mtrain: A Convenience Tool for Machine Translation

mtrain: A Convenience Tool for Machine Translation Samuel Läubli* and Mathias Müller* and Beat Horat and Martin Volk {laeubli,mmueller,horat,volk}@cl.uzh.ch Institute of Computational Linguistics University of Zurich Abstract We present mtrain, a convenience tool for machine translation. It wraps existing machine translation libraries and scripts to ease their use. mtrain is written purely in Python 3, well-documented, and freely available.1 Machine translation libraries usually focus on core model training, while data preparation and auto- matic evaluation are left to the user. This presents a barrier to experimental reproducibility, rapid pro- totyping, and entry to the field from neighbour- ing disciplines. In the spirit of the Experimen- tal Management System for Moses (Koehn, 2010), our tool is meant to automate these tasks. mtrain is designed to handle most aspects of a machine translation experiment: it manages pre- processing, model training, and automatic evalua- tion. Preprocessing involves automatically split- ting a data set into training, validation, and test sets; tokenization; casing; byte-pair encoding; and normalization. On top of these standard prepro- cessing steps, mtrain can also deal with inline XML markup and intelligently transfer XML tags to translations (Müller, 2017). Our tool provides training automation for statis- tical phrase-based models with Moses (Koehn et al., 2007) and neural RNN encoder-decoder mod- els with Nematus (Sennrich et al., 2017). Af- ter training, mtrain offers automatic evaluation of translation quality. It outputs the well-known BLEU, TER, and METEOR metrics (Clark et al., 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/ZurichNLP/mtrain ∗equal contribution 2011). Given a folder that contains trained mod- els, the separate component mtrans can be used to translate from files or standard input. All steps can be configured with config files or command line options, but default settings already lead to functional baseline systems, making it eas- ier for inexperienced users to use the tool. Going forward, we consider wrapping additional machine translation libraries that are native Python 3, such as Sockeye (Hieber et al., 2017). References Clark, Jonathan H, Chris Dyer, Alon Lavie, and Noah A Smith. 2011. Better hypothesis testing for statistical machine translation: Controlling for optimizer insta- bility. In Proceedings of ACL, pages 176–181. Hieber, Felix, Tobias Domhan, Michael Denkowski, David Vilar, Artem Sokolov, Ann Clifton, and Matt Post. 2017. Sockeye: a toolkit for neural machine translation. arXiv preprint, arXiv:0902.0885. Koehn, Philipp, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. 2007. Moses: Open source toolkit for statistical machine translation. In Pro- ceedings of ACL, pages 177–180. Koehn, Philipp. 2010. An experimental management system. The Prague Bulletin of Mathematical Lin- guistics, 94:87–96. Müller, Mathias. 2017. Treatment of markup in sta- tistical machine translation. In Proceedings of the Third Workshop on Discourse in Machine Transla- tion, pages 36–46. Sennrich, Rico, Orhan Firat, Kyunghyun Cho, Alexan- dra Birch, Barry Haddow, Julian Hitschler, Marcin Junczys-Dowmunt, Samuel Läubli, Antonio Valerio Miceli Barone, Jozef Mokry, and Maria Nadejde. 2017. Nematus: a …

Content truncated to 500 words