DirectQE: Direct Pretraining for Machine Translation Quality Estimation

Machine Translation Quality Estimation (QE) is a task of predicting the quality machine translations without relying on any reference. Recently, predictor-estimator framework trains predictor as feature extractor, which leverages extra parallel corpora QE labels, achieving promising performance. However, we argue that there are gaps between and estimator in both data training objectives, preclude models from benefiting large number more directly. We propose novel called DirectQE provides direct pretraining for tasks. In DirectQE, generator trained to produce pseudo closer real data, detector pretrained these with objectives akin task. Experiments widely used benchmarks show outperforms existing methods, using such BERT. also give extensive analyses showing how fixing two contributes our improvements.