The ModernMT Project

The ModernMT Project Nicola Bertoldi Davide Caroselli MMT Srl - Trento, Italy www.modernmt.eu Marcello Federico Abstract This short presentation introduces Mod- ernMT: an open-source project 1 that in- tegrates real-time adaptive neural machine translation into a single easy-to-use prod- uct. 1 Neural Machine Translation Neural machine translation (MT) technology has been widely adopted by the translation industry, to produce ready-to-use drafts or high quality trans- lations via post-editing. Deployed engines are ei- ther generic, such as Google Translate, or cus- tomized, like the MT@EC engine, which is partic- ularly suited to the EU policy documents. Generic MT works well on average, but it can be bad at handling domain specific terminology or style. On the other hand, custom MT can definitely be more accurate but does not scale. Actually, for the trans- lation industry the granularity of a domain can be very fine – i.e. specific customers might use ter- minology in their own original way – thus, ending up with thousands of domains. Customization be- comes quickly unfeasible, for both computational and infrastructure costs. This is where ModernMT comes in! 2 Adaptive NeuralMT ModernMT is a new open-source MT software that consolidates the current state-of-the-art MT tech- nology into a single and easy-to-use product. Mod- ernMT adapts to the context in real-time and is ca- pable of learning from and evolving through inter- action with users, with the final aim of increasing MT-output utility for the translator in a real profes- sional environment. This is achieved by augment- ing a generic neural MT system with an internal c⃝ 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1https://github.com/ModernMT/MMT. dynamic memory, storing all available user transla- tion memories (TMs). This process is completely transparent when users work with a CAT tool, like MateCat (Federico et al., 2014). The ModernMT memory is kept in sync with the user TMs. When ModernMT receives a translation query, it quickly analyses its context, recalls from its memory the most related translation examples, and instantly adapts its neural network to the query. 3 Performance Our adaptation approach (Farajian et al., 2017) has proven to deliver the translation quality of cus- tomized machine translation at the maintenance costs of a generic system. Thanks to careful op- timization, the whole adaptation process, which is run for every sentence, just takes a fraction of sec- ond. Further, experiments have also shown that our context-based and memory-based adaptation method impacts positively on the translation of ter- minology. 4 Software as a Service ModernMT is also provided as a SaaS solution for enterprises. Benefits include cutting-edge innova- tions, baseline models trained on billions of words of premium data, support for nine language pairs (and more to come). Finally, professional transla- tors can instead benefit from all the advantages of ModernMT in MateCat2, a free CAT tool that is perfectly integrated with our service. References Federico M. et al. 2014. The MATECAT Tool, Proc. COLING, Dublin, Ireland. Farajian M. A. et …

Content truncated to 500 words