Answerer in Questioner's Mind: Information Theoretic Approach to\n Goal-Oriented Visual Dialog

Abstract

Goal-oriented dialog has been given attention due to its numerous\napplications in artificial intelligence. Goal-oriented dialogue tasks occur\nwhen a questioner asks an action-oriented question and an answerer responds\nwith the intent of letting the questioner know a correct action to take. To ask\nthe adequate question, deep learning and reinforcement learning have been\nrecently applied. However, these approaches struggle to find a competent\nrecurrent neural questioner, owing to the complexity of learning a series of\nsentences. Motivated by theory of mind, we propose "Answerer in Questioner's\nMind" (AQM), a novel information theoretic algorithm for goal-oriented dialog.\nWith AQM, a questioner asks and infers based on an approximated probabilistic\nmodel of the answerer. The questioner figures out the answerer's intention via\nselecting a plausible question by explicitly calculating the information gain\nof the candidate intentions and possible answers to each question. We test our\nframework on two goal-oriented visual dialog tasks: "MNIST Counting Dialog" and\n"GuessWhat?!". In our experiments, AQM outperforms comparative algorithms by a\nlarge margin.\n