Contributions to Attributed Probabilistic Finite State Bi-Automata for Dialogue Management

Task-oriented Spoken Dialogue Systems (SDSs), also known as Conversational Assistants, have been generating a great deal of interest in recent years, as they can be used to automate repetitive and low-value tasks and processes which use a natural communication channel. One basic component of every task-oriented SDS is the Dialogue Manager (DM), which is responsible for tracking the current state of the conversation and for deciding the next action of the system.

This dissertation intends to improve a data-driven framework based on stochastic finite-state transducers for DM modelling in task-oriented SDSs: the Attributed Probabilistic Finite State Bi-Automata (A-PFSBA). Several contributions are presented that enhance the A-PFSBA based DM in different aspects. First, its model generalisation mechanism is improved to better employ context, the semantic relation between dialogue states and the spatial relations of the dialogue state space. Second, the A-PFSBA theoretical framework is extended for policy-making. In the same way, multiple policies with different degrees of complexity are implemented following this formulation. Third, a simple-yet-effective algorithm is proposed to incrementally learn an initial DM, which can be adjusted to work under uncertainty. Finally, the potential of the A-PFSBA framework to be deployed in data scarcity and zero-data scenarios and its capability to bridge the gap between data-driven and rule-based paradigms for DM development is tested. 

The presented contributions have been validated using two well-known corpora: the Let's Go corpus and the Dialogue State Tracking Challenge 2 corpus. In order to validate the viability of the A-PFSBA framework in industrial scenarios, three applications that employ the A-PFSBA formulation and which have been validated by real users are also presented.