A method for in-depth comparative evaluation: How (dis)similar are outputs of pos taggers, dependency parsers and coreference resolvers really?