The QT21 Combined Machine Translation System for English to Latvian

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 348–357 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics The QT21 Combined Machine Translation System for English to Latvian Jan-Thorsten Peter1, Hermann Ney1, Ondˇrej Bojar2, Ngoc-Quan Pham3, Jan Niehues 3, Alex Waibel3, Franck Burlot4, Franc¸ois Yvon4, M¯arcis Pinnis5, Valters ˇSics5, Joost Bastings6, Miguel Rios6, Wilker Aziz6, Philip Williams7, Fr´ed´eric Blain8, Lucia Specia8 1RWTH Aachen University, Aachen, Germany 2Charles University, Prague, Czech Republic 3Karlsruhe Institute of Technology, Karlsruhe, Germany 4LIMSI, CNRS, Universit´e Paris Saclay, 91 403 Orsay, France 5Tilde, Riga, Latvia 6University of Amsterdam, Amsterdam, Netherlands 7University of Edinburgh, Edinburgh, UK 8University of Sheffield, Sheffield, UK Abstract This paper describes the joint submis- sion of the QT21 projects for the English→Latvian translation task of the EMNLP 2017 Second Conference on Ma- chine Translation (WMT 2017). The sub- mission is a system combination which combines seven different statistical ma- chine translation systems provided by the different groups. The systems are combined using either RWTH’s system combination approach, or USFD’s consensus-based system- selection approach. The final submission shows an improvement of 0.5 BLEU compared to the best single system on newstest2017. 1 Introduction Quality Translation 21 (QT21) is a European ma- chine translation research project with the aim of substantially improving statistical and machine learning based translation models for challenging languages and low-resource scenarios. Members of the QT21 project have jointly built a combined statistical machine translation system, in order to achieve high-quality machine transla- tion from English into Latvian. Core components of the QT21 combined sys- tem for the WMT 2017 shared task for ma- chine translation of news 1 are seven individual 1http://www.statmt.org/wmt17/ translation-task.html English→Latvian translation engines which have been set up by different project partners. The outputs of all these individual engines are combined using the system combination approach as implemented in Jane, RWTH’s open source sta- tistical machine translation toolkit (Freitag et al., 2014a). The Jane system combination is a mature implementation which previously has been suc- cessfully employed in other collaborative projects and for different language pairs (Peter et al., 2016; Freitag et al., 2013, 2014b,c). As an alternative way of combining our sys- tems, all outputs have been merged as the form of a n-best list and a consensus-based system-selection applied to obtain as best translation hypothesis the candidate that is most similar to the most likely translations amongst those systems. 2 Preprocessing The training data was pre-processed using a cus- tom language-specific tokeniser and the Moses truecaser (truecase.perl). For tokenisation, we used the Tilde’s regular expression-based to- keniser for Latvian and English that takes into ac- count language-specific characteristics (e.g., ab- breviations, contractions, date, time, and numer- ical expressions, etc.) and non-translatable enti- ties (e.g., phone numbers, e-mail addresses, XML tags, URLs, file paths, various identifiers and codes, etc.). Only the first word in each sentence was truecased. The data (backtranslation included) is fur- ther cleaned using a simple language identifier from Shuyo (2010). We simply removed sentence pairs whose …

Content truncated to 500 words