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