This paper describes the OpenNMT submissions to the WNGT 2020 efficiency shared task. We explore training and acceleration of Transformer models with various sizes that are trained in a teacher-student setup. We also present a custom and optimized C++ inference engine that enables fast CPU and GPU decoding with few dependencies. By combining additional optimizations and parallelization techniques, we create small, efficient, and high-quality neural machine translation models.

We present a system description of the OpenNMT Neural Machine Translation entry for the WNMT 2018 evaluation. In this work, we developed a heavily optimized NMT inference model targeting a high-performance CPU system. The final system uses a combination of four techniques, all of them lead to significant speed-ups in combination: (a) sequence distillation, (b) architecture modifications, (c) precomputation, particularly of vocabulary, and (d) CPU targeted quantization. This work achieves the fastest performance of the shared task, and led to the development of new features that have been integrated to OpenNMT and available to the community.

Knowledge distillation has recently been successfully applied to neural machine translation. It allows for building shrunk networks while the resulting systems retain most of the quality of the original model. Despite the fact that many authors report on the benefits of knowledge distillation, few have discussed the actual reasons why it works, especially in the context of neural MT. In this paper, we conduct several experiments aimed at understanding why and how distillation impacts accuracy on an English-German translation task. We show that translation complexity is actually reduced when building a distilled/synthesised bi-text when compared to the reference bi-text. We further remove noisy data from synthesised translations and merge filtered synthesised data together with original reference, thus achieving additional gains in terms of accuracy.

Training efficiency is one of the main problems for Neural Machine Translation (NMT). Deep networks need for very large data as well as many training iterations to achieve state-of-the-art performance. This results in very high computation cost, slowing down research and industrialisation. In this paper, we propose to alleviate this problem with several training methods based on data boosting and bootstrap with no modifications to the neural network. It imitates the learning process of humans, which typically spend more time when learning “difficult” concepts than easier ones. We experiment on an English-French translation task showing accuracy improvements of up to 1.63 BLEU while saving 20% of training time.