Towards the calibration of translation errors: the why and the how