Structuring Terminology: between Lexicons and Domain Knowledge Representation

This article presents a revisited view on the project DB-MAT and its approach to model multilingual terminology using a single knowledge base. 1 German-Bulgarian terminology in DB-MAT The project DB-MAT (1992-1995) aimed at the design and implementation of a translators' workbench providing linguistic and domain explanations to human translators, within the paradigm of the knowledge-based Machine Aided Translation (MAT) [1]. Most generally, the innovative idea is to integrate in the MAT workbench a domain model (a knowledge base of conceptual graphs) and to generate on the fly explanations, when the translator highlights unknown terms in the source text to be translated. The project had to deal with German and Bulgarian languages, which opened the question how to link the corresponding entries of the bilingual lexicon to the Knowledge Base (KB) entities. Figure 1 presents an early model of the pointers between the lexicon and the KB items. In principle, keeping phrasal lexicons is an acceptable strategy to support multilingual terminology (in the 80ies, several projects and prototypes of the so-called "knowledgebased term banks" seemed to approach the issue in a similar way). However, especially in DB-MAT, it became clear that the picture shown at Figure 1 has two potential "defects". First, it contains repeating information in the lexicons (see for instance all Bulgarian noun phrases including the word court). Second, the Bulgarian noun phrases are to be declined during the surface verbalisation according to complicated grammar rules (since the articles in Bulgarian are augmented at the end of the noun or the preceding adjective); so it turned out that the generation grammar for Bulgarian would work more easily with non-phrasal lexicons (since declination rules are to be supported anyway). In this way, to avoid repeating information in the (phrasal) lexicons and to provide more uniform treatment and even some elegance in the process of multilingual 1 between Hamburg University (NATS) and the Linguistic Modelling Department, Central Laboratory for Parallel Processing. Bulgarian Academy of Sciences, funded by Volkswagen Foundation (Germany)