Smart Reply (SR) systems present a user with a set of replies, of which one can be selected in place of having to type out a response. To perform well at this task, a system should be able to effectively present the user with a diverse set of options, to maximise the chance that at least one of them conveys the user’s desired response. This is a significant challenge, due to the lack of datasets containing sets of responses to learn from. Resultantly, previous work has focused largely on post-hoc diversification, rather than explicitly learning to predict sets of responses. Motivated by this problem, we present a novel method SimSR, that employs model-based simulation to discover high-value response sets, through simulating possible user responses with a learned world model. Unlike previous approaches, this allows our method to directly optimise the end-goal of SR–maximising the relevance of at least one of the predicted replies. Empirically on two public datasets, when compared to SoTA baselines, our method achieves up to 21% and 18% improvement in ROUGE score and Self-ROUGE score respectively.

Language models pre-trained on large self-supervised corpora, followed by task-specific fine-tuning has become the dominant paradigm in NLP. These pre-training datasets often have a one-to-many structure—e.g. in dialogue there are many valid responses for a given context. However, only some of these responses will be desirable in our downstream task. This raises the question of how we should train the model such that it can emulate the desirable behaviours, but not the undesirable ones. Current approaches train in a one-to-one setup—only a single target response is given for a single dialogue context—leading to models only learning to predict the average response, while ignoring the full range of possible responses. Using text-based games as a testbed, our approach, PASA, uses discrete latent variables to capture the range of different behaviours represented in our larger pre-training dataset. We then use knowledge distillation to distil the posterior probability distribution into a student model. This probability distribution is far richer than learning from only the hard targets of the dataset, and thus allows the student model to benefit from the richer range of actions the teacher model has learned. Results show up to 49% empirical improvement over the previous state-of-the-art model on the Jericho Walkthroughs dataset.