Sentence Level Human Translation Quality Estimation with Attention-based Neural Networks

This paper explores the use of Deep Learning methods for automatic estimation\nof quality of human translations. Automatic estimation can provide useful\nfeedback for translation teaching, examination and quality control.\nConventional methods for solving this task rely on manually engineered features\nand external knowledge. This paper presents an end-to-end neural model without\nfeature engineering, incorporating a cross attention mechanism to detect which\nparts in sentence pairs are most relevant for assessing quality. Another\ncontribution concerns of prediction of fine-grained scores for measuring\ndifferent aspects of translation quality. Empirical results on a large human\nannotated dataset show that the neural model outperforms feature-based methods\nsignificantly. The dataset and the tools are available.\n