2019/07/05 by Chinnadhurai Sankar, Sujith Ravi, Sankar, Chinnadhurai +1
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Natural Language Processing Techniques #Speech and dialogue systems #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1907.02848
openalex publication_date 2019/07/05 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
Open domain dialog systems face the challenge of being repetitive and\nproducing generic responses. In this paper, we demonstrate that by conditioning\nthe response generation on interpretable discrete dialog attributes and\ncomposed attributes, it helps improve the model perplexity and results in\ndiverse and interesting non-redundant responses. We propose to formulate the\ndialog attribute prediction as a reinforcement learning (RL) problem and use\npolicy gradients methods to optimize utterance generation using long-term\nrewards. Unlike existing RL approaches which formulate the token prediction as\na policy, our method reduces the complexity of the policy optimization by\nlimiting the action space to dialog attributes, thereby making the policy\noptimization more practical and sample efficient. We demonstrate this with\nexperimental and human evaluations.\n