Sem Resumo

Conversational agents or chatbots are increasingly employed in commercial applications to answer questions and to recommend items. Despite their success, they usually behave as black-boxes from the user perspective, typically failing to produce high quality human-computer interactions. Thus interpretability is a major concern for the next generation of recommendation systems. This work addresses challenges related to the development of a recommendation system that can explain its own suggestions. Furthermore, this work evaluates the impact of dierent explanation generation techniques both in simulated interactions and in tests with human subjects. This work present novel model-agnostic methods that address challenges of explanation generation in the context of knowledge embedding based conversational recommendation systems, such as: explanation delity, graph incompleteness, time to response constraints and reasons against generation. Finally, this research evaluates the technical feasibility of such methods with simulated experiments and shows preliminary on user perception of the generated explanations.