Domain Adaptation for MT: A Study with Unknown and Out-of-Domain Tasks

Domain Adaptation for MT: A Study with Unknown and Out-of-Domain Tasks Hoang Cuong Agolo, Inc. cuong.hoang@agolo.com Abstract Translation quality could degrade non- gracefully outside the desired domain for MT. Meanwhile, translation requests are often unknown and potentially out-of- domain in practice. This paper shows that having an ecosystem with a range of pre- trained domain-specific MT systems can reduce the effect: a translation task can be out of scope of most pre-trained MT systems, but a few others can be capable of handling the task. But how to obtain the best translation from an ecosystem for such translation requests? We contribute two frameworks to address the problem. Experiments show that our frameworks give the performance in the middle be- tween top rank MT systems with reason- ably large-scale ecosystems. 1 Introduction Translation models have been developed under the assumption that we know the domain at test time in advance, and the domain is strictly relevant to our training data. However, we inevitably will come across test data that is sampled from a different dis- tribution to our training data when using the mod- els in the wild. Another critical thing is that the domain of test data is often unknown in practice (e.g. Google Translate and Microsoft Translators receive translation requests from their users with- out knowing in advance their interests). We have not had a solution for this well-known problem yet. Machine Translation (MT) has been c⃝ 2019 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. advanced by new models, including using Neu- ral Machine Translation (NMT) instead of Sta- tistical Machine Translation (SMT). The hope is that a better translation model would improve the translation in all settings/situations. This, how- ever, is not true. Translation quality could degrade nongracefully outside the desired domain for both NMT and SMT. In fact, it has been known that NMT suffers even harder than SMT when the test data is out-of-domain (Koehn and Knowles, 2017; Chu and Wang, 2018). We also improve MT by using domain adaptation methods (i.e. improving translation system from having a small seed in- domain data such as system interpolation, instance weighting and data selection). In practice, this is not a thorough solution because we do not know the domain of user translation requests in advance. The contribution of this work is to pro- vide a simple, easy-and-fast-to-deploy, translation model-agnostic1 solution to the challenging prob- lem. Our approach is to construct an “ecosystem” with a range of pre-trained domain-specific MT systems, each specialized in a certain domain (e.g. Speech, Financial, Food, etc.). Our intuition is that having such an ecosystem could reduce the de- crease in translation quality for an outside domain. That is, an out-of-domain translation task can be out of scope of most pre-trained MT systems in the ecosystem. However, with the diversity of do- mains in a reasonably large ecosystem, we hope there is a chance to have certain pre-trained sys- tems in the ecosystem …

Content truncated to 500 words