Linguistic representation of Finnish in a limited domain speech-to-speech translation system

226 EAMT 2005 Conference Proceedings Linguistic representation of Finnish in a limited domain speech-to-speech translation system Marianne Santaholma University of Geneva, ISSCO/TIM/ETI, Switzerland Marianne.Santaholma@eti.unige.ch Abstract. This paper describes the development of Finnish linguistic resources for use in MedSLT, an Open Source medical domain speech-to-speech translation system. The paper describes the collection of medical sub-domain corpora for Finnish, the creation of the Fin- nish generation grammar by adapting the original English grammar, the composition ofthe domain specific Finnish lexicon and the definition of interlingua to Finnish mapping rules for multilingual translation. It is shown that Finnish can be effectively introduced into the existing MedSLT framework and that despite the differences between English and Finnish, the Finnish grammar can be created by manual adaptation from the original English gram- mar. An initial evaluation of English to Finnish speech-to-speech translation is also pre- sented. 1. Introduction The basic architecture of a speech-to-speech translation system typically includes several com- ponents. Speech-to-speech translation systems are composed of a source language speech rec- ognition module, followed by a translation mo- dule which converts the recognized string into the target language, followed by a text-to-speech output module. These components may be based on different kinds of architectures. For example translations may be obtained using a variety of translation methodologies, such as rule-based, statistical or example-based translation engines. In past years statistical methods have been commonly used in speech systems. This even to the point that it may have given the impression that rule-based methods are no longer relevant. The general success of statistical methods over rule-based methods is based principally on the general robustness of the statistical systems and on the overall easiness of system development. However in some special fields, like for exam- ple in the medical domain, reliability of the sys- tem is the primary design goal, and it is valued higher than robustness. This suggests that in these domains rule-based methods can offer ad- vantages. MedSLT is an Open Source project which is developing a generic platform for building this kind of rule-based system where reliability is a crucial issue (See Rayner & Bouillon, 2002; Rayner et al, 2004). To com- pare rule-based methods to statistical methods there exist two versions of the system, one based on grammar-based language modeling (GLM) and one on statistical language model- ing (SLM). These versions are trained on the same corpus, and evaluated on a test corpus col- lected using both versions of the system. The experiments show that in terms of number of sentences translated, the GLM and SLM scored equally well. However, (Rayner et al, 2004) concluded that the GLM was preferable in terms of presenting a more predictable inter- face. A rule-based spoken translation system im- plies several different resources: a description of the source language (SL) and of the target language (TL) and a set of translation rules, for example transfer rules or interlingua mapping rules. Since in general the development of lin- guistic resources used in translation systems is laborious and time consuming, …

Content truncated to 500 words