Differentiable Expected BLEU for Text Generation

Neural text generation models such as recurrent networks are typically trained by maximizing data log-likelihood based on cross entropy. Such training objective shows a discrepancy from test criteria like the BLEU metric. Recent work optimizes expected BLEU under the model distribution using policy gradient, while such algorithm can suffer from high variance and become impractical. In this paper, we propose a new Differentiable Expected BLEU (DEBLEU) objective that permits direct optimization of neural generation models with gradient descent. We leverage the decomposability and sparsity of BLEU, and reformulate it with moderate approximations, making the evaluation of the objective and its gradient efficient, comparable to common cross-entropy loss. We further devise a simple training procedure with ground-truth masking and annealing for stable optimization. Experiments on neural machine translation and image captioning show our method significantly improves over both cross-entropy and policy gradient training.