MACHINE TRANSLATION SEEN AS INTERACTIVE MULTILINGUAL TEXT GENERATION
HAROLD L. SOMERS, Danny Jones · 1992
In this paper we describe the design of what we hope is a radically new Machine Translation (MT) system, for use by a monolingual user wishing to construct texts in a specific and fairly restricted domain, which will then be generated multilingually. The envisaged user is the original author of a text to be composed on-line (and interactively). The system is rooted in a specific domain, and this aspect permits the system to interpret the user's input, and even make predictions about message content, both of which will contribute to high quality of translation. The domain that we are working in at present is the generation of multilingual job advertisements in semi-technical fields such as computer design, management or graphic arts, where the new career flexibility offered by the forthcoming open market suggests an obvious application. The domain has the advantage of moderately fixed text formats with some degree of flexibility in content, a combination which suits the particular approach we are taking. Other domains with similar properties might include: • in a multinational conglomerate, exchange of messages between e.g. suppliers concerning orders for and availability of goods, spare parts, etc., or offers and tenders for services, equipment and so on. • customers' and/or field engineers' initial maintenance/break-down reports on equipment (one can think of major manufacturers of electronic appliances based in an English-speaking country, but exported world-wide). • communication between military or police forces (e.g. in NATO, or Interpol, or the need for Anglo-French communication in the Channel Tunnel). The mode of operation is as an interactive system with a flexible human-machine interface (systemand/or user-driven) where the system cooperates with the user in order to produce a guaranteed quality translation for transmission. The system's linguistic and domain-specific knowledge bases are closely related, in that translation equivalents are tied to situational contexts. Furthermore, its domain knowledge serves to guide the user towards acceptable and standardized message structures. The system is nevertheless flexible enough to permit more than a rigid 'phrase-book' type of message generation. The system will accept input in English and generate texts in English, Spanish and Greek.