Building A Chatbot In Healthcare Domain Based On Dialog Policy Learning

Since December 2019,COVID-19 has spread the world and the number of infections and deaths around the world has increased dramatically.By the Mid-April 2020,the cumulative number of infections worldwide has exceeded one million,and the cumulative number of deaths has exceeded one hundred thousand.As the epidemic continues to develop,researchers all over the world are endeavoring to contribute their parts to this battle.Based on the current background of the global fight against the epidemic,this thesis applies dialog policy learning in build a chatbot that will be used in the domain of public healthcare services.It provides services such as consultation about this coronavirus,hotline help,and appointment with the hospital to help fight the global epidemic.In the application of chatbot for COVID-19,most of the previous ones apply FAQ or rule-based dialog policy.The innovation of this thesis applies dialog policy learning with supervised learning and deep reinforcement learning in building an anti-epidemic chatbot.As the brain of a chatbot,the dialog policy determines whether a chatbot system is succeed or failed.For task-oriented chatbots,supervised learning and deep reinforcement learning are two popular research approaches of dialog policy learning.Compared with FAQ and rule-based dialog policy,both approaches the above will make chatbots more flexible and intelligent due to their way of data-driven.This thesis is devoted to the realization of a chatbot serving the public healthcare domain.It proposes to use an approach based on supervised learning and another one based on deep reinforcement learning to solve the learning problem of dialog policy for chatbots.The above two dialog policy learning approaches can allow chatbots to expand to more domains and complete more service tasks.Combined with the task of appointment with the hospital in healthcare domain,this thesis has clarified and analyzed the model design,model construction,model implementation,experience results,advantages,disadvantages,and applicability of these two dialog policy learning approaches in detail.This thesis mainly covers the following:(1)Construct a dialog policy for the healthcare chatbot based on the approach of supervised learning.This dialog policy learning approach uses a hybrid code networks.Due to the addition of some artificial rule interventions,compared with other similar approaches to complete the task of dialog policy learning based on supervised learning,the hybrid code networks greatly reduces the amount of data required for training,while retaining the key benefit of inferring a latent representation of dialog state.In this thesis,the dialog policy based on supervised learning will try to "imitate" the way of humancomputer dialog in the corpus as much as possible.This dialog policy model has advantages of being lightweight and easy to implement.And the performance after training on small-scale datasets is also very good.(2)Construct a dialog policy for the healthcare chatbot based on the approach of deep reinforcement learning.This dialog policy learning approach uses the deep reinforcement learning algorithm DQN as an agent.Unlike the dialog policy trying to imitate the corpus as much as possible based on the approach of supervised learning,the dialog policy learning based on the approach of deep reinforcement learning are more exploratory and flexible.This thesis uses ?-greedy exploration strategy for exploration and utilization when training DQN,which explores the reward values in different dialog states as much as possible.The dialog policy model based on deep reinforcement learning shows a good advantage in the personalized policy exploration and the average dialog turns which are required to complete a task.(3)Construct dialog datasets and verify the performance of dialog policy learning approaches.This thesis constructs dialog datasets used in the fight against epidemic combined with the background of the epidemic raging.Based on the constructed datasets,this thesis has used the above two dialog policy learning approaches to train and obtained two dialog policy models.In the test phase,both models completes the task of appointment with the hospital with success rates of more than 97.8%;and in the interaction phase with the user,compared with the rule-based dialog policy,the decision-making performance of the healthcare chatbot using the above two models is more efficient and intelligent.