A Re-examination of Machine Learning Approaches for Sentence-Level MT Evaluation