Based On The Research And Realization Of The Online Shopping Mall Service Dialogue System

Dialogue system is a research hotspot in the field of science and technology and industry at present,especially the non-contact interaction during the epidemic has been widely concerned and recognized by the society,which has a positive performance in the fight against the epidemic.Usually,the dialog system can not only help users to complete various tasks through multiple rounds of dialog interaction.At present,the dialogue system can be divided into task type,question and answer type and chat type,but now the dialogue system is often integrated development.Therefore,the research content of this paper is based on natural language understanding and three types of dialogue management,hoping to improve the previous system functions through three ways of dialogue.This paper realizes the online mall service dialogue system.It mainly realizes the natural language comprehension module of joint training and the task type,retrieval type and generative three-type dialogue management module.The main research contents include:(1)This paper found that the data set expansion did not improve the performance of the model,but instead reduces the model evaluation index dialogue corpus as we get from the Internet all kinds of "noise",such as exist together meaningless "ah","well," replied,for retrieval model and generate interference model training existence can not be ignored.We sift through a lot of conversations.It provides the quality assurance for the experiments in the following chapters.(2)In this paper,we use two datasets,ATIS and snips,and introduce the capsule neural network model to carry out relevant experiments on the joint training of intention recognition and slot recognition.We compare the performance of each model in different datasets,and also pay attention to the performance under low resource conditions..(3)In the chapter of dialog management,the retrieval model and the single-round dialog generation model are mainly studied.Retrieval we used the bm25 "coarse screen" to reduce the workload of the later models,and then experimented with a variety of machine learning and deep learning-related models.The generated model compares AME with the traditional Seq2 seq model,and an attention mechanism is added to reduce the number of generic responses.(4)Finally,this system adopts the capsule of the neural network of natural language understanding,use methods and retrieval model based on rules,generation model to complete the dialogue management module,management module in the conversation to determine whether the default task type intentions,if they are entering the task-oriented dialogue management,we are useing the template type module generated reply.Otherwise,leave it to other models.If the reliability of the retrieval model is low,it is left to the generated model to complete the conversation.