Research On Guided Dialog Technology Based On Recommendation

With the vigorous development of deep learning technology,the research and application of dialogue systems have grown rapidly,such as Siri,Cortana,Google Assistant,etc.Correspondingly,these products have also become emotional partners for elders and children,online customer service,virtual assistants in mobile terminals,and smart guides for online shopping and consumption.Different from the virtual assistants,the application of smart shopping guides has just received attention from the industry and academia.Smart shopping guides such as Amazon,Ebay,Taobao,JD,and Yelp usually play the role of shop assistants or consultants,recommending products for mobile terminal users,shops,restaurants,services,etc.make their daily lives more convenient.Therefore,these products usually need to introduce a recommendation system to complete the act of recommendation,so the task of recommendation-oriented dialogue is gradually extended.Different from task-oriented dialogue,first,the user's goal is usually ambiguous,and it is expected to be determined through recommendation-oriented dialogue.Second,in a recommendation-oriented dialogue,the recommendation object can be the final result,or an intermediate action to clarify or confirm the user's intention,thereby allowing multiple recommendations in the dialogue session.Most importantly,unlike task-oriented dialogue systems that assume that users fully understand the slots and values of tasks,users in recommendationoriented dialogue systems are usually not familiar with the slots and values of recommended objects.Therefore,there may be an information gap between the sentence generated by the recommendation-oriented dialogue system and the user's understanding,which makes the system uncertain about the user's intention.In order to study how to better interact with users and bridge the information gap between the dialogue system and users,we will study three systems: dialogue recommendation based on simple reinforcement learning,guided dialogue based on simple reinforcement learning,and,guided dialogue based on hierarchical reinforcement learning.Conversational recommendation based on simple reinforcement learning will ask questions based on the attribute space of the product,and use the result of the question as the current state for action selection,and the strategy system will decide to ask questions or make recommendations.It should be noted that once the system selects the recommended action,no matter what kind of feedback the user makes,the session will end.The guided dialogue based on simple reinforcement learning integrates the recommendation system and the dialogue system through a strategy system,and adds recommended actions to the dialogue decision.When the user does not understand the slot terminology or does not know his preference,the system can choose to recommend to the user based on the information obtained in the current dialogue to guide the user to clarify his preference,improve the success rate of the dialogue,and reduce the turn of the dialogue.The guided dialogue based on hierarchical reinforcement learning.,hierarchical reinforcement learning is mainly used to solve the problem of dimensional disasters caused by too large state or action space,and good results have been achieved in the experiment.