Labelled Dependencies in Machine Translation Evaluation.pdfVIP

Labelled Dependencies in Machine Translation Evaluation Karolina Owczarzak Josef van Genabith Andy Way National Centre for Language Technology School of Computing, Dublin City University Dublin 9, Ireland {owczarzak,josef,away}@computing.dcu.ie the translation on the basis of what syntactic Abstract features it shares with the reference. Our method follows and substantially extends We present a method for evaluating the the earlier work of Liu and Gildea (2005), who use quality of Machine Translation (MT) syntactic features and unlabelled dependencies to output, using labelled dependencies evaluate MT quality, outperforming BLEU on produced by a Lexical-Functional segment-level correlation with human judgement. Grammar (LFG) parser. Our dependency- Dependencies abstract away from the particulars of based method, in contrast to most popular the surface string (and syntactic tree) realization string-based evaluation metrics, does not and provide a “normalized” representation of unfairly