Two-step translation with grammatical post-processing

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 426–432, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Two-step translation with grammatical post-processing∗ David Mareˇcek, Rudolf Rosa, Petra Galuˇsˇc´akov´a and Ondˇrej Bojar Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, Prague {marecek,rosa,galuscakova,bojar}@ufal.mff.cuni.cz Abstract This paper describes an experiment in which we try to automatically correct mistakes in grammatical agreement in English to Czech MT outputs. We perform several rule-based corrections on sentences parsed to dependency trees. We prove that it is possible to improve the MT quality of majority of the systems par- ticipating in WMT shared task. We made both automatic (BLEU) and manual evaluations. 1 Introduction This paper is a joint report on two English-to-Czech submissions to the WMT11 shared translation task. The main contribution is however the proposal and evaluation of a rule-based post-processing system DEPFIX aimed at correcting errors in Czech gram- mar applicable to any MT system. This is somewhat the converse of other approaches (e.g. Simard et al. (2007)) where a statistical system was applied for the post-processing of a rule-based one. 2 Our phrase-based systems This section briefly describes our underlying phrase- based systems. One of them (CU-BOJAR) was sub- mitted directly to the WMT11 manual evaluation, the other one (CU-TWOSTEP) was first corrected by the proposed method (Section 3 below) and then submitted under the name CU-MARECEK. ∗This research has been supported by the European Union Seventh Framework Programme (FP7) under grant agreement n◦ 247762 (Faust), n◦ 231720 (EuroMatrix Plus), and by the grants GAUK 116310 and GA201/09/H057. 2.1 Data for statistical systems Our training parallel data consists of CzEng 0.9 (Bojar and ˇZabokrtsk´y, 2009), the News Commen- tary corpus v.6 as released by the WMT11 orga- nizers, the EMEA corpus, a corpus collected from the transcripts of TED talks (http://www.ted.com), the parallel news and separately some of the par- allel web pages of the European Commission (http://ec.europa.eu), and the Official Journal of the European Union as released by the Apertium con- sortium (http://apertium.eu/data). A custom web crawler was used for the European Commission website. English and Czech websites were matched according to their URLs. Unfortu- nately, Czech websites very often contain untrans- lated parts of English texts. Because of this, we aimed especially at the news articles, which are very often translated correctly and also more relevant for the shared task. Texts were segmented using train- able tokenizer (Klyueva and Bojar, 2008) and dedu- plicated. Processed texts were automatically aligned by Hunalign (Varga and others, 2005). The data from the Official Journal were first con- verted from XML to plain text. The documents were paired according to their filenames. To better han- dle the nature of these data, we decided to divide the documents into two classes based on the aver- age number of words per sentence: “lists” are docu- ments with less than 2.8 words per sentence, other documents are called “texts”. The corresponding “lists” were …

Content truncated to 500 words