MICE: a middleware layer for MT

MICE: Adapting MT through Middleware Joachim Van den Bogaert, Heidi Depraetere, Tom Vanallemeersch CrossLang Kerkstraat 106 9050 Gentbrugge Belgium {first.lastname}@crosslang.com Abstract The MICE project (2018–2020) will de- liver a middleware layer for improving the output quality of the eTranslation sys- tem of EC’s Connecting Europe Facility through additional services, such as do- main adaptation and named-entity recog- nition. It will also deliver a user portal, allowing for human post-editing. 1 Objectives The MICE project (Middleware for Customer eTranslation), which is funded by the CEF Telecom programme (Connecting Europe Facility) and runs from October 2018 to September 2020, delivers a middleware layer for the improvement of the eTranslation machine translation (MT) system. The latter is developed by DG Translation, supports all 24 official EU languages, and is provided by the CEF Automated Translation building block of DG CNECT as a service to Digital Service Infra- structures (DSIs) of the EC and to public administrations of Member States. The project consortium of MICE consists of two companies, CrossLang (coordinator) and Tilde, and two public organisations, NBN (Bureau for Standardisation, Belgium) and RIK (Centre of Registers and Information Systems, Estonia). The middleware layer consists of the follow- ing services: • domain adaptation; • terminology resolution; • named-entity recognition; • document filtering; • normalisation. © 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, at- tribution, CC-BY-ND. MICE also provides a human and automated post-editing (PE) environment for CEF eTransla- tion output, based on MateCat.1 This will help users to dynamically enhance the MT output and aggregate data for further system improvement. The tests in the project involve several lan- guages (English, Dutch, French, Estonian), do- mains (standards and e-Business/e-Land register information) and countries (Belgium, Estonia). Domain-specific neural MT systems will be made available by the project consortium. 2 Architecture MICE will expose its middleware layer for customisation through APIs and a user portal, in order to increase its impact and usability. Tasks will be performed in real-time or offline, depending on user preference. The input consists of text snippets (messages in plain text of maximally 5,000 characters) or full text documents (Microsoft Office, open document formats, etc.). The MICE architecture, which is shown in Figure 1, is compliant with the eDelivery building block. The MICE project will create a reference im- plementation for the automated translation of standards and e-Business/e-Land register infor- mation in Belgium and Estonia. It will be exten- sible in order to allow for future add-ons of MT- related services, such as automated domain de- tection or combination of MT systems. 3 Test results In both use cases (NBN, RIK), similar tests for domain adaptation were performed. We will illustrate them based on the NBN use case. 1 http://www.matecat.com In-domain English–French and English–Dutch MT systems were built by extracting and align- ing sentences from PDF files of NBN. In addi- tion, parallel synthetic data were created by backtranslating text from in-domain monolingual target language data, and other parallel text was obtained by scraping …

Content truncated to 500 words