Date of Award
Yuan, Xiaobu (Computer Science)
CC BY-NC-ND 4.0
This thesis applies the theory of naturalistic decision making (NDM) in human physcology model for the study of dialogue management system in major approaches from the classical approach based upon finite state machine to most recent approach using partially observable markov decision process (POMDP). While most of the approaches use various techniques to estimate system state, POMDP-based system uses the belief state to make decisions. In addition to the state estimation POMDP provides a mechanism to model the uncertainty and allows error-recovery. However, applying Markovian over the belief-state space in the current POMDP models cause significant loss of valuable information in the dialogue history, leading to untruthful management of user's intention. Also there is a need of adequate interaction with users according to their level of knowledge. To improve the performance of POMDP-based dialogue management, this thesis proposes an enabling method to allow dynamic control of dialogue management. There are three contributions made in order to achieve the dynamism which are as follows: Introduce historical belief information into the POMDP model, analyzing its trend and predicting the user belief states with history information and finally using this derived information to control the system based on the user intention by switching between contextual control modes. Theoretical derivations of proposed work and experiments with simulation provide evidence on dynamic dialogue control of the agent to improve the human-computer interaction using the proposed algorithm.
Dhanapal, Rajaprabhu, "An Approach for Contextual Control in Dialogue Management with Belief State Trend Analysis and Prediction" (2012). Electronic Theses and Dissertations. 319.