DTED: Evaluation of Machine Translation Structure Using Dependency Parsing and Tree Edit Distance
Martin McCaffery, Mark-Jan Nederhof · 2016
We present DTED, a submission to the WMT 2016 Metrics Task using structural information generated by dependency parsing and evaluated using tree edit distances.In this paper we apply this system to translations produced during WMT 2015, and compare our scores with human rankings from that year.We find moderate correlations, despite the human judgements being based on all aspects of the sentences while our metric is based only on word order.