Active Learning for Neural Machine Translation

Neural machine translation (NMT) normally requires a large bilingual corpus to train a high-translation-quality model. However, building such parallel corpora for many low-resource language pairs is rather expensive. In this paper, we propose to select informative source sentences to build a parallel corpus under the active learning framework so as to reduce the cost of manual translation as much as possible. Particularly, we propose two novel and effective sentence selection methods for active learning: selection based on semantic similarity and decoder probability. Experiments on Indonesian-English and Chinese-English show that our selection approaches are superior to random selection and two conventional selection methods.