English-Czech MT in 2008

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 125–129, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics English-Czech MT in 2008 ∗ Ondˇrej Bojar, David Mareˇcek, V´aclav Nov´ak, Martin Popel, Jan Pt´aˇcek, Jan Rouˇs, Zdenˇek ˇZabokrtsk´y Charles University, Institute of Formal and Applied Linguistics Malostransk´e n´am. 25, Praha 1, CZ-118 00, Czech Republic {bojar,marecek,novak,ptacek,zabokrtsky}@ufal.mff.cuni.cz {popel,jan.rous}@matfyz.cz Abstract We describe two systems for English-to- Czech machine translation that took part in the WMT09 translation task. One of the systems is a tuned phrase-based system and the other one is based on a linguisti- cally motivated analysis-transfer-synthesis approach. 1 Introduction We participated in WMT09 with two very dif- ferent systems: (1) a phrase-based MT based on Moses (Koehn et al., 2007) and tuned for English→Czech translation, and (2) a complex system in the TectoMT platform (ˇZabokrtsk´y et al., 2008). 2 Data 2.1 Monolingual Data Our Czech monolingual data consist of (1) the Czech National Corpus (CNC, versions SYN200[056], 72.6%, Kocek et al. (2000)), (2) a collection of web pages downloaded by Pavel Pecina (Web, 17.1%), and (3) the Czech mono- lingual data provided by WMT09 organizers (10.3%). Table 1 lists sentence and token counts (see Section 2.3 for the explanation of a- and t- layer). Sentences 52 M with nonempty t-layer 51 M a-nodes (i.e. tokens) 0.9 G t-nodes 0.6 G Table 1: Czech monolingual training data. ∗ The work on this project was supported by the grants MSM0021620838, 1ET201120505, 1ET101120503, GAUK 52408/2008, MˇSMT ˇCR LC536 and FP6-IST-5-034291-STP (EuroMatrix). 2.2 Parallel Data As the source of parallel data we use an internal release of Czech-English parallel corpus CzEng (Bojar et al., 2008) extended with some additional texts. One of the added sections was gathered from two major websites containing Czech sub- titles to movies and TV series1. The matching of the Czech and English movies is rather straight- forward thanks to the naming conventions. How- ever, we were unable to reliably determine the se- ries number and the episode number from the file names. We employed a two-step procedure to au- tomatically pair the TV series subtitle files. For every TV series: 1. We clustered the files on both sides to remove duplicates 2. We found the best matching using a provi- sional translation dictionary. This proved to be a successful technique on a small sample of manually paired test data. The process was facilitated by the fact that the correct pairs of episodes usually share some named entities which the human translator chose to keep in the original English form. Table 2 lists parallel corpus sizes and the distri- bution of text domains. English Czech Sentences 6.91 M with nonempty t-layer 6.89 M a-nodes (i.e. tokens) 61 M 50 M t-nodes 41 M 33 M Distribution: [%] [%] Subtitles 68.2 Novels 3.3 Software Docs 17.0 Commentaries/News 1.5 EU (Legal) Texts 9.5 Volunteer-supplied 0.4 Table 2: Czech-English data sizes and sources. 1www.opensubtitles.org and titulky.com 125 2.3 Data Preprocessing using TectoMT platform: Analysis and Alignment As …

Content truncated to 500 words