Yandex School of Data Analysis Russian-English Machine Translation System for WMT14

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 66–70, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Yandex School of Data Analysis Russian-English Machine Translation System for WMT14 Alexey Borisov and Irina Galinskaya Yandex School of Data Analysis 16, Leo Tolstoy street, Moscow, Russia {alborisov, galinskaya}@yandex-team.ru Abstract This paper describes the Yandex School of Data Analysis Russian-English system submitted to the ACL 2014 Ninth Work- shop on Statistical Machine Translation shared translation task. We start with the system that we developed last year and in- vestigate a few methods that were success- ful at the previous translation task includ- ing unpruned language model, operation sequence model and the new reparameter- ization of IBM Model 2. Next we propose a {simple yet practical} algorithm to trans- form Russian sentence into a more easily translatable form before decoding. The al- gorithm is based on the linguistic intuition of native Russian speakers, also fluent in English. 1 Introduction The annual shared translation task organized within the ACL Workshop on Statistical Machine Translation (WMT) aims to evaluate the state of the art in machine translation for a variety of lan- guages. We participate in the Russian to English translation direction. The rest of the paper is organized as follows. Our baseline system as well as the experiments concerning the methods already discussed in lit- erature are described in Section 2. In Section 3 we present an algorithm we use to transform the Rus- sian sentence before translation. In Section 4 we discuss the results and conclude. 2 Initial System Development We use all the Russian-English parallel data avail- able in the constraint track and the Common Crawl English monolingual corpus. 2.1 Baseline We use the phrase-based Moses statistical ma- chine translation system (Koehn et al., 2007) with mostly default settings and a few changes (Borisov et al., 2013) made in the following steps. Data Preprocessing includes filtering out non Russian-English sentence pairs and correction of spelling errors. Phrase Table Smoothing uses Good-Turing scheme (Foster et al., 2006). Consensus Decoding selects the translation with minimum Bayes risk (Kumar and Byrne, 2004). Handling of Unknown Words comprises incor- poration of proper names from Wiki Headlines parallel data provided by CMU1 and translitera- tion. We improve the transliteration algorithm in Section 2.4. Note that unlike last year we do not use word alignments computed for the lemmatized word forms. 2.2 Language Model We use 5-gram unpruned language model with modified Kneser-Ney discount estimated with KenLM toolkit (Heafield et al., 2013). 2.3 Word alignment Word alignments are generated using the fast_align tool (Dyer et al., 2013), which is much faster than IBM Model 4 from MGIZA++ (Gao and Vogel, 2008) and outperforms the latter in terms of BLEU. Results are given in Table 1. 2.4 Transliteration We employ machine transliteration to generate ad- ditional translation options for out-of-vocabulary 1http://www.statmt.org/wmt14/ wiki-titles.tgz 66 MGIZA++ fast_align Run Time 22 h 14 m 2h 49 m Perplexity – ru→en 97.00 90.37 – en→ru 209.36 216.71 BLEU – …

Content truncated to 500 words