Hierarchical Phrase-Based MT at the Charles University for the WMT 2010 Shared Task

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 212–215, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Hierarchical Phrase-Based MT at the Charles University for the WMT 2010 Shared Task Daniel Zeman Charles University in Prague, Institute of Formal and Applied Linguistics (ÚFAL) Univerzita Karlova v Praze, Ústav formální a aplikované lingvistiky (ÚFAL) Malostranské náměstí 25, Praha, CZ-11800, Czechia zeman@ufal.mff.cuni.cz Abstract We describe our experiments with hier- archical phrase-based machine translation for WMT 2010 Shared Task. We provide a detailed description of our configuration and data so the results are replicable. For English-to-Czech translation, we experi- ment with several datasets of various sizes and with various preprocessing sequences. For the other 7 translation directions, we just present the baseline results. 1 Introduction Czech is a language with rich morphology (both inflectional and derivational) and relatively free word order. In fact, the predicate-argument struc- ture, often encoded by fixed word order in English, is usually captured by inflection (especially the system of 7 grammatical cases) in Czech. While the free word order of Czech is a problem when translating to English (the text should be parsed first in order to determine the syntactic functions and the English word order), generating correct in- flectional affixes is indeed a challenge for English- to-Czech systems. Furthermore, the multitude of possible Czech word forms (at least order of magnitude higher than in English) makes the data sparseness problem really severe, hindering both directions. There are numerous ways how these issues could be addressed. For instance, parsing and syntax-aware reordering of the source-language sentences can help with the word order differ- ences (same goal could be achieved by a reorder- ing model or a synchronous context-free grammar in a hierarchical system). Factored translation, a secondary language model of morphological tags or even a morphological generator are some of the possible solutions to the poor-to-rich translation is- sues. Our submission to the shared task should reveal where a pure hierarchical system stands in this jun- gle and what of the above mentioned ideas match the phenomena the system suffers from. Although our primary focus lies on English-to-Czech trans- lation, we also report the accuracy of the same system on moderately-sized corpora for the other three languages and seven translation directions. 2 The Translation System Our translation system belongs to the hierarchi- cal phrase-based class (Chiang, 2007), i.e. phrase pairs with nonterminals (rules of a synchronous context-free grammar) are extracted from sym- metrized word alignments and subsequently used by the decoder. We use Joshua, a Java-based open- source implementation of the hierarchical decoder (Li et al., 2009), release 1.1.1 Word alignment was computed using the first three steps of the train-factored-phrase- model.perl script packed with Moses2 (Koehn et al., 2007). This includes the usual combination of word clustering using mkcls3 (Och, 1999), two- way word alignment using GIZA++4 (Och and Ney, 2003), and alignment symmetrization using the grow-diag-final-and heuristic (Koehn et al., 2003). For language modeling we use the SRILM toolkit5 …

Content truncated to 500 words