi-supervised learning prob- lem. From this perspective, self-training, a classi- cal semi-supervised learning approach, emerges a prospective research direction on domain adapta- tion (Zou et al., 2019; Liu et al., 2021). Self-training consists of a series of loops over the pseudo labeling phase and model retraining phase. In the pseudo labeling phase, self-training takes the model’s prediction as the pseudo labels for the unlabeled data from the target domain. Based on these pseudo-labeled instances, self-training re- trains the current model in the model retraining phase. The trained model can be adapted to the target domain by repeating these two phases. …

Content truncated to 500 words