LIMSI@WMT’15 : Translation Task

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 145–151, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. LIMSI @ WMT’15 : Translation Task Benjamin Marie1,2,3, Alexandre Allauzen1,2, Franck Burlot1, Quoc-Khanh Do1,2, Julia Ive1,2,4, Elena Knyazeva1,2, Matthieu Labeau1,2, Thomas Lavergne1,2, Kevin L¨oser1,2, Nicolas P´echeux1,2, Franc¸ois Yvon1 1LIMSI-CNRS, 91 403 Orsay, France 2Universit´e Paris-Sud, 91 403 Orsay, France 3Lingua et Machina 4Centre Cochrane franc¸ais firstname.lastname@limsi.fr Abstract This paper describes LIMSI’s submis- sions to the shared WMT’15 translation task. We report results for French-English, Russian-English in both directions, as well as for Finnish-into-English. Our submis- sions use NCODE and MOSES along with continuous space translation models in a post-processing step. The main novelties of this year’s participation are the follow- ing: for Russian-English, we investigate a tailored normalization of Russian to trans- late into English, and a two-step process to translate first into simplified Russian, fol- lowed by a conversion into inflected Rus- sian. For French-English, the challenge is domain adaptation, for which only mono- lingual corpora are available. Finally, for the Finnish-to-English task, we explore unsupervised morphological segmentation to reduce the sparsity of data induced by the rich morphology on the Finnish side. 1 Introduction This paper documents LIMSI’s participation to the machine translation shared task for three language pairs: French-English and Russian-English in both directions, as well as Finnish-into-English. Each of these tasks poses its own challenges. For French-English, the task differs slightly from previous years as it considers user-generated news discusssions. While the domain remains the same, the texts that need to be translated are of a less formal type. To cope with the style shift, new monolingual corpora have been made avail- able; they represent the only available in-domain resources to adapt statistical machine translation (SMT) systems. For Russian-English, the main source of diffi- culty is the processing of Russian, a morphologi- cally rich language with a much more complex in- flectional system than English. To mitigate the ef- fects of having too many Russian word forms, we explore ways to normalize Russian prior to trans- lation into English, so as to reduce the number of forms by removing some ”redundant” morpholog- ical information. When translating into Russian, we consider a two-step scenario. A conventional SMT system is first built to translate from En- glish into a simplified version of Russian; a post- processing step then restores the correct inflection wherever needed. Finally, for Finnish-into-English, we report pre- liminary experiments that explore unsupervised morphological segmentation techniques to reduce the sparsity issue induced by the rich morphology of Finnish. 2 Systems Overview Our experiments use NCODE1, an open source im- plementation of the n-gram approach, as well as MOSES, which implements a vanilla phrase-based approach.2 For more details about these toolkits, the reader can refer to (Koehn et al., 2007) for MOSES and to (Crego et al., 2011) for NCODE. 2.1 Tokenization and word alignments Tokenization for French and English text relies on in-house text processing tools (D´echelotte et al., 2008). All bilingual corpora provided by …

Content truncated to 500 words