Multi-Agent Mutual Learning at Sentence-Level and Token-Level for Neural Machine Translation

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1715–1724 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1715 Multi-Agent Mutual Learning at Sentence-Level and Token-Level for Neural Machine Translation Baohao Liao Yingbo Gao Hermann Ney Human Language Technology and Pattern Recognition Group Computer Science Department RWTH Aachen University D-52056 Aachen, Germany baohao.liao@rwth-aachen.de {ygao|ney}@cs.rwth-aachen.de Abstract Mutual learning, where multiple agents learn collaboratively and teach one another, has been shown to be an effective way to dis- till knowledge for image classification tasks. In this paper, we extend mutual learning to the machine translation task and operate at both the sentence-level and the token-level. Firstly, we co-train multiple agents by us- ing the same parallel corpora. After conver- gence, each agent selects and learns its poorly predicted tokens from other agents. The poorly predicted tokens are determined by the acceptance-rejection sampling algorithm. Our experiments show that sequential mutual learn- ing at the sentence-level and the token-level improves the results cumulatively. Absolute improvements compared to strong baselines are obtained on various translation tasks. On the IWSLT’14 German-English task, we get a new state-of-the-art BLEU score of 37.0. We also report a competitive result, 29.9 BLEU score, on the WMT’14 English-German task. 1 Introduction Neural machine translation (NMT) has achieved significant progress over recent years (Sutskever et al., 2014; Bahdanau et al., 2015; Gehring et al., 2017; Vaswani et al., 2017; Edunov et al., 2018). Conventional training of the NMT models with hard targets limits the models’ generalization abil- ity (Szegedy et al., 2016; Pereyra et al., 2017). This has led to a rapid growth of research in developing more regularized models. Teacher-student (T/S) learning (Li et al., 2014; Hinton et al., 2015; Meng et al., 2019) is an effective method to handle this problem. It has been widely applied in many cases, e.g. model compression (Li et al., 2014; Hinton et al., 2015), domain adaptation (Li et al., 2017; Meng et al., 2018) and low-resource machine trans- lation (Chen et al., 2017). T/S learning is a strategy that trains a student model with both hard targets and soft posteriors produced by a pre-trained teacher model (Li et al., 2014). Because training with soft targets provides smoother output distribution, T/S learning could outperform the single model training (Li et al., 2014; Hinton et al., 2015; Meng et al., 2018). However, does a teacher always outperform a student? In order to evaluate the pros and cons of different models, we conduct experiments on two different architectures. Table 1 shows the trans- lation quality from various models. Arch1 out- performs Arch2 for the translation tense, whereas Arch2 outperforms Arch1 for certain word trans- lation. Besides, the same architecture but initial- ized differently also has tiny translation differences. This phenomenon shows that a certain architecture may not always be suitable to be a teacher. In this paper, we propose a two-step multi-agent mutual learning scheme, where one agent learns from other agents at the sentence-level as the first step, which we call sentence-wise …

Content truncated to 500 words