Yandex School of Data Analysis Machine Translation Systems for WMT13

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 99–103, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Yandex School of Data Analysis machine translation systems for WMT13 Alexey Borisov, Jacob Dlougach, Irina Galinskaya Yandex School of Data Analysis 16, Leo Tolstoy street, Moscow, Russia {alborisov,jacob,galinskaya}@yandex-team.ru Abstract This paper describes the English-Russian and Russian-English statistical machine translation (SMT) systems developed at Yandex School of Data Analysis for the shared translation task of the ACL 2013 Eighth Workshop on Statistical Machine Translation. We adopted phrase-based SMT approach and evaluated a number of different techniques, including data fil- tering, spelling correction, alignment of lemmatized word forms and translitera- tion. Altogether they yielded +2.0 and +1.5 BLEU improvement for ru-en and en- ru language pairs. We also report on the experiments that did not have any positive effect and provide an analysis of the prob- lems we encountered during the develop- ment of our systems. 1 Introduction We participated in the shared translation task of the ACL 2013 Workshop on Statistical Machine Translation (WMT13) for ru-en and en-ru lan- guage pairs. We provide a detailed description of the experiments carried out for the development of our systems. The rest of the paper is organized as follows. Section 2 describes the tools and data we used. Our Russian→English and English→Russian se- tups are discussed in Section 3. In Section 4 we report on the experiments that did not have any positive effect despite our expectations. We pro- vide a thorough analysis of erroneous outputs in Section 5 and draw conclusions in Section 6. 2 Tools and data 2.1 Tools We used an open source SMT system Moses (Koehn et al., 2007) for all our experiments ex- cluding the one described in Section 4.1 due to its performance constraints. To overcome the limita- tion we employed our in-house decoder. Language models (LM) were created with an open source IRSTLM toolkit (Federico et al., 2008). We computed 4-gram LMs with modified Kneser-Ney smoothing (Kneser and Ney, 1995). We used an open source MGIZA++ tool (Gao and Vogel, 2008) to compute word alignment. To obtain part of speech (POS) tags we used an open source Stanford POS tagger for English (Toutanova et al., 2003) and an open source suite of language analyzers, FreeLing 3.0 (Carreras et al., 2004; Padró and Stanilovsky, 2012), for Rus- sian. We utilized a closed source free for non- commercial use morphological analyzer, Mystem (Segalovich, 2003), that used a limited dictionary to obtain lemmas. We also made use of the in-house language rec- ognizer based on (Dunning, 1994) and a spelling corrector designed on the basis of the work of Cucerzan and Brill (2004). We report all results in case-sensitive BLEU (Papineni et al., 2002) using mt-eval13a script from Moses distribution. 2.2 Data Training data We used News Commentary and News Crawl monolingual corpora provided by the organizers of the workshop. Bilingual training data comprised English- Russian parallel corpus release by Yandex1, News Commentary and Common Crawl corpora pro- vided …

Content truncated to 500 words