An Efficient Interlingua Translation System for Multi-lingual Document Production

[MT Summit III, Washington, DC, USA, July 1991] An Efficient Interlingua Translation System for Multi-lingual Document Production* Teruko Mitamura Eric H. Nyberg, 3rd Jaime G. Carbonell Center for Machine Translation Carnegie Mellon University Pittsburgh, PA 15213 Abstract Knowledge-based interlingual machine translation systems produce semantically accurate translations, but typically require massive knowledge acquisi- tion. This paper describes KANT, a system that reduces this requirement to produce practical, scal- able, and accurate KBMT applications. First, the set of requirements is discussed, then the full KANT architecture is illustrated, and finally results from a fully implemented prototype are presented. 1 Introduction Knowledge-based machine translation holds great promise for the development of accurate, high-quality translation sys- tems. However, it has been our experience that constructing large-scale knowledge-based translation systems requires a vast amount of knowledge acquisition effort from linguistic and domain experts. It is also the case that translation systems that use a wide variety of complex knowledge sources tend to require more processing time in order to translate texts. In large-scale translation domains, more domain knowledge is required, and the amount of potential ambiguity in processing increases significantly. In this paper we describe how a knowledge-based trans- lation system can surmount these problems, while providing fast, accurate, high-quality translation in large-scale domains. We present a knowledge-based translation architecture that combines principled constraints on syntactic coverage, an adequate amount of semantic interpretation, an independent generation module, and the use of semi-automated knowledge acquisition tools in support of large-scale multi-lingual trans- lation applications. Then we illustrate this architecture with a fully-implemented prototype for efficient, high-accuracy translation. 2 Objectives We assume the knowledge-based or interlingua framework for translation, as illustrated in Figure 1. The source text is first analyzed into a language-independent intermediate rep- resentation, or Interlingua. Then the target language text is produced from the Interlingua representation (Carbonell and Tomita, 1987). *Proceedings of Machine Translation Summit III, Washington, DC, July 2-4, 1991. Figure 1: Interlingua Translation Our active research plan is to design and implement a large- scale system for multi-lingual document production, based on the prototype system discussed in Section 3. Once imple- mented on a large scale, the complete system will support authoring in a single controlled source language (technical English) and provide on-demand translation of the source text into multiple languages at remote locations. As a result, our design for an Interlingua-based machine translation system must provide efficient, on-line translation to multiple output languages, requiring no post-editing. In this section we de- scribe each of our objectives in turn, along with the features required in a translation system that meets these objectives. Then, we describe the fully-implemented prototype system that demonstrates their feasibility. • No Post-Editing. A system that requires no post-editing must achieve a high degree of semantic and grammatical accuracy at every stage of processing, and maintain a high level of stylistic quality in the generation of target language text. • No Human Help in Disambiguation, Once a document has been authored, a fully-automated system can resolve ambiguity that arises …

Content truncated to 500 words