Online Learning over Time in Adaptive Neural Machine Translation

Proceedings of Recent Advances in Natural Language Processing, pages 411–420 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_047 411 Online Learning over Time in Adaptive Neural Machine Translation Thierry Etchegoyhen∗ David Ponce∗ Harritxu Gete Ugarte Victor Ruiz G´omez Vicomtech Foundation, Basque Research and Technology Alliance (BRTA) {tetchegoyhen,adponce,hgete,vruiz}@vicomtech.org Abstract Adaptive Machine Translation purports to dy- namically include user feedback to improve translation quality. In a post-editing scenario, user corrections of machine translation output are thus continuously incorporated into trans- lation models, reducing or eliminating repet- itive error editing and increasing the useful- ness of automated translation. In neural ma- chine translation, this goal may be achieved via online learning approaches, where network parameters are updated based on each new sample. This type of adaptation typically re- quires higher learning rates, which can affect the quality of the models over time. Alterna- tively, less aggressive online learning setups may preserve model stability, at the cost of reduced adaptation to user-generated correc- tions. In this work, we evaluate different on- line learning configurations over time, measur- ing their impact on user-generated samples, as well as separate in-domain and out-of-domain datasets. Results in two different domains indi- cate that mixed approaches combining online learning with periodic batch fine-tuning might be needed to balance the benefits of online learning with model stability. 1 Introduction Machine Translation (MT) quality has increased significantly in recent years, notably with the ad- vent of modern Neural Machine Translation (NMT) approaches (Bahdanau et al., 2015; Vaswani et al., 2017). Despite this progress, machine translated output requires post-editing in many cases, a pro- cess which is made more taxing when the same errors are repeated by MT systems segment after segment. To tackle this issue, adaptive approaches to ma- chine translation aim to incorporate user feedback, oftentimes in post-editing scenarios (Turchi et al., ∗Equal contribution. 2017), although on-the-fly adaptation is also rele- vant for interactive machine translation (Peris et al., 2017). In NMT, responsive model adaptation can be achieved via online learning approaches, where network parameters are updated based on each new sample of user-edited data. To perform this type of adaption from single data points, higher optimiser learning rates (LR) are typically required, which can affect the quality of the models over time. Al- ternatively, less aggressive online learning setups may preserve model stability, at the cost of reduced adaptation to user-generated corrections. In this work, we study the evolution of online learning over time in a post-editing scenario, to determine optimal configurations in terms of both adaptation to continuous user input and model sta- bility. For this purpose, we examine the behaviour of four different gradient-descent optimisers in two different domains, with varying learning rates, and evaluate their behaviour as the number of samples increases over time. To measure system responsiveness to user input over time, we evaluate online learning on dynam- ically increasing sets of samples formed by simu- lated user corrections. To determine model stability as online learning is performed, we also measure the quality of the MT models on static …

Content truncated to 500 words