Capturing Lexical Variation in MT Evaluation Using Automatically Built Sense-Cluster Inventories