The MIT-LL/AFRL IWSLT-2007 MT system

The MIT-LL/AFRL IWSLT-2007 MT System Wade Shen, Brian Delaney† MIT Lincoln Laboratory Information Systems and Technology Group 244 Wood St. Lexington, MA 02420, USA {swade,bdelaney}@ll.mit.edu Tim Anderson, Ray Slyh Air Force Research Laboratory Human Effectiveness Directorate 2255 H St. Wright-Patterson AFB, OH 45433 {Timothy.Anderson,Raymond.Slyh}@wpafb.af.mil Abstract The MIT-LL/AFRL MT system implements a standard phrase-based, statistical translation model. It incorporates a number of extensions that improve performance for speech- based translation. During this evaluation our efforts focused on the rapid porting of our SMT system to a new language (Arabic) and novel approaches to translation from speech in- put. This paper discusses the architecture of the MIT- LL/AFRL MT system, improvements over our 2006 system, and experiments we ran during the IWSLT-2007 evaluation. Specifically, we focus on 1) experiments comparing the per- formance of confusion network decoding and direct lattice decoding techniques for machine translation of speech, 2) the application of lightweight morphology for Arabic MT pre- processing and 3) improved confusion network decoding. 1. Introduction During the evaluation campaign for the 2007 International Workshop on Spoken Language Translation (IWSLT-2007) our experimental efforts centered on 1) the improvement of statistical MT when processing speech input and 2) robust methods for coping with minimal training data. In this paper we describe improvements over our base- line system. Refer to [5] for more detail regarding the imple- mentation of these stages of processing. The conditions of this year’s evaluation have changed as have the data sets and languages for which we’ve submitted system. As such, the processing stages described in 2 have been adapted accord- ingly. The remainder of this paper is structured as follows. In Section 2, we discuss the baseline system and minor im- provements to this standard statistical MT architecture that we incorporate. In sections 3, 4 and 5 we describe improve- ments to our baseline system including the design and imple- mentation of a new, lattice-based decoder for speech input †This work is sponsored by the Air Force Research Laboratory under Air Force contract FA8721-05-C-0002. Opinions, interpretations, conclu- sions and recommendations are those of the authors and are not necessarily endorsed by the United States Government. and the implementation of light morphological analysis for Arabic MT preprocessing. Section 6 shows results from a variety of development experiments using the provided data sets. In these experi- ments, we compare the performance of different ASR-based MT strategies using the same baseline system on both the Arabic and Italian speech translation tasks. We also report on experiments comparing the light Arabic morphological processing we employed to more sophisticated methods and their relative performance impact on MT quality. 1.1. IWSLT-2007 Data Usage We submitted systems for Chinese, Arabic and Italian-to- English language pairs. In each case, we used only data sup- plied by the evaluation (and in Italian, the named-entity list) for each language pair for training and optimization. From these data, we extract word/character alignments. These alignments are then expanded using slightly modified ver- sions of standard heuristics. This process is described in …

Content truncated to 500 words