Rewriter-Evaluator Architecture for Neural Machine Translation
This work addresses the problem of improving NMT translation quality for users by introducing a multi-pass decoding architecture with a learned termination policy, representing an incremental improvement over existing multi-pass methods.
The paper introduces a Rewriter-Evaluator architecture for Neural Machine Translation (NMT) that iteratively refines translations and uses an evaluator to determine termination. This approach notably improves NMT model performance on Chinese-English and English-German translation tasks, outperforming previous baselines.
Encoder-decoder has been widely used in neural machine translation (NMT). A few methods have been proposed to improve it with multiple passes of decoding. However, their full potential is limited by a lack of appropriate termination policies. To address this issue, we present a novel architecture, Rewriter-Evaluator. It consists of a rewriter and an evaluator. Translating a source sentence involves multiple passes. At every pass, the rewriter produces a new translation to improve the past translation and the evaluator estimates the translation quality to decide whether to terminate the rewriting process. We also propose prioritized gradient descent (PGD) that facilitates training the rewriter and the evaluator jointly. Though incurring multiple passes of decoding, Rewriter-Evaluator with the proposed PGD method can be trained with a similar time to that of training encoder-decoder models. We apply the proposed architecture to improve the general NMT models (e.g., Transformer). We conduct extensive experiments on two translation tasks, Chinese-English and English-German, and show that the proposed architecture notably improves the performances of NMT models and significantly outperforms previous baselines.