The MITLL-AFRL IWSLT 2016 Systems

The MITLL-AFRL IWSLT 2016 Systems† Michaeel Kazi1, Elizabeth Salesky1, Brian Thompson1, Jonathan Taylor1, Jeremy Gwinnup2‡, Timothy Anderson2, Grant Erdmann2, Eric Hansen2, Brian Ore2, Katherine Young2, Michael Hutt2 1MIT Lincoln Laboratory 2Air Force Research Laboratory Human Language Technology Group Airman Systems Directorate 244 Wood Street 2255 H Street Lexington, MA 02420, USA Wright-Patterson AFB, OH 45433, USA Abstract This report summarizes the MITLL-AFRL MT and ASR systems and the experiments run during the 2016 IWSLT evaluation campaign. Building on lessons learned from pre- vious years’ results, we refine our ASR systems and exam- ine the explosion of neural machine translation systems and techniques developed in the past year. We experiment with a variety of phrase-based, hierarchical and neural-network ap- proaches in machine translation and utilize system combina- tion to create a composite system with the best characteristics of all attempted MT approaches. 1. Introduction During the evaluation campaign for the 2016 International Workshop on Spoken Language Translation (IWSLT16) [1] our experimental efforts in machine translation (MT) focused on the extension of our efforts from WMT16[2] and IWSLT15[3] and the exploration of many new neural machine translation (NMT) techniques including the refinement and improvement of our in-house NMT techniques, advanced selection tech- niques for parallel training data and the combination of this myriad of systems and techniques via system combination. Our Automatic Speech Recognition (ASR) systems largely remain the same as last year, with the exception of training with this year’s additional data. 2. Machine Translation For our efforts in the machine translation task this year we acknowledge the recent explosion of neural machine transla- tion (NMT) techniques and leverage our previous experience with phrase-based and hierarchical machine translation sys- tems to create a best-of-breed machine translation system via the technique of system combination. †This material is based upon work supported by the Air Force Research Laboratory in part under Air Force Contracts No. (FA8721-05-C-0002 and/or FA8702-15-D-0001). Any opinions, findings and conclusions or rec- ommendations expressed in this material are those of the authors and do not necessarily reflect the views of the Air Force Research Laboratory. ‡Corresponding author: jeremy.gwinnup.1@us.af.mil 2.1. Data Used As in previous years, we made use of all data sources avail- able to train various aspects of our MT systems: the WIT3 Corpus [4], the new QED (formerly AMARA) Corpus [5], the new Parallel UN Corpus [6], and the existing Multi-UN Corpus [7]. We also make use of all the data sources avail- able to WMT16[8]. 2.2. Preprocessing As in previous years, we use processes documented in [9] to clean the training data. The QED corpus required more processing as detailed in §2.2.3. 2.2.1. Arabic Preprocessing For two MT systems, we used MADAMIRA [10] v1.0 with the ATB scheme. For our other MT Systems, we applied Farasa[11] Ara- bic morphological processing to most of our MT systems. We experimented with variations of this processing, includ- ing the separation of stems and affixes. An additional vari- ation was to break each fragment into individual tokens in- stead of grouping …

Content truncated to 500 words