Abstract
A major bottleneck for the large diffusion of data-driven conversational agents is that conversational domains are subject to continuous changes, which soon make initial dialogue models inadequate to manage new situations. In the current context, updating training data is usually carried on manually, and, in addition, there are no tools for simulating the impact of a certain domain change on the performance of the dialogue system. This position paper advocates that substantial progress in the capacity to simulate domain changes is based on the ability to automatically adapt training and test dialogues to those changes. We discuss the potential of a simulation framework for task-oriented dialogues, as well as the research challenges that need to be addressed.