Learning to Stop in Structured Prediction for Neural Machine Translation
This work addresses a specific bottleneck in neural machine translation for practitioners, offering incremental improvements to beam search optimization.
The paper tackled the problem of beam search lacking principled stopping criteria in neural machine translation, which leads to a bias toward longer hypotheses, and proposed a novel ranking method and structured prediction loss that improved length and BLEU scores in experiments on synthetic and real language data.
Beam search optimization resolves many issues in neural machine translation. However, this method lacks principled stopping criteria and does not learn how to stop during training, and the model naturally prefers the longer hypotheses during the testing time in practice since they use the raw score instead of the probability-based score. We propose a novel ranking method which enables an optimal beam search stopping criteria. We further introduce a structured prediction loss function which penalizes suboptimal finished candidates produced by beam search during training. Experiments of neural machine translation on both synthetic data and real languages (German-to-English and Chinese-to-English) demonstrate our proposed methods lead to better length and BLEU score.