Design And Implementation Of Cognitive Service Bot Intelligent Interaction System

With the development of human-machine dialogue technology,more and more companies have launched their cognitive service products,such as VPA,smart speakers,and shopping guide robots.However,in these practical applications,most of the machines passively respond to user expressions,lacking knowledge of user preferences and involuntary dialogue consciousness.Therefore,how to make the machine actively and naturally guide the dialogue direction to recommend services to meet the needs of users is very important.In order to solve the problem of lack of user cognition,inability to guide the dialogue and error generation of replies,the machine needs to plan the dialogue target sequence and generate appropriate replies based on the integration of external knowledge.Therefore,this article mainly studies the two tasks of dialogue goal planning and automatic reply generation.In this paper,dialogue goal planning is divided into two sub-problems: goal sequence planning and dialogue goal detection.The dialogue goal sequence is generated according to the user portrait and personal knowledge base before the dialogue starts;the change of the current dialogue goal is detected in real time while the dialogue is in progress and re-planned accordingly Target sequence,so that the machine can prepare service resources to meet user needs in advance.The selection of the current dialogue target in the human-machine dialogue depends on whether the target is completed or not.Therefore,this paper divides the dialogue target detection into two potentially related subtasks: target completion estimation and current target prediction.The joint learning method is used to model one of the two subtasks.Between the links.To this end,this paper proposes the dialogue goal planning framework DGPF.For sub-problem 1,DGPF uses three different strategies,preference-driven priority,knowledge-driven priority,and multilayer perceptron sequence to generate the target sequence.For sub-problem 2,DGPF uses the Chinese pre-training model ERNIE and the graph embedding model Node2 Vec to encode the user expression text and the previous round of targets into vectors respectively,and input the encoded vectors into the two sub-networks of Completion Network and Prediction Network to predict whether the dialogue target is Complete and judge the current dialogue goal.This paper proposes a method of reply generation based on ontology concept.This templated method is suitable for question answering scenarios oriented to vertical domain knowledge graphs.First,answer retrieval and reasoning are performed,and then the answers are filled into the defined verbal template and returned to the user as the response text.This paper proposes two end-to-end response generation models: Pointer Net model based on pointer network and KGSelect Net model based on knowledge selection.These two Seq2Seq-based models are suitable for goal-driven open-domain dialogue scenarios.Pointer Net uses three GRU-based encoders to encode user expressions,dialogue goals,and related knowledge.Among them,the dialogue goal is the output of DGPF,and the related knowledge is the SPO triples that belong to the same field as the dialogue goal after being screened by DGPF.Pointer Net is based on the pointer network to copy the text in the triplet when generating the reply.KGSelect Net is an improvement of Pointer Net.It uses the attention mechanism to calculate the score of each piece of knowledge in the user's expression,and finds the most appropriate knowledge for response generation.