Continual Learning of Neural Machine Translation within Low Forgetting Risk Regions

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1707 - 1718 December 7-11, 2022 ©2022 Association for Computational Linguistics Continual Learning of Neural Machine Translation within Low Forgetting Risk Regions Shuhao Gu1,2∗, Bojie Hu3, Yang Feng1,2† 1 Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) 2 University of Chinese Academy of Sciences 3 Tencent Minority-Mandarin Translation, Beijing, China {gushuhao19b,fengyang}@ict.ac.cn, bojiehu@tencent.com Abstract This paper considers continual learning of large-scale pretrained neural machine transla- tion model without accessing the previous train- ing data or introducing model separation. We argue that the widely used regularization-based methods, which perform multi-objective learn- ing with an auxiliary loss, suffer from the mis- estimate problem and cannot always achieve a good balance between the previous and new tasks. To solve the problem, we propose a two- stage training method based on the local fea- tures of the real loss. We first search low forget- ting risk regions, where the model can retain the performance on the previous task as the pa- rameters are updated, to avoid the catastrophic forgetting problem. Then we can continually train the model within this region only with the new training data to fit the new task. Specif- ically, we propose two methods to search the low forgetting risk regions, which are based on the curvature of loss and the impacts of the parameters on the model output, respectively. We conduct experiments on domain adapta- tion and more challenging language adaptation tasks, and the experimental results show that our method can achieve significant improve- ments compared with several strong baselines. 1 Introduction The current large-scale pretrained neural machine translation (NMT) models, such as GMNMT (John- son et al., 2017), mBART50-nn (Tang et al., 2020), and M2M-100 (Fan et al., 2021), are generally trained with large amounts of data from different domains and languages, so the model can learn good semantic representation and mapping relation- ship at the same time. Based on such models, we hope that they can continually acquire new knowl- edge from new translation tasks, e.g., new domains ∗The work was done when the first author was an intern at Tencent. †Corresponding author: Yang Feng. Reproducible code: https://github.com/ictnlp/LFR-NMT. Figure 1: An illustration to indicate the difference be- tween the regularization-based method and our method. R denotes the task loss. θ denotes the model parame- ters. → and ← denote the parameter update direction during continual learning. [θmin, θmax] denotes the low forgetting risk regions. and languages, while preserving the previously learned knowledge, i.e., continual learning (CL). However, catastrophic forgetting is the biggest chal- lenge of continual learning, which means the model will forget the previously learned knowledge while learning new knowledge (Gu and Feng, 2020). One solution to avoid forgetting is to mix the previous and the new training data to train the model jointly. However, considering that the training data of the large-scale pre-trained NMT model is usually very large, this method will bring more training con- sumption …

Content truncated to 500 words