Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Statistical dialogue management systems

Poulakis Konstantinos

Full record

---

Summary

Modern Statistical Dialogue Managers (SDMs) have made significant strides in providing robust and efficient human-machine interaction. This progress is based both on the large amount of data and on the development of innovative algorithms based on augmented learning technique. Extensive use of these systems in real-world voice systems (Spoken Dialogue Systems (SDS)) can reduce development and maintenance costs and increase systems tolerance for uncertainty due to both environmental and subsystem errors used, such as Automatic Speech Recognition (ASR) or Natural Language Understanding (NLU). In such an uncertain environment in which each decision on the dialogue situation is made serially with direct dependence on previous decisions, one of the most appropriate models used is based on Partially Observable Markov Decision Process (POMDP). In practice, the large number of situations and actions in the dialogue, as well as the dimension of the observations, make it computationally impossible to optimize the model. Therefore, the practical implementation of POMDP-based systems requires the development of effective algorithms and approaches. In this diploma thesis, we attempt a detailed overview of the methods and techniques that have been developed to create SDM. We first focus on POMDPs-based systems and look at different methods of representing the dialog state space in order to reduce computational costs and improve results. Then, a comparison is made between different learning methods, both linear and non-linear based on deep neural networks. The results from a series of experiments conducted in the PyDial environment using artificial data from a simulator show that the technology is very promising.

Available Files

Services