Tuning a phrase-based statistical translation system for the IWSLT 2005 Chinese to English and Arabic to English tasks

Tuning a phrase-based statistical translation system for the IWSLT 2005 Chinese to English and Arabic to English tasks Marta R. Costa-juss`a and Jos´e A. R. Fonollosa TALP Research Center Universitat Polit`ecnica de Catalunya, Barcelona {mruiz,adrian}@gps.tsc.upc.edu Abstract Nowadays, most of the statistical translation systems are based on phrases (i.e. groups of words). We describe a phrase-based system using a modified method for the phrase extraction which deals with larger phrases while keeping a reasonable number of phrases. Also, different alignments to extract phrases are allowed and additional features are used which lead to a clear improvementin the performance of translation. Finally, the system manages to do reordering. We report results in terms of translation accuracy by using the BTEC corpus in the tasks of Chi- nese to English and Arabic to English, in the framework of IWSLT’05 evaluation. 1. Introduction From the initial word-based translation models [3], research on statistical machine translation has been strongly improved. At the end of the last decade the use of context in the translation model (phrase-based approach) supposed a clear improvement in translation quality ( [17], [16], [8]). Statistical Machine Translation (SMT) is based on the assumption that every sentence e in the target language is a possible translation of a given sentence f in the source language. The main difference between two possible translations of a given sentence is a probability assigned to each, which has to be learned from a bilingual text corpus. Thus, the translation of a source sentence f can be formulated as the search of the target sentence e that maximizes the translation probability P(e|f), ˜e = argmax e P(e|f) (1) If we use Bayes rule to reformulate the translation probability, we obtain, ˜e = argmax e P(f|e)P(e) (2) This translation model is known as the source-channel approach [2] and it consists on a language model P(e) and a separate translation model P(f|e) [6]. In the last few years, new systems tend to use se- quences of words, commonly called phrases [7], aiming at introducing word context in the translation model. As alternative to the source-channel approach the decision rule can be modeled through a log-linear maximum en- tropy framework. ˜e = argmax e � M � m=1 λmhm(e, f) � (3) The features functions, hm, are the system mod- els (translation model, language model and others) and weights, λi, are typically optimized to maximize a scor- ing function [12]. It is derived from the Maximum En- tropy approach as shown in [1] and has the advantage that additional features functions can be easily integrated in the overall system. This paper addresses a modification of the phrase- extraction algorithm in [13] and results in Chinese to En- glish and Arabic to English tasks are reported. It also combines several alignments before extracting phrases and interesting features. It is organized as follows. Sec- tion 2 explains the SMT system: the phrase extraction, its modification and shows the different features which have been taken into account and, briefly, the decoding; sec- tion …

Content truncated to 500 words