DEMOCRAT: deciding between multiple outputs created by automatic translation

We present DEMOCRAT, a system which DEcides between Multiple Outputs CReated by Automatic Translation, specifically free online MT systems, and tries to construct from them a consensus translation which, it is hoped, will take the best elements of the contributing systems and produce and output as good as or better than any of the individual MT systems on their own. We review the small number of previous implementations of this variation of multi-engine MT, contrasting our own algorithm with those reported elsewhere. Other implementations all use language models and require extensive training, whereas DEMOCRAT is “plug-and-play”. Also, they have been evaluated only on single language pairs and texts, whereas we have experimented with a larger variety, and found the results to be variable, but consistent. We therefore consider what factors contribute to DEMOCRAT’s performance.