THE NEW GENERATION BEHAVIORTRAN: DESIGN PHILOSOPHY AND SYSTEM ARCHITECTURE

In many conventional machine translation systems, the translation outputs are usually strongly affected by the syntactic information of the source sentences and thus tend to produce literal translations that are not natural to the native speakers. In this paper, we introduce the design philosophy and system architecture of the new generation BehaviorTran, which will enable an MT system to operate with high modularity and to acquire its translation knowledge from a bilingual corpus with a two-way training method. In such a paradigm, the knowledge bases only provide static descriptions on the legal forms of the constructs, while ambiguity resolution and preference evaluation are governed by sets of statistical parameters. This makes it easier to adapt the system to specific user styles and maintain different parameter sets for different customers. Thus, it is expected to be a promising paradigm for producing satisfactory translations.