Abstract
Responses generated by neural conversational models tend to lack informativeness and diversity. We present a novel adversarial learning method, called Adversarial Information Maximization (AIM) model, to address these two related but distinct problems. To foster response diversity, we leverage adversarial training that allows distributional matching of synthetic and real responses. To improve informativeness, we explicitly optimize a variational lower bound on pairwise mutual information between query and response. Empirical results from automatic and human evaluations demonstrate that our methods significantly boost informativeness and diversity.