Accuracy and robustness in measuring the lexical similarity of semantic role fillers for automatic semantic MT evaluation