The AFRL-MITLL WMT15 System: There’s More than One Way to Decode It!

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 112–119, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. The AFRL-MITLL WMT15 System: There’s More than One Way to Decode It! Jeremy Gwinnup†, Timothy Anderson, Grant Erdmann, Katherine Young†, Christina May† Air Force Research Laboratory jeremy.gwinnup.ctr,timothy.anderson.20, grant.erdmann,katherine.young.1.ctr, christina.may.3.ctr@us.af.mil Michaeel Kazi‡, Elizabeth Salesky‡, Brian Thompson‡ MIT Lincoln Laboratory michaeel.kazi,elizabeth.salesky, brian.thompson@ll.mit.edu Abstract This paper describes the AFRL-MITLL statistical MT systems and the improve- ments that were developed during the WMT15 evaluation campaign. As part of these efforts we experimented with a num- ber of extensions to the standard phrase- based model that improve performance on the Russian to English translation task cre- ating three submission systems with differ- ent decoding strategies. Out of vocabulary words were addressed with named entity postprocessing. 1 Introduction As part of the 2015 Workshop on Machine Translation (WMT15) shared translation task, the MITLL and AFRL human language technol- ogy teams participated in the Russian–English translation task. Our machine translation sys- tems represent enhancements to both our sys- tems from IWSLT2014 (Kazi et al., 2014) and WMT14 (Schwartz et al., 2014), the addition of hierarchical decoding systems (Hoang and Koehn, 2008), neural network joint models (Devlin et al., 2014) and the utilization of Drem (Erdmann and Gwinnup, 2015), a method of scaled derivative- free trust-region optimization, during the system tuning process. 2 System Description We submitted systems for the Russian-to-English machine translation shared task. In all submitted systems, we used either phrase-based or hierarchi- cal variants of the moses decoder (Koehn et al., †This work is sponsored by the Air Force Research Labo- ratory under Air Force contract FA-8650-09-D-6939-029. ‡This work is sponsored by the Air Force Research Labo- ratory under Air Force contract FA-8721-05-C-0002. 2007). As in previous years, our submitted sys- tems used only the constrained data supplied when training. 2.1 Data Usage In training our Russian–English systems we uti- lized the following corpora to train translation and language models: Yandex1, Commoncrawl (Smith et al., 2013), LDC Gigaword English v5 (Parker et al., 2011) and News Commentary. The Wikipedia Headlines corpus2 was reserved to train named en- tity recognizers. 2.2 Data Preprocessing As with our WMT14 submission systems, prepro- cessing to address issues with the training data was required to ensure optimal system performance. Unicode characters in the private use, control char- acter(C0, C1, zero-width, non-breaking, joiner, directionality and paragraph markers), and unal- located ranges were removed. Punctuation nor- malization and tokenization using Moses prepro- cessing scripts were then applied before lower- casing the data. The Commoncrawl corpus was further processed as in Schwartz et al. (2014) to exclude wrong-language text and to normalize mixed-alphabet spellings. 2.3 Factored Data Generation We generated a class-factored version of the paral- lel Russian–English training data by using mkcls to produce 600 word classes for each side of the data. The factored data was then used to create a factored translation model and an in-domain class language model (Brown et al., 1993) for the En- glish portion. 1https://translate.yandex.ru/corpus?lang=en …

Content truncated to 500 words