Valency and MT: Recent Developments in the METAL System

V~TJ~CY AND MT: RECENT DEVELOPMENTS IN THE METAL SYSTEM Rudi Gebruers Siemens-METAL project Katholieke Universiteit Leuven Maria Theresiastraat 21 B-3000 LEUVEN BELGIUM ABSTRACT This paper describes a valency model, developed within the Belgian METAL project, aimed at enhancing the modularity and multilinguality of the METAL system. The introduction provides background, section 1 discusses the existing valency framework, and section 2 presents the alternative model. The final section deals with some results and problems with this model. 0. Introduction The task of MT is to map between equivalent linguistic objects. One of the central design questions in MT is that of the best method to decompose the translation relation. The ideal would be to have a system that produces a (natural) language-independent representation from a source language (SL) text, which could then be synthesized in any target language (TL). However, this ideal not being feasible for real-world texts, it has become customary to adopt a model where a transfer module, specific to one language pair, defines a mapping between language-dependent structural representations. In principle it should be possible to design a 'transfer' model in such a way that the analysis module for mapping surface strings onto structural representations and the synthesis module for mapping structural representations onto surface strings remain the same, regardless of the TL and SL, respectively. The advantage of this 'multilingual' design is that existing modules will not be seriously affected by the addition of a new language to the system. A still more attractive, but also more ambitious, design would be one in which the same grammar can be used for both parsing and generating, and the same translation rules for translating between two languages in either direction (see Jin and Simmons, 1986 for an example of a 'symmetric' translation system). Whereas early MT systems blended the rules of grammar and the analysis procedure for efficiency reasons, it has also become customary, given current system optimization techniques, to make a clear separation between programming logic and data on the one hand, and linguistic logic and data on the other. This separation is convenient for the division of labour between the linguist and the programmer, and it enables the former to revise and complete his rule systems without the latter having to constantly change his programs. The METAL automatic translation system tries to be multilingual in the above sense. More- over, it makes an attempt at separating soft- ware and lingware (= linguistic knowledge written in a specialised user language). In the following, I will show how the adoption of a new kind of valency framework, developed at the K.U.Leuven in the course of the last two years, enhances the multilinguality and modularity even further. For the sake of clarity, I will first review the relevant aspects of the valency framework in the current METAL system. I. Valency in the METAL system Since the main claim to be advanced in this paper bears on the relation between a valency framework and the general design of an MT system, …

Content truncated to 500 words