MATMT 2008 workshop Mixing Approaches to Machine Translation

Constructing a classifier that distinguishes machine translations from human translations is a promising approach to automatically evaluating machine-translated sentences. We developed a classifier with this approach that distinguishes translations based on word-alignment distributions between source sentences and human/machine translations. We used Support Vector Machines as machinelearning algorithms for this classifier. Our experimental results revealed that our method of evaluation had a weak correlation with human evaluations. We further found that our method outperformed wellknown automatic-evaluation metrics with respect to correlation with the manual evaluation, and that it could identify the qualitative characteristics of machine translations, which greatly help improve their quality.