New Language Pairs in TectoMT

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 98–104, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. New Language Pairs in TectoMT Ondˇrej Dušek,∗ Luís Gomes,‡ Michal Novák,∗ Martin Popel,∗ and Rudolf Rosa∗ ∗Charles University in Prague, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics {odusek,mnovak,popel,rosa}@ufal.mff.cuni.cz ‡University of Lisbon, Faculty of Sciences, Department of Informatics luis.gomes@di.fc.ul.pt Abstract The TectoMT tree-to-tree machine transla- tion system has been updated this year to support easier retraining for more transla- tion directions. We use multilingual stan- dards for morphology and syntax annota- tion and language-independent base rules. We include a simple, non-parametric way of combining TectoMT’s transfer model outputs. We submitted translations by the English- to-Czech and Czech-to-English TectoMT pipelines to the WMT shared task. While the former offers a stable performance, the latter is completely new and will require more tuning and debugging. 1 Introduction The TectoMT tree-to-tree machine translation (MT) system (Žabokrtský et al., 2008) has been competing in WMT translation tasks since 2008 and has seen a number of improvements. Un- til now, the only supported translation direction was English to Czech. This year, as a part of the QTLeap project,1 we have enhanced TectoMT and its underlying natural language processing (NLP) framework, Treex (Popel and Žabokrtský, 2010), to support more language pairs. We simpli- fied the training pipeline to be able to retrain the translation models faster, and we use abstracted language-independent rules with the help of Inter- set (Zeman, 2008) where possible. Together with our partners on the QTLeap project, we have implemented translation systems for other language pairs (English to and from Dutch, Spanish, Basque, and Portuguese) which are not part of WMT shared Translation Task this year. However, we were also able to submit the results of a newly built Czech-English translation 1http://qtleap.eu system in the shared task. The performance of the current version leaves a lot of room for improve- ment, but proves the potential of TectoMT for dif- ferent language pairs. The original TectoMT system for English- Czech translation has seen just small changes, e.g., adding specialized translation models for se- lected pronouns (Novák et al., 2013a; Novák et al., 2013b) and fine-tuning of a handful of rules. Therefore, its performance is virtually identical to that of the last year’s version. This paper is structured as follows: in Section 2, we introduce the TectoMT basic architecture. In Section 3, we describe the improvements to Tec- toMT that were added for an easier support of new language pairs. Section 4 then details the Czech-to-English TectoMT system submitted to WMT15. We discuss TectoMT’s performance in the task and examine the most severe error sources in Section 5. Section 6 then concludes the paper. 2 The TectoMT Translation System TectoMT (Žabokrtský et al., 2008) is a tree-to- tree MT system system consisting of an analysis- transfer-synthesis pipeline, with transfer on the level of deep syntax. It is based on the Prague Tec- togrammatics theory (Sgall et al., 1986) and dis- tinguishes …

Content truncated to 500 words