CUNI in WMT14: Chimera Still Awaits Bellerophon

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 195–200, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics CUNI in WMT14: Chimera Still Awaits Bellerophon Aleˇs Tamchyna, Martin Popel, Rudolf Rosa, Ondˇrej Bojar Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, Prague, Czech Republic surname@ufal.mff.cuni.cz Abstract We present our English→Czech and English→Hindi submissions for this year’s WMT translation task. For English→Czech, we build upon last year’s CHIMERA and evaluate several setups. English→Hindi is a new language pair for this year. We experimented with reverse self-training to acquire more (synthetic) parallel data and with modeling target-side morphology. 1 Introduction In this paper, we describe translation systems sub- mitted by Charles University (CU or CUNI) to the Translation task of the Ninth Workshop on Statis- tical Machine Translation (WMT) 2014. In §2, we present our English→Czech systems, CU-TECTOMT, CU-BOJAR, CU-DEPFIX and CU- FUNKY. The systems are very similar to our sub- missions (Bojar et al., 2013) from last year, the main novelty being our experiments with domain- specific and document-specific language models. In §3, we describe our experiments with English→Hindi translation, which is a translation pair new both to us and to WMT. We unsuccess- fully experimented with reverse self-training and a morphological-tags-based language model, and so our final submission, CU-MOSES, is only a basic instance of Moses. 2 English→Czech Our submissions for English→Czech build upon last year’s successful CHIMERA system (Bojar et al., 2013). We combine several different ap- proaches: • factored phrase-based Moses model (§2.1), • domain-adapted language model (§2.2), • document-specific language models (§2.3), • deep-syntactic MT system TectoMT (§2.4), • automatic post-editing system Depfix (§2.5). We combined the approaches in several ways into our four submissions, as made clear by Ta- ble 1. CU-TECTOMT is the stand-alone TectoMT translation system, while the other submissions are Moses-based, using TectoMT indirectly to pro- vide an additional phrase-table. CU-BOJAR uses a factored model and a domain-adapted language model; in CU-DEPFIX, Depfix post-processing is added; and CU-FUNKY also employs document- specific language models. CU-TECTOMT CU-BOJAR CU-DEPFIX CU-FUNKY TectoMT (§2.4) D D D D Factored Moses (§2.1) D D D Adapted LM (§2.2) D D D Document-specific LMs (§2.3) D Depfix (§2.5) D D Table 1: EN→CS systems submitted to WMT. 2.1 Our Baseline Factored Moses System Our baseline translation system (denoted “Base- line” in the following) is similar to last year – we trained a factored Moses model on the concatena- tion of CzEng (Bojar et al., 2012) and Europarl (Koehn, 2005), see Table 2. We use two fac- tors: tag, which is the part-of-speech tag, and stc, which is “supervised truecasing”, i.e. the surface form with letter case set according to the lemma; see (Bojar et al., 2013). Our factored Moses sys- tem translates from English stc to Czech stc | tag in one translation step. Our basic language models are identical to last year’s submission. We added an adapted language 195 Tokens [M] Corpus Sents [M] …

Content truncated to 500 words