Research On Modeling Method Of Affective Dialogue Management And Its Application

As a new direction in the field of artificial intelligence,affective computing aims to build a harmonious human-computer interaction environment and to make the computer have higher,and more comprehensive intelligence,by enabling the computer to recognize,understand,express and adapt to human emotions.Affective dialogue system is an important way to realize the multi-modal human-computer emotional interaction,which is the core subject of affective computing.Affective dialogue management is the core module of affective dialogue system,this dissertation mainly studies its modeling method and application.The starting point of this study is to build several effective and reasonable affective dialogue management models for specific dialogue areas,based on the existing dialogue management model and taking the user's emotional information factors into consideration.The main research contents of this dissertation include:(1)We propose a new affective dialogue management model based on two finite state machines(TFSM).We construct two finite state machines to model the user and the system respectively,and simulate the dialogue process as an information exchange between the two state machines.In this way,we build a framework of affective dialogue system which aims to provide the service of information inquiry and affective interaction.We further discuss the implementation methods of information inquiry and emotional response modules,including task-based multi-topic information inquiry and sub-state machine for several rounds of human computer affective interaction.In addition,we employ the return value function of partially observable Markov decision process(POMDP)model to analyze and evaluate the system performance of TFSM-based dialogue management model.The experimental results not only show that a separate finite state machine user model can help the system predict user's dialogue states,but also confirm that TFSM-based dialogue management model outperforms the conventional FSM model.(2)We propose a novel factored POMDP model for affective dialogue management.In order to effectively respond to user's affective state in dialogue process,user's state space is factored into two distinct components:goal and emotion,and the system's action space is for the first time factored into two parts,i.e.,goal response and emotion response.The reward function is accordingly updated by weighted sum of the two-part rewards.On this basis,we establish user's goal model,user's emotion model and system's observation model,and then propose a factored POMDP-based affective dialogue management model,which aims to respond to both the user's intention and emotion in a dialogue.We take intelligent music player as an example to illustrate the practical application of the model,and evaluate the influence of key parameters on the system performance.The simulation results demonstrate that the proposed model is rational and feasible for designing intelligent music player.(3)We study the portability of POMDP based affective dialogue management model.For the specific dialogue domain of affective tutoring system,in order to fulfill the intended task of concepts learning and affective interaction,we divide the user's state space into three components:goal,congnitive and emotion,while the system's action space is still factored into two parts:goal response and emotion response.On this basis,we extend the previous factored POMDP model,and propose a new factored POMDP based dialogue management model to build an affective tutoring system.The simulation results show that the affective tutoring system based on the proposed model is more beneficial to those learners with poor learning ability,and is suitable for concepts learning of larger scale.All these results to some extent prove the validity and feasibility of the proposed model.(4)We systematically review the advance of modeling method and application of dialogue management,and analyze in details the model principles,advances,limitations and advantages of several common dialogue management models,including finite state machine,slot-filling,Markov decision process(MDP),POMDP,example-based,plan-based,Bayesian network-based,and so on.These studies laid the theoretical foundation for the modeling method innovation of affective dialogue management.