Source-Language Features and Maximum Correlation Training for Machine Translation Evaluation

We propose three new features for MT evaluation: source-sentence constrained n-gram precision, source-sentence reordering metrics, and discriminative unigram precision, as well as a method of learning linear feature weights to directly maximize correlation with human judgments. By aligning both the hypothesis and the reference with the sourcelanguage sentence, we achieve better correlation with human judgments than previously proposed metrics. We further improve performance by combining individual evaluation metrics using maximum correlation training, which is shown to be better than the classification-based framework.