Abstract
Despite important progress, conversational systems often generate dialogues that sound un-natural to humans. We conjecture that the reason lies in the different training and testing conditions: agents are trained in a controlled “lab” setting but tested in the “wild”. During training, they learn to utter a sentence given the ground-truth dialogue history generated by human annotators. On the other hand, during testing, the agents must interact with each other, and hence deal with noisy data. We propose to fill this gap between the training and testing environments by training the model with mixed batches containing both samples of human and machine-generated dialogues. We assess the validity of the proposed method on GuessWhat?!, a visual referential game. We show that our method improves the linguistic quality of the generated dialogues, and it leads to higher accuracy of the guessing task; simple perturbations of the ground-truth dialogue history that mimic machine-generated data do not account for a similar improvement. Finally, we run a human evaluation experiment on a sample of machine-machine dialogues to complement the quantitative analysis. This experiment shows that also human annotators successfully exploit dialogues generated by a model trained with mixed batches to solve the task. Hence, the mixed-batch training does not cause a language drift. Moreover, we find that the new training regime allows human annotators to be significantly more confident when selecting the target object, showing that the generated dialogues are informative. © 2022 Associazione Italiana di Linguistica Computazionale.