MorphoLogic‘s Submission for the WMT 2009 Shared Task

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 155–159, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics MorphoLogic’s submission for the WMT 2009 Shared Task Attila Novák MorphoLogic Kardhegy utca 5, Budapest 1116, Hungary novak@morphologic.hu Abstract In this article, we describe the machine translation systems we used to create MorphoLogic’s submissions to the WMT09 shared Hungarian to English and English to Hungarian shared transla- tion tasks. We used our rule based MetaMorpho system to generate our pri- mary submission. In addition, we created a hybrid system where the Moses de- coder is used to rank translations or as- semble partial translations created by MetaMorpho. Our third system was a purely statistical morpheme based system for the Hungarian to English task. 1 Introduction This year, MorphoLogic submitted translations for the WMT09 shared Hungarian to English and English to Hungarian translation tasks. Our pri- mary submissions were translated by MetaMor- pho, a purely rule based machine translation sys- tem (Prószéky and Tihanyi, 2002). Since last year’s workshop we improved the Hungarian to English grammar of MetaMorpho by making more efficient the handling of certain structural ambiguities and making the way the system han- dles long sentences more robust. The way Metamorpho selects the translation to output is not optimal whether or not a full parse for the source sentence could be obtained by its parser.1 Thus we decided to experiment with a hybrid system where translations and partial translations produced by MetaMorpho are ranked or assembled by the Moses decoder (Koehn et al., 2007) using a target language model. 1 In the first case, simply the first translation is output instead of considering all possible translations and selecting the best, while in the second case, the algo- rithm that combines the partial translations does not check how well the target language side of the pieces fit together. In addition, we created a purely statistical morpheme based system (also using Moses) for the Hungarian to English task. However, results obtained with the latter setup have been clearly inferior in quality to those produced by the rule based system both in terms of BLEU score and subjective human judgment. 2 The MetaMorpho translation system MetaMorpho is a rule based system the architec- ture of which differs from that of most well- known rule based systems: it does not contain a separate transfer component. Its grammar oper- ates with pairs of patterns (context-free rules en- riched with features) that consist of one source pattern used during bottom-up parsing and one or more target patterns that are applied during top- down generation of the translation. The architec- ture of the grammar is completely homogeneous: the same formalism is used to represent general rules of grammar, more-or-less idiomatic phrases and fully lexicalized items, these differ only in the degree of underspecification. The translation of the parsed structures is al- ready determined during parsing the source lan- guage input. The actual generation of the target language representations does not involve …

Content truncated to 500 words