The University of Washington machine translation system for the IWSLT 2007 competition

The University of Washington Machine Translation System for the IWSLT 2007 Competition Katrin Kirchhoff and Mei Yang Department of Electrical Engineering University of Washington {katrin,yangmei}@ee.washington.edu Abstract This paper presents the University of Washington’s sub- mission to the 2007 IWSLT benchmark evaluation. The UW system participated in two data tracks, Italian-to-English and Arabic-to-English. Our main focus was on incorporating out-of-domain data, which contributed to improvements for both language pairs in both the clean text and ASR output conditions. In addition, we compared supervised and semi- supervised preprocessing schemes for the Arabic-to-English task and found that the semi-supervised scheme performs competitively with the supervised algorithm while using a fraction of the run-time. 1. Introduction We describe University of Washington’s machine translation system for the 2007 IWSLT competition. Our system partici- pated in two tracks, the Italian-to-English Challenge task and the Arabic-to-English Classical task. For Italian-to-English translation our main focus was on utilizing out-of-corpus training resources and to determine the relative benefit of having small amounts of in-domain training data vs. larger amounts of unrelated data. For Arabic-to-English transla- tion, we also investigated using text resources unrelated to the BTEC travel task. In addition, we compared linguis- tic methods for tokenization vs. a semi-supervised algorithm that seeks to improve initial tokenization by utilizing unan- notated data. The following sections describe the basic setup of the translation system (Section 2), the Italian-to-English System (Section 3), the Arabic-to-English system (Section 4), and experiments and results (Section 5). 2. Translation System 2.1. Translation model Our system is a phrase-based statistical MT system based on a log-linear probability model: e∗ = argmaxep(e|f) = argmaxe{ K � k=1 λkφk(e, f)} (1) Given and English sentence e and a foreign sentence f, φ(e, f) is a feature function defined on both sentences, and λ is a feature weight. In particular, we use the following feature functions: • two phrase-based translation scores, one for each translation direction • two lexical translation scores, one for each translation direction • word count penalty • phrase count penalty • distortion penalty • language model score • a data source feature Phrasal and lexical translation scores are computed as shown below in Equations 2 and 3. For a segmentation of source and target sentences into phrases, f = ¯f1, ¯f2, ..., ¯fM and e = ¯e1, ¯e2, ..., ¯eM, the phrasal translation score for ¯e given ¯f is computed as P(¯e| ¯f) = count(¯e, ¯f) count( ¯f) (2) i.e. as the relative frequency estimate from the phrase- segmented training corpus. The lexical score is computed as Scorelex(¯e| ¯f) = J � j=1 1 |{j|a(i) = j}| I � a(i)=j p(fj|ei) (3) where j ranges over words in phrase ¯f and i ranges over words in phrase ¯e. Phrases are extracted from the word-aligned training cor- pus using the heuristic technique described in [6]. For word alignment we use an HMM-based word to phrase alignment model [1]. Under this model, target phrases are generated by individual source words, including the NULL word. The alignment of …

Content truncated to 500 words