The ITC-irst statistical machine translation system for IWSLT-

The ITC-irst Statistical Machine Translation System for IWSLT-2004 Nicola Bertoldi, Roldano Cattoni, Mauro Cettolo, Marcello Federico ITC-irst, via Sommarive 18, Povo (TN) - Italy {bertoldi,cattoni,cettolo,federico}@itc.it Abstract Focus of this paper is the system for statistical machine trans- lation developed at ITC-irst. It has been employed in the evaluation campaign of the International Workshop on Spo- ken Language Translation 2004 in all the three data set condi- tions of the Chinese-English track. Both the statistical model underlying the system and the system architecture are pre- sented. Moreover, details are given on how the submitted runs have been produced. 1. Introduction This paper reports on the participation of ITC-irst to the eval- uation campaign organized by the International Workshop on Spoken Language Translation (IWSLT) 2004. The Statisti- cal Machine Translation (SMT) system developed at ITC-irst was applied to all the three data set conditions of the Chinese- English track. The ITC-irst SMT system implements an extension of the IBM Model 4 as a log-linear interpolation of statistical mod- els, which estimate probabilities in terms of phrases. The use of phrases rather than words has recently emerged as a mean to cope with the limited context that Model 4 exploits to guess word translation (lexicon model) and word positions (distortion model) [1, 2, 3, 4, 5, 6, 7]. While parameters of the models are estimated exploiting statistics of phrase pairs extracted from word alignments, the weights of the interpolation are optimized through a train- ing procedure which directly aims at minimizing translation errors on a development set. Decoding is implemented in terms of a dynamic program- ming algorithm. The paper is organized as follows. Next section details the statistical model underlying the system. Sections 3 and 4 briefly describe the search and the segmentation algorithms, respectively. Section 5 gives an overview of the system ar- chitecture. Finally, in Section 6 experimental set-ups of the evaluation campaign runs and results are presented and dis- cussed. 2. Statistical Machine Translation The advantages of the statistical translation approach are ad- vocated by the many papers on the subject, which followed its first introduction. Of course, there have been also attempts to overcome some of its shortcomings, e.g. the use of limited context within the foreign string to guess word translations and word positions. Recently, several research labs have re- ported improvements in translation accuracy by shifting from word- to phrase-based SMT. In particular, statistical phrase- based translation models have recently emerged, which rely on statistics of phrase pairs. Phrase pairs statistics can be au- tomatically extracted from word-aligned parallel corpora [5]. In the following subsections, we introduce the SMT frame- work and the Model 4. Then, we briefly describe a method for extracting phrase pairs. Finally, a novel phrase-based translation framework is presented which is tightly related to Model 4. 2.1. Log-linear Model As originally proposed by [8], the most likely translation of a foreign source sentence f into English is obtained by search- ing for the sentence with the highest posterior probability: e∗ = arg …

Content truncated to 500 words