Comparing Rule-based and Statistical MT Output

This paper describes a comparison between a statistical and a rule-based MT system. The first section describes the setup and the evaluation results; the second section analyses the strengths and weaknesses of the respective approaches, and the third tries to define an architecture for a hybrid system, based on a rule-based backbone and enhanced by statistical intelligence. This contribution originated in a project called “Translation Quality for Professionals” (TQPro) which aimed at developing translation tools for professional translators. One of the interests in this project was to find a baseline for machine translation quality, and to extend MT quality beyond it. The baseline should compare stateof-the-art techniques for both statistical packages and rulebased systems, and draw conclusions from the comparison. This paper presents some insights into the results of this work.