Differentiable Sampling with Flexible Reference Word Order for Neural Machine Translation
This addresses exposure bias in machine translation, offering a more accurate and efficient training method, though it is incremental as it builds on existing scheduled sampling approaches.
The paper tackled the misalignment assumption in scheduled sampling for neural machine translation by introducing a differentiable sampling algorithm that optimizes the probability of aligning the reference with the sampled output using soft alignments. Experiments on IWSLT tasks showed BLEU improvements over baselines, with simpler training and better performance at smaller beam sizes.
Despite some empirical success at correcting exposure bias in machine translation, scheduled sampling algorithms suffer from a major drawback: they incorrectly assume that words in the reference translations and in sampled sequences are aligned at each time step. Our new differentiable sampling algorithm addresses this issue by optimizing the probability that the reference can be aligned with the sampled output, based on a soft alignment predicted by the model itself. As a result, the output distribution at each time step is evaluated with respect to the whole predicted sequence. Experiments on IWSLT translation tasks show that our approach improves BLEU compared to maximum likelihood and scheduled sampling baselines. In addition, our approach is simpler to train with no need for sampling schedule and yields models that achieve larger improvements with smaller beam sizes.