OctaveMT: Putting Three Birds into One Cage

OctaveMT: Putting Three Birds into One Cage Juan A. Alonso Lucy Software Ibérica (ULG Group) Copèrnic 42-44 1r 08021 Barcelona, Spain juan.alonso@ulgroup.com Albert Llorens Lucy Software Ibérica (ULG Group) Copèrnic 42-44 1r 08021 Barcelona, Spain albert.llorens@ulgroup.com Abstract This product presentation describes the in- tegration of the three MT technologies currently used – rule-based (RBMT), Sta- tistical (SMT) and Neural (NMT) – into one scalable single platform, OctaveMT. MT clients can access all three types of MT engines, whether on a user specified basis or depending on several translation parameters (language-direction, domain, etc.) 1 Introduction Historically, Lucy Software and Services (a com- pany of the United Language Group) has been fo- cusing its development efforts on its RBMT sys- tem. However, during the last few years, we started to develop and use SMT technology and during the last months we have also been working on the NMT area. Our mid-term goal is to have an operational RBMT–NMT hybrid engine. The aim of this presentation is to introduce and describe the integration of all three MT technolo- gies into one single product platform, OctaveMT. 2 System Architecture The system architecture is depicted in Figure 1. The platform keystone is the LT Task Scheduler component, a portable and scalable task distribution system offering high performance for many kinds of services. It accepts translation requests from one or more MT Clients through a RESTful API. These translation requests are stored in the Task Pool component of the Task Scheduler. © 2018 The authors. This article is licensed under a Creative Commons 3.0 license, no derivative works, at- tribution, CC-BY-ND. The translation tasks are then handled by one or more MT engines. Each engine has an eServant component that monitors its activity; when it is idle, it fetches one request from the Task Pool. This task is then fed through the deformatter, the segmenter, the tokenizer and, finally, the en- gine dispatcher. The dispatcher sends the seg- mented text to the back-end engine type specified in the translation task (RBMT, SMT or NMT). Af- ter that, the translated text is sent back to the refor- matter, and finally delivered to the originator MT Client through the Task Scheduler. 3 Advantages of this Approach By re-using common sub-components for the three types of translation engines, tasks such as document format handling and conversion, which typically are a problem for raw SMT & NMT en- gines, can be properly handled. Additionally, this approach allows to use standard load-balancing techniques to build distributed high-performance MT infrastructures. P´erez-Ortiz, S´anchez-Mart´ınez, Espl`a-Gomis, Popovi´c, Rico, Martins, Van den Bogaert, Forcada (eds.) Proceedings of the 21st Annual Conference of the European Association for Machine Translation, p. 339 Alacant, Spain, May 2018.