Cas-GAN: An Approach Of Dialogue Policy Learning Based On Gcn And Rl Techniques

Dialogue management systems are commonly applied in daily life,such as online shopping,hotel booking,and driving booking.In dialogue management systems,the user is interacting with the system by multi-turn dialogues.The efficient dialogue management policy help systems to respond to the user in an effective way.Policy learning is a complex task to build the dialogue system.In order to build a goal-oriented dialogue agent,different approaches have been proposed in the last decade to train a system with an efficient dialogue policy.The Generative adversarial network(GAN)is consisting of two networks,generator and the discriminator.The major role of generator is the generation of fake data from real data and focuses on the optimization of the policy learning process.The discriminator network will receive the output from the trained policy learning and it will generate the reward.The reward output can be false or true depending on the values from 0 to 1.GAN is also applied in dialogue generation in several existing works,it helps in building the dialogue agents by selecting the optimal policy learning.The efficient dialogue policy learning benefits to improve the quality(fluency,diversity)of generated dialogues.Reinforcement learning(RL)algorithms are used to optimize the policies,because the sequence is discrete.In present study,we have proposed a new technique,the Cascade Generative Adversarial Network(Cas-GAN),which is combing the GCN and RL together for dialog generation.The Cas-GAN can model the relations between the dialogues(sentences)by using graph convolutional networks(GCN).The graph nodes are consisting of different high level and low-level nodes(that represents the vertices and edges of graph).Then we use the maximum loglikelihood(MLL)approach to train the parameters and choose the best nodes.RL is used to calculate rewards,since the values of the nodes are not continuous.The experimental results compared with the Seq-GAN network,have shown the efficiency of our proposed model Cas-GAN.