Recent developments in machine learning along with a general shift in the public attitude towards digital personal assistants has opened new frontiers for conversational systems. Nevertheless, building data-driven multi-domain conversational agents that act optimally given a dialogue context is an open challenge. The first step towards that goal is developing an efficient way of learning a dialogue policy in new domains. Secondly, it is important to have the ability to collect and utilise human-human conversational data to bootstrap an agent's knowledge. The work presented in this thesis demonstrates how a neural dialogue manager fine-tuned with reinforcement learning presents a viable approach for learning a dialogue policy efficiently and across many domains.

The thesis starts by introducing a dialogue management module that learns through interactions to act optimally given a current context of a conversation. The current shift towards neural, parameter-rich systems does not fully address the problem of error noise coming from speech recognition or natural language understanding components. A Bayesian approach is therefore proposed to learn more robust and effective policy management in direct interactions without any prior data. By putting a distribution over model weights, the learning agent is less prone to overfit to particular dialogue realizations and a more efficient exploration policy can be therefore employed. The results show that deep reinforcement learning performs on par with non-parametric models even in a low data regime while significantly reducing the computational complexity compared with the previous state-of-the-art.

The deployment of a dialogue manager without any pre-training on human conversations is not a viable option from an industry perspective. However, the progress in building statistical systems, particularly dialogue managers, is hindered by the scale of data available. To address this fundamental obstacle, a novel data-collection pipeline entirely based on crowdsourcing without the need for hiring professional annotators is introduced. The validation of the approach results in the collection of the Multi-Domain Wizard-of-Oz dataset (MultiWOZ), a fully labeled collection of human-human written conversations spanning over multiple domains and topics. The proposed dataset creates a set of new benchmarks (belief tracking, policy optimisation, and response generation) significantly raising the complexity of analysed dialogues.

The collected dataset serves as a foundation for a novel reinforcement learning (RL)-based approach for training a multi-domain dialogue manager. A Multi-Action and Slot Dialogue Agent (MASDA) is proposed to combat some limitations: 1) handling complex multi-domain dialogues with multiple concurrent actions present in a single turn; and 2) lack of interpretability, which consequently impedes the use of intermediate signals (e.g., dialogue turn annotations) if such signals are available. MASDA explicitly models system acts and slots using intermediate signals, resulting in an improved task-based end-to-end framework. The model can also select concurrent actions in a single turn, thus enriching the representation of the generated responses. The proposed framework allows for RL training of dialogue task completion metrics when dealing with concurrent actions. The results demonstrate the advantages of both 1) handling concurrent actions and 2) exploiting intermediate signals: MASDA outperforms previous end-to-end frameworks while also offering improved scalability.