Generating relevant/conditioned responses in dialog is challenging, and requires not only proper modelling of context in the conversation, but also the ability to generate fluent sentences during inference. In this paper, we propose a two-step framework based on →