Omnifluent English-to-French and Russian-to-English Systems for the 2013 Workshop on Statistical Machine Translation

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 158–163, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics OmnifluentTM English-to-French and Russian-to-English Systems for the 2013 Workshop on Statistical Machine Translation Evgeny Matusov, Gregor Leusch Science Applications International Corporation (SAIC) 7990 Science Applications Ct. Vienna, VA, USA {evgeny.matusov,gregor.leusch}@saic.com Abstract This paper describes OmnifluentTM Trans- late – a state-of-the-art hybrid MT sys- tem capable of high-quality, high-speed translations of text and speech. The sys- tem participated in the English-to-French and Russian-to-English WMT evaluation tasks with competitive results. The features which contributed the most to high translation quality were training data sub-sampling methods, document-specific models, as well as rule-based morpholog- ical normalization for Russian. The latter improved the baseline Russian-to-English BLEU score from 30.1 to 31.3% on a held- out test set. 1 Introduction Omnifluent Translate is a comprehensive multilin- gual translation platform developed at SAIC that automatically translates both text and audio con- tent. SAIC’s technology leverages hybrid machine translation, combining features of both rule-based machine and statistical machine translation for im- proved consistency, fluency, and accuracy of trans- lation output. In the WMT 2013 evaluation campaign, we trained and tested the Omnifluent system on the English-to-French and Russian-to-English tasks. We chose the En–Fr task because Omnifluent En– Fr systems are already extensively used by SAIC’s commercial customers: large human translation service providers, as well as a leading fashion de- signer company (Matusov, 2012). Our Russian-to- English system also produces high-quality transla- tions and is currently used by a US federal govern- ment customer of SAIC. Our experimental efforts focused mainly on the effective use of the provided parallel and monolin- gual data, document-level models, as well using rules to cope with the morphological complexity of the Russian language. While striving for the best possible translation quality, our goal was to avoid those steps in the translation pipeline which would make a real-time use of the Omnifluent sys- tem impossible. For example, we did not integrate re-scoring of N-best lists with huge computation- ally expensive models, nor did we perform system combination of different system variants. This al- lowed us to create a MT system that produced our primary evaluation submission with the translation speed of 18 words per second1. This submission had a BLEU score of 24.2% on the Russian-to- English task2, and 27.3% on the English-to-French task. In contrast to many other submissions from university research groups, our evaluation system can be turned into a fully functional, commer- cially deployable on-line system with the same high level of translation quality and speed within a single work day. The rest of the paper is organized as follows. In the next section, we describe the core capabilities of the Omnifluent Translate systems. Section 3 explains our data selection and filtering strategy. In Section 4 we present the document-level trans- lation and language models. Section 5 describes morphological transformations of Russian. In sec- tions 6 we present an extension to the system that allows for automatic spelling correction. In …

Content truncated to 500 words