Batch-level GANs to promote dialogue response variety

Abstract

Exploiting large pretrained transformers has become one of the most popular approaches for dialogue modelling. Nonetheless, due to their lack of robustness and explainability, we believe it is necessary to keep exploring alternative methodologies. In this work, we focus on Generative Adversarial Networks (GANs) for open-domain dialogue generation. In particular, we extend the idea of conventional GAN discriminators, which operate at a single response level, to the batch level. Our proposed discriminator evaluates how human a set of responses are for the corresponding dialogue contexts. We show that batch-level GANs outperform response-level GANs and a MLE baseline in terms of variability, without hurting the semantic coherence, according to our metrics. We believe that our proposal could benefit future work in GAN-related research, as well as other AI systems that employ discriminators.