Knowledge acquisition for dialogue agents using reinforcement learning on graph representations