Interlingua vs. Transfer? Knowledge Sharing Across Projects

INTERLINGUA VS TRANSFER? KNOWLEDGE SHARING ACROSS PROJECTS Nadia Mesli IAI Martin-Luther-Straße 14 D-66111 Saarbrücken Germany email: nadia@iai.uni-sb.de Abstract Sharing knowledge across projects is often considered impossible, especially if the sys- tems involved present differences in underlying theory, representation, or even program- ming language. This paper, taking a collaboration between the CAT2 and PANGLOSS MT projects as an example, demonstrates that it is possible to combine resources devel- oped separately to support the processing of different modules in a joint system. Despite the differences in the two MT methods, the transfer-based CAT2 and the interlingual PANGLOSS systems proved to share similarities in their linguistic representations that simplified the interfacing between the two systems. Referring to this result the paper presents a discussion of the transfer/interlingua continuum and attempts to bridge the gap between the two MT alternatives. 1 Introduction As Machine Translation (MT) matures, and as systems become more modular and well specified, pressure increases on system builders to re-use previously built modules instead of developing new ones for every new system. The natural question arises: can one put together a parser for one language and a generator for another to produce an MT system from scratch? What is required of the system? When combining two systems using different MT methods, one may well wonder whether the two systems involved must share some similarities in their linguistic representations, or whether they can offer interfacing facilities that simplify the mapping from one system into the other. This paper presents the result of a six-month exchange program between the University of Southern California (USA) and the University of the Saarland (Germany). It is an extract of the final report [Mesli 94]. The aim of the project was to test the possibility of combining the CAT2 MT system developed at the Institut der Gesellschaft zur Förderung der Angewandten Informationsforschung (IAI) of the University of the Saarland as a sideline to the EUROTRA MT project with modules of the PANGLOSS MT system developed at the Information Sciences Institute (ISI) of the University of Southern California. Our purpose was to investigate the possibility of linking together two such different, separately developed systems as the transfer-based CAT2 and the interlingual knowledge-based PANGLOSS 169 systems, despite differences in underlying method of translation, representation, and even pro- gramming language (CAT2 is written in Prolog and PANGLOSS primarily in Lisp). Despite the differences in the two MT systems and methods involved, such a linkup seemed to be possible from the outset, since both systems are modularized, containing several levels for potential interfacing. That CAT2 is transfer-based rather than interlingua-based means that the concept of an interface between source language and target language is inherent in the architecture of the system. Because of the collaboration between different research groups within the PANGLOSS MT project, the PANGLOSS MT system is also designed for having an interface that accepts any kind of informa- tion that need to be made available to PENMAN's sentence generator. For this purpose, it uses the normal interface to …

Content truncated to 500 words