Towards Text-Based Machine Translation

Towards Text-Based Machine Translation Jörg Schütz and Bärbel Ripplinger IAI Martin-Luther-Straße 14 D-66111 Saarbrücken [joerg, babs]@iai.uni-sb.de Abstract The increasing use of world wide networking facilities for international information exchange and service provision demands for specific applications of machine translation (MT) tools which are beyond their traditional tasks. The new areas of application of MT will have to be centered around various human-to-human communication media taking into account the different cultures inherent in natural language. Thus, MT has to aim for cross-cultural communication when being used as an information and service provider between different natural languages on the information super-highways. Most of the information electronically exchanged or accessed deals with a specialized vocabulary (terminology). In this paper we will argue for terminology-based MT and discuss the various knowledge sources, which will also ensure an appropriate basis for text-based MT. We have restricted our work to the domain of telecommunications, in particular that of satellite communication. 1 Introduction and Motivation Over the last decade (1984-1994) machine translation systems, commercial systems as well as research prototypes, have steadily improved regarding their linguistic capabilities. However, they still lack the command of language a human translator possesses, in particular with regard to the interpretation of the textual units to be translated in their contextual, situational and cultural background environments. This includes, for example, the correct identification of ambiguous subject/object positions, as in 'Fernübertragungsausrüstungen umfassen auch Modulationsgeräte.' (Telecommunication equipment also comprises modulating equipment.), of metaphorical meanings, like 'Der Prozeß wurde abgeschossen.' (The process was killed), or of pro-forms within and across sentence boundaries. Our aim is to provide for comprehensive knowledge resources which are organized so that sublanguage information and extra-linguistic information about a specific domain is accessible in a concise, efficiently machine-tractable form, and which are formalized so as to ensure consistency across organizations of related grammatical and lexical strata. However, the depth of the natural language (NL) analysis is restricted to the needs of the translation task, i.e. the access to the different information sources is defined by the application. To achieve this, different knowledge sorts are used to build a sublanguage information repository which can be applied within a constraint-based natural language processing (NLP) framework. The approach has been implemented and tested in the Advanced Language Engineering Platform (ALEP) environment ([ALEP, 1993]), a general purpose NLP development platform, based on an 185 object-centered architecture and a typed feature logic based linguistic formalism, promoted by the European Commission (EC) for the Linguistic Research and Engineering (LRE) action line and the forthcoming Fourth Framework Programme. The primary driving force of our MT approach is the conceptual organization of the domain of telecommunications. Its purpose is to provide domain-specific constraints that ensure the control of the analysis, translation and generation process of sublanguage expressions. On the one hand, this is done by providing a model of the domain - the ontology - which represents the concepts of the domain and their generic and partitive relationships. On the other hand, knowledge about the terminology of the domain …

Content truncated to 500 words