Applying Referenceless MT Evaluation Metrics to a Multi-Engine MT System

Recent work in the area of automatic machine translation evaluation has led to the development of metrics that have equivalent levels of performance with standard metrics, yet do not rely on the existence of human generated gold standard reference translations. To achieve this performance, alternative machine generated translation are used in place of the human references. A seemingly natural application of such work is to utilize it as part of a multi-engine machine translation system by allowing the evaluation metric to select the best candidate from a list composed of possible sentences from a number of different MT systems. This possibility is explored, and a number of factors affecting the performance of the resulting MEMT systems are identified. No significant gains have yet been made with this method, however.