Online and Multitask Learning for Machine Translation Quality Estimation in Real-world Scenarios

We investigate the application of different supervised learning approaches to machine translation quality estimation in realistic conditions where training data are not available or are heterogeneous with respect to the test data. Our experiments are carried out with two techniques: online and multitask learning. The former is capable to learn and self-adapt to user feedback, and is suitable for situations in which training data is not available. The latter is capable to learn from data coming from multiple domains, which might considerably differ from the actual testing domain. Two focused experiments in such challenging conditions indicate the good potential of the two approaches.