NICE: Neural Integrated Custom Engines

NICE: Neural Integrated Custom Engines Daniel Mar´ın Buj1, Daniel Ib´a˜nez Garc´ıa1, Zuzanna Parcheta1, Francisco Casacuberta2 daniel.marin@cdt.europa.eu {daniel.ibanez,zuzanna.parcheta}@ext.cdt.europa.eu fcn@prhlt.upv.es 1 Translation Centre for the Bodies of the European Union 2 PRHLT Research Center, Universitat Politecnica de Val`encia Abstract In this paper, we present a machine translation system implemented by the Translation Centre for the Bodies of the European Union. The main goal of this project is to create domain-specific machine translation engines to support machine translation services and applications for the Translation Centre’s clients. In this article, we explain the entire implementation process of NICE: Neural Integrated Custom Engines. We describe the problems identified and the solutions provided, and present the final results for different language pairs. Finally, we describe the work that will be done on this project in the future. 1 Project description Set up in 1994, the Translation Centre for the Bodies of the European Union (CdT) delivers an average of 750,000 pages a year to over 60 European Union institutions, agencies and bodies across Europe. It has grown steadily, hand in hand with an increasing number of official European Union (EU) languages. To meet the needs of its clients and to cope with very specialised fields and growing translation volumes, the CdT has decided to enhance its services with state-of-the-art technologies such as neural machine translation (NMT) (Wu, 2016; Castilho, 2017). The business goal of this project is to provide raw machine translation of source texts that c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. enable translators to produce final translations that are indistinguishable from human translations with less effort than it would take to produce the same translations from scratch (Jia, 2019). Also, we aim to create engines that are fully integrated into CdT’s translation management system and fine-tuned for specific needs, such as post-editing particular document types, which cannot be isachieved by existing systems. Finally, the purpose to keep maximum confidentiality in the inference process by assuring an adapted, on-premise infrastructure. In this work, we focus on two different domains: intellectual property (IP) and public health (PH). Although the scope of the project includes all 24 official EU languages, each domain has its requirements in terms of language coverage. Thus, we targeted specific pairs for the development phase of the engines, with English being the common language for all models. The practice adopted for the development of machine translation engines included extensive preprocessing of data. After such data preparation, a generic model (GEN) was trained using data from all available domains. Then, the generic model was fine-tuned with in-domain data (IND). After training the IND model, we tested it using fixed test sets, and with five standard metrics. high-qualityAfter the automatic evaluation against references, human translators assessed another set of representative samples by applying predefined metrics at a segment level, such as adequacy and fluency (Koehn, 2006), and by post-editing the raw output to measure the potential productivity gains (Levenshtein distance (Marg, 2016)). …

Content truncated to 500 words