Experimenting with phrase-based statistical translation within the IWSLT Chinese-to-English shared translation task

Experimenting with Phrase-Based Statistical Translation within the IWSLT 2004 Chinese-to-English Shared Translation Task Philippe Langlais, Michael Carl, Oliver Streiter RALI/DIRO IAI NUK University of Montreal Saarbr¨ucken National University of Kaohsiung Canada Germany Taiwan felipe@iro.umontreal.ca carl@iai.uni-sb.de ostreiter@nuk.edu.tw Abstract This paper describes the system we built for the Chinese to English track of the IWSLT 2004 evaluation campaign. A one month effort was devoted to this exercise, starting from scratch and making use as much as possible of freely avail- able packages. We show that a decent phrase-based transla- tion engine can be built within this short time frame. 1. Introduction Machine Translation is a very active field nowadays strongly anchored into a paradigm of performance. Evaluation exer- cises such as those conducted within the TIDES project are pushing system designers to constantly improve their sys- tems. Currently, many of the top-performing systems are phrase-based statistical machine translation (SMT) engines. The fact that SMT systems are among the best ones in those evaluation exercises is not surprizing considering the pecu- liarities of the translation tasks considered. The popularity of phrase-based models (PBMs) in SMT is neither a surprise, since PBMs allow to a certain extent to cope with local word reordering across languages, as well as to account for local context modelling. [1] also credit PBMs for being somehow tolerant to tokenization errors, an interesting characteristic when dealing with languages such as Chinese, the source lan- guage under consideration in this study. This effervescent activity comes with some bonus. Sev- eral freely available valuable packages (e.g. Giza++ [2], Pharaoh [3], SRILM [4]) make possible the fast develop- ment of a phrase-based translation engine. Other packages allow to quickly evaluate a system according to a gold stan- dard (e.g. MTEVAL http://www.nist.gov/speech/ tests/mt/mt2001/resource and GTM http:// nlp.cs.nyu.edu/GTM). This paper reports on the one month effort we spent building a system for the Chinese-to- English track of the IWSLT workshop, relying intensively on these packages. 2. Phrase-Based Models Very recently, several authors [5, 6] proposed at the same time an astonishingly simple but powerful model which we designate hereafter as Flat Phrase-Based model (FPBM). A FPBM is simply a collection of pairs of sequences of words with one or several scores (or probabilities) attached to them. The main difference between a FPBM and an alignment tem- plate (AT) model [7] being that the former does not attempt to model internal reordering of phrases. Thus, FPBMs as such do not have generalization capabilities. Zens and Ney [8] give an experimental comparison of both models on three different test sets. On the German-English Verbmobil task, the AT engine outperforms the PB engine they tested, while on the other tasks — the Spanish-English Xerox task, and the French-English Hansards task — they observed the opposite. Tom`as et al. [9] recently revisited the AT model and report that combining it with a FPBM brings some improvements. The recipe given in [5, 6] for acquiring a FPBM is sim- ple: use a word-alignment to identify in an heuristic way an …

Content truncated to 500 words