Towards structured neural spoken dialogue modelling

Resumen

In this thesis, we try to alleviate some of the weaknesses of the current approaches to dialogue modelling,one of the most challenging areas of Artificial Intelligence. We target three different types of dialogues(open-domain, task-oriented and coaching sessions), and use mainly machine learning algorithms to traindialogue models. One challenge of open-domain chatbots is their lack of response variety, which can betackled using Generative Adversarial Networks (GANs). We present two methodological contributions inthis regard. On the one hand, we develop a method to circumvent the non-differentiability of textprocessingGANs. On the other hand, we extend the conventional task of discriminators, which oftenoperate at a single response level, to the batch level. Meanwhile, two crucial aspects of task-orientedsystems are their understanding capabilities because they need to correctly interpret what the user islooking for and their constraints), and the dialogue strategy. We propose a simple yet powerful way toimprove spoken understanding and adapt the dialogue strategy by explicitly processing the user's speechsignal through audio-processing transformer neural networks. Finally, coaching dialogues shareproperties of open-domain and task-oriented dialogues. They are somehow task-oriented but, there is norush to complete the task, and it is more important to calmly converse to make the users aware of theirown problems. In this context, we describe our collaboration in the EMPATHIC project, where a VirtualCoach capable of carrying out coaching dialogues about nutrition was built, using a modular SpokenDialogue System. Second, we model such dialogues with an end-to-end system based on TransferLearning.