Phrase-Based and Deep Syntactic English-to-Czech Statistical Machine Translation

Proceedings of the Third Workshop on Statistical Machine Translation, pages 143–146, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Phrase-Based and Deep Syntactic English-to-Czech Statistical Machine Translation ∗ Ondˇrej Bojar and Jan Hajiˇc Institute of Formal and Applied Linguistics ´UFAL MFF UK, Malostransk´e n´amˇest´ı 25 CZ-11800 Praha, Czech Republic {bojar,hajic}@ufal.mff.cuni.cz Abstract This paper describes our two contributions to WMT08 shared task: factored phrase-based model using Moses and a probabilistic tree- transfer model at a deep syntactic layer. 1 Introduction Czech is a Slavic language with very rich morphol- ogy and relatively free word order. The Czech morphological system (Hajiˇc, 2004) defines 4,000 tags in theory and 2,000 were actually seen in a big tagged corpus while the English Penn Treebank tagset contains just about 50 tags. In our parallel corpus (see below), the English vocabulary size is 148k distinct word forms but more than twice as big in Czech, 343k distinct word forms. When translating to Czech from an analytic lan- guage such as English, target word forms have to be chosen correctly to produce a grammatical sen- tence and preserve the expressed relations between elements in the sentence, e.g. verbs and their modi- fiers. This year, we have taken two radically different approaches to English-to-Czech MT. Section 2 de- scribes our setup of the phrase-based system Moses (Koehn et al., 2007) and Section 3 focuses on a sys- tem with probabilistic tree transfer employed at a deep syntactic layer and the new challenges this ap- proach brings. ∗The work on this project was supported by the grants FP6- IST-5-034291-STP (EuroMatrix), MSM0021620838, MˇSMT ˇCR LC536, and GA405/06/0589. 2 Factored Phrase-Based MT to Czech Bojar (2007) describes various experiments with factored translation to Czech aimed at improving target-side morphology. We use essentially the same setup with some cleanup and significantly larger target-side training data: Parallel data from CzEng 0.7 (Bojar et al., 2008), with original sentence-level alignment and tokeniza- tion. The parallel corpus was taken as a monolithic text source disregarding differences between CzEng data sources. We use only 1-1 aligned sentences. Word alignment using GIZA++ toolkit (Och and Ney, 2000), the default configuration as available in training scripts for Moses. We based the word align- ment on Czech and English lemmas (base forms of words) as provided by the combination of tag- gers and lemmatizers by Hajiˇc (2004) for Czech and Brants (2000) followed by Minnen et al. (2001) for English. We symmetrized the two GIZA++ runs us- ing grow-diag-final heuristic. Truecasing. We attempted to preserve meaning- bearing case distinctions. The Czech lemmatizer produces case-sensitive lemmas and thus makes it easy to cast the capitalization of the lemma back on the word form.1 For English we approximate the same effect by a two-step procedure.2 1We change the capitalization of the form to match the lemma in cases where the lemma is lowercase, capitalized (uc- first) or all-caps. For mixed-case lemmas, we keep the form intact. 2We first collect a lexicon of the most typical “shapes” for each word …

Content truncated to 500 words