aaivu/aaivu-dialogue-policy-optimization

Dialogue policy optimization in low resource setting # Dialogue policy optimization in low resource setting ![project] ![research] ## Summary The dialogue policy optimization in task oriented conversational agents, employed in low resource setting, is an open research project. We have developed a novel approach for dialogue policy optimization using Reinforcement Learning. The methodology is based on Self-play and a novel sampling technique that prioritizes failed dialogues over successful ones. ## Description A conversational agent or a typical dialogue system is a software which interprets and responds to statements, made by users in real time, in free-form natural language, with the purpose of retrieving desired information relevant to the task to which the agent is oriented. The dialogue manager of such a system uses a dialogue policy to decide what the agent should say at each point in the dialogue. Designing an efficient dialogue policy, that could essentially be used in real world dialogue system applications, encompasses numerous challenges. Accuracy of user responses being highly variable, user responses being relatively unconstrained, the user alone being aware, frequently, of his/her goal are just a few challenges to start off with. Thus, dialogue policy optimization is regarded as an open-research problem. Currently many state of the art methods, hinged on Reinforcement Learning (RL), are being deployed to address this problem. RL is an area of machine learning, where the agent is rewarded or punished by the user according to the state it returns for the given input. The agent continuously learns by taking suitable actions to maximize reward in a particular situation. When viewing dialogue management as a RL task, the reward function - which the agent tries to maximize - would be fulfilling the user's goal as efficiently as possible. However, RL based methods tend to overfit in a low resource setting due to sample deficiency. To tackle this issue, we introduce a novel probability based …