Milica Gasic:
Neural dialogue models: growing up confident and independent

Abstract:
Neural approaches to dialogue modelling have produced state of the art results in each dialogue subtask, ranging from dialogue state tracking to user modelling. One of the biggest advantages of neural architectures for dialogue is their potential for end-to-end dialogue modelling. However, these, in many ways powerful, models have their Achilles’ heels. Their weaknesses are particularly apparent in three learning aspects, namely independence, confidence and interaction. In this talk, I will show how each of these aspects can be addressed in neural approaches to dialogue state tracking, dialogue policy estimation and user modelling.