Adapting finite-state translation to the TransType2 project

Adapting finite-state translation to the TransType2 project Elsa Cubel, Jorge Gonz´alez, Antonio L. Lagarda, Francisco Casacuberta, Alfons Juan and Enrique Vidal Institut Tecnol`ogic d’Inform`atica Universitat Polit`ecnica de Val`encia E-46071 Val`encia, Spain {ecubel, jgonza, alagarda, fcn, ajuan, evidal}@iti.upv.es Abstract Machine translation can play an impor- tant role nowadays, helping communica- tion between people. One of the projects in this field is TransType2 1. Its purpose is to develop an innovative, interactive machine translation system. TransType2 aims at facilitating the task of producing high-quality translations, and make the translation task more cost-effective for hu- man translators. To achieve this goal, stochastic finite-state transducers are being used. Stochastic finite-state transducers are generated by means of hybrid finite-state and statistical alignment techniques. Viterbi parsing procedure with stochastic finite-state transducers have been adapted to take into account the source sentence to be translated and the target prefix given by the human translator. Experiments have been carried out with a corpus of printer manuals. The first results showed that with this preliminary proto- type, users can only type a 15% of the words instead the whole complete trans- lated text. 1TransType2 - Computer Assisted Translation, RTD project by the European Commission under the IST Programme (IST- 2001-32091). 1 Introduction The aim of the TransType2 project (TT2) (Schlum- bergerSema S.A. et al., 2001) is to develop a Com- puter Assisted Translation (CAT) system that will help to solve a very pressing social problem: how to meet the growing demand for high-quality trans- lation. The innovative solution proposed by TT2 is to embed a data driven Machine Translation (MT) en- gine within an interactive translation environment. In this way, the system combines the best of two paradigms: the CAT paradigm, in which the human translator ensures high-quality output, and the MT paradigm (Brown et al., 1990), in which the machine ensures significant productivity gains. Until now, translation technology has not been able to keep pace with the demands for high- quality translation. TT2 has the ability to signifi- cantly increase translator productivity and thus has enormous commercial potential. Six different ver- sions of the system will be developed for translation between English and French, Spanish or German. To ensure that TT2 meets the needs of translators, two professional translation agencies are in charge of evaluation of the successive prototypes. Stochastic finite-state transducers (SFST) have proved to be adequate models for MT in limited-domain applications (Vidal, 1997; Amen- gual et al., 2000; Casacuberta et al., 2001). SFSTs are interesting for their simplicity and the possibil- ity of inferring models automatically from bilingual training corpus. They allow a very efficient search of new test data (Vidal, 1997) and make it possible to work with text-input and speech-input translation (Amengual et al., 2000). Moreover, hybrid finite-state techniques and statistical translation techniques can be used to pro- duce efficient SFSTs. The learning of SFST can be improved if word-aligned training pairs are used (i.e. using statistical alignment models). This paper is concerned with the use of SF- STs for computer-assisted translation. The …

Content truncated to 500 words