Machine Translation and Monolingual Postediting: The AFRL WMT-14 System

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 186–194, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Machine Translation and Monolingual Postediting: The AFRL WMT-14 System Lane O.B. Schwartz Air Force Research Laboratory lane.schwartz@us.af.mil Timothy Anderson Air Force Research Laboratory timothy.anderson.20@us.af.mil Jeremy Gwinnup SRA International† jeremy.gwinnup.ctr@us.af.mil Katherine M. Young N-Space Analysis LLC† katherine.young.1.ctr@us.af.mil Abstract This paper describes the AFRL sta- tistical MT system and the improve- ments that were developed during the WMT14 evaluation campaign. As part of these efforts we experimented with a number of extensions to the stan- dard phrase-based model that improve performance on Russian to English and Hindi to English translation tasks. In addition, we describe our efforts to make use of monolingual English speakers to correct the output of ma- chine translation, and present the re- sults of monolingual postediting of the entire 3003 sentences of the WMT14 Russian-English test set. 1 Introduction As part of the 2014 Workshop on Machine Translation (WMT14) shared translation task, the human language technology team at the Air Force Research Laboratory participated in two language pairs: Russian-English and Hindi-English. Our machine translation sys- tem represents enhancements to our system from IWSLT 2013 (Kazi et al., 2013). In this paper, we focus on enhancements to our pro- cedures with regard to data processing and the handling of unknown words. In addition, we describe our efforts to make use of monolingual English speakers to correct the output of machine translation, and present the results of monolingual postediting of the entire 3003 sentences of the WMT14 Russian- English test set. Using a binary adequacy clas- sification, we evaluate the entire postedited †This work is sponsored by the Air Force Research Laboratory under Air Force contract FA-8650-09-D- 6939-029. test set for correctness against the reference translations. Using bilingual judges, we fur- ther evaluate a substantial subset of the post- edited test set using a more fine-grained ade- quacy metric; using this metric, we show that monolingual posteditors can successfully pro- duce postedited translations that convey all or most of the meaning of the original source sen- tence in up to 87.8% of sentences. 2 System Description We submitted systems for the Russian-to- English and Hindi-to-English MT shared tasks. In all submitted systems, we use the phrase-based moses decoder (Koehn et al., 2007). We used only the constrained data sup- plied by the evaluation for each language pair for training our systems. 2.1 Data Preparation Before training our systems, a cleaning pass was performed on all data. Unicode charac- ters in the unallocated and private use ranges were all removed, along with C0 and C1 con- trol characters, zero-width and non-breaking spaces and joiners, directionality and para- graph markers. 2.1.1 Hindi Processing The HindEnCorp corpus (Bojar et al., 2014) is distributed in tokenized form; in order to ensure a uniform tokenization standard across all of our data, we began by detokenized this data using the Moses detokenization scripts. In addition to normalizing various extended Latin punctuation marks to their …

Content truncated to 500 words