Conversational AI 11-Statistical Data-Driven Dialogue Systems
While rule-based dialogue systems are still widely used, particularly in commercially deployed dialogue systems, an alternative approach involving machine learning from data has come to dominate current dialogue systems research and is also being increasingly applied in commercial systems. There have been two main phases in the history of statistical data-driven dialogue systems. In the first phase, systems continued to be developed using the modular architecture but efforts were directed toward the use of machine learning techniques to optimize the components of the architecture, in particular the Natural Language Understanding (NLU), Dialogue Manager (DM), and Natural Language Generation (NLG) components.
MOTIVATING THE STATISTICAL DATA-DRIVEN APPROACH
In a statistical data-driven dialogue system the processes involved in the various components of the system are modeled probabilistically. For example, the system’s dialogue strategy is based on uncertain information that has been derived from the output of the ASR and NLU components. As a result, the system’s belief state, i.e., its beliefs about the current state of the dialogue, in particular its beliefs about what the user has said, are uncertain and DM has to maintain a distribution over multiple hypotheses of the dialogue state. Learning in a statistical data-driven dialogue system is data-driven. The data can take various forms. On the one hand, it can be data from previous dialogues that are similar in domain to that of the current system.