CORAL: Contextual Response Retrievability Loss Function for Training Dialog Generation Models

Abstract

Natural Language Generation (NLG) represents a large collection of tasks in the field of NLP. While many of these tasks have been tackled well by the cross-entropy (CE) loss, the task of dialog generation poses a few unique challenges for this loss function. First, CE loss assumes that for any given input, the only possible output is the one available as the ground truth in the training dataset. In general, this is not true for any task, as there can be multiple semantically equivalent sentences, each with a different surface form. This problem gets exaggerated further for the dialog generation task, as there can be multiple valid responses (for a given context) that not only have different surface forms but are also not semantically equivalent. Second, CE loss does not take the context into consideration while processing the response and, hence, it treats all ground truths with equal importance irrespective of the context. But, we may want our final agent to avoid certain classes of responses (e.g. bland, non-informative or biased responses) and give relatively higher weightage for more context-specific responses. To circumvent these shortcomings of the CE loss, in this paper, we propose a novel loss function, CORAL, that directly optimizes recently proposed estimates of human preference for generated responses. Using CORAL, we can train dialog generation models without assuming non-existence of response other than the ground-truth. Also, the CORAL loss is computed based on both the context and the response. Our experiments show that, against various large and small scale baselines, our model is able to obtain significantly higher scores for the human preference estimate. While large-scale dialog models have shown much promise and aptitude for dialog generation, it is important to continue the search for more suitable loss functions for training dialog systems. Extensive comparisons on two benchmark datasets show that the proposed methods outperform strong state-of-the-art baseline models of different sizes. We will make the code and trained model checkpoints publicly available upon publication of this paper.Preprint. Under review.