The MIT-LL/AFRL MT System

The MIT-LL/AFRL MT System Wade Shen, Brian Delaney, and Tim Anderson MIT Lincoln Laboratory 244 Wood St. Lexington, MA 02420 {swade,bdelaney}@ll.mit.edu Air Force Research Laboratory 2255 H St. Wright-Patterson AFB, OH 45433 Timothy.Anderson@wpafb.af.mil Abstract The MIT-LL/AFRL MT system is a statistical phrase-based translation system that implements many modern SMT training and decoding techniques. Our system was designed with the long term goal of dealing with corrupted ASR input for Speech-to-Speech MT applications. This paper will discuss the architecture of the MIT-LL/AFRL MT system, and experiments with manual and ASR transcription data that were run as part of the IWSLT-2005 Chinese-to-English evaluation campaign.1 1. Introduction In recent years, the development of statistical methods for machine translation has resulted in high quality translations that can be used in real applications with increasing confidence. Specific advancements include: • Extracting word alignments from parallel corpora [1] [2] • Learning and modeling the translation of phrases [4] [5] • Combining and optimizing model parameters [6] [7] [8] • Decoding and rescoring techniques [9] [10] Our system draws from these advances and implements a number of these techniques including log-linear model combination and minimum error rate training to translate foreign language sentences. We developed our system during preparation for IWSLT-2005 to serve as a platform for future research. Most of the components of our system have been developed in-house in order to facilitate future experimentation. In subsequent sections, we will discuss the details translation system including our alignment and language models and methods we’ve implemented for optimization and decoding. The basic translation training and decoding processes are shown in Figure 1. We start with a word alignment extracted from a training set using GIZA++. These alignments are expanded and phrases are counted to form the phrase translation model. Language models are then trained from the English side of training set (and possibly with other English texts, if available). Using development bitexts separated from the training set, we then employ a minimum error rate training process to 1 This work is sponsored by the United States Air Force Research Laboratory under Air Force Contract FA8721-05-C-0002. Opinions, interpretations, conclusions and recommendations are those of the author and are not necessarily endorsed by the United States Government. optimize model parameters utilizing a held out development set. These trained parameters and models can then be applied to test data during decoding and rescoring phases of the translation process. Figure 1: Basic Statistical Translation Architecture 2. Translation Models The basic phrase-translation model we employ is described in detail [4] and [5]. We use GIZA++ [1], [3] and [4] to generate alignments between foreign language and English language words/tokens in both directions (f-to-e and e-to-f). An expanded alignment is then computed using heuristics that interpolate between the intersection and union of these bidirectional word alignments as detailed in [4] and [5]. Then phrases are extracted from the expanded alignments to build the phrase model. We introduced two minor modifications to this basic process in our system: 1. Prior to iterative expansion of …

Content truncated to 500 words