A Comprehensive and Practical Model of Memory-Based Machine Translation

Abstract

This paper presents a comprehensive and practical model of memory-based machine translation (MBMT), also called example-based machine translation (EBMT). Although MBMT is gaining increasing interest, models proposed so far are incomplete in that these models focus on only one aspect of translation, and ignore involvement of rules which are available as expert knowledge. Goals of this paper are to present (1) a comprehensive architecture for MBMT to serve as a basis for discussion on how real systems should be built, (2) a novel method for integrating memory-based approach and rule-based approach, which would enable handling linguisitc phenomena, such as zero-pronoun, and use of expert knowledge available as rules, (3) an optimal representation of translation examples, so that users can incrementally customize the system, (4) a robust matching scheme, using a version of Dynamic Programming, to attain robust and broadcoverage translation. A part of the model has been implemented on massively parallel machines. At the same time, the preliminary evaluation confirmed the improvments on the translation quality and user acceptance over pure MBMTs.