Lightweight Word-Level Confidence Estimation for Neural Interactive Translation Prediction
Rebecca E. Knowles, Philipp Koehn · Edinburgh Research Explorer (University of Edinburgh) · 2018
In neural interactive translation prediction, a system provides translation suggestions (“auto-complete” functionality) for human translators. These translation suggestions may be rejected by the translator in predictable ways; being able to estimate confidence in the quality of translation suggestions could be useful in providing additional information for users of the system. We show that a very small set of features (which are already generated as byproducts of the process of translation prediction) can be used in a simple model to estimate confidence for interactive translation prediction