On the practice of error analysis for machine translation evaluation