Learning part-of-speech taggers with inter-annotator agreement loss