A Comparison of Statistical and Neural MT in a Multi-Product and Multilingual Software Company - User Study

A Comparison of Statistical and Neural MT in a Multi-Product and Multilingual Software Company - User Study Nander Speerstra Machine Translation Researcher, Infor Baron van Nagellstraat 89 Barneveld, Netherlands Nander.Speerstra@infor.com Abstract Over the last 4 years, Infor has been im- plementing machine translation (MT) in its translation process. In this paper, the results of both statistical and neural MT projects are provide to give an insight in the advantages and disadvantages of MT use in a large company. We also offer a look into the future of MT within our com- pany and to strengthen the implementation of MT in our translation process. 1 Introduction In the last few years, we have seen a change of di- rection regarding machine translation approaches. In different domains, more research is being fo- cussed on neural machine translation (NMT) in comparison to phrase-based statistical machine translation: in both the research environment (Bo- jar et al., 2016) and commercial companies like Google (Wu et al., 2016) and Microsoft (Awadalla et al., 2018) NMT is increasingly important. In the context of commercial translations, the continuous improvement of (N)MT has not passed unnoticed. More and more language service providers (LSPs) are implementing machine trans- lation into their translation workflows and in ad- dition, translation teams in large companies are in- vesting in machine translation as part of their trans- lation processes. As a large global software development com- pany, Infor1 translates its products into many lan- guages. This paper summarizes the results of the c⃝ 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1http://www.infor.com/ investigations into the potential benefits of ma- chine translation for a company with many prod- ucts, many target languages and very different translation circumstances per product. This study consists of two main parts: SMT and NMT. First, we give a description of our experiments, after which the results of the experiments are described. Lastly, the results and impact on our company are discussed. We had 2 main goals for this user study: to find out the current importance of (S)MT in our com- pany and the potential benefits of moving to NMT in the future. These goals are discussed in Section 4. 2 Experiments 2.1 Background Infor is an enterprise software company that cur- rently markets more than 125 different products, translating any number of these into 49 separate languages. The translation process involves both internal translators (up to 15 languages) and LSPs. A visual representation of the MT workflow is presented in Figure 1. Once the documentation is finished by technical writers, the translatable files are pre-processed: sentences that have been translated in previous versions of the product are re-used to prevent re-translation of already trans- lated content. Subsequently, machine translation and an automatic post editing script is run to fix some of SMT’s errors. From here, the post-editing and translation are done by vendors or internal lin- guists, who also perform a quality check. For many products, both the …

Content truncated to 500 words