Syntax and Semantics in Quality Estimation of Machine Translation

We employ syntactic and semantic information in estimating the quality of machine translation from a new data set which contains source text from English customer support forums and target text consisting of its machine translation into French. These translations have been both post-edited and evaluated by professional translators. We find that quality estimation using syntactic and semantic information on this data set can hardly improve over a baseline which uses only surface features. However, the performance can be improved when they are combined with such surface features. We also introduce a novel metric to measure translation adequacy based on predicate-argument structure match using word alignments. While word alignments can be reliably used, the two main factors affecting the performance of all semantic-based methods seems to be the low quality of semantic role labelling (especially on ill-formed text) and the lack of nominal predicate annotation.