Research On The Technologies Of Dialog System Based On Multitask Learning And Knowledge Graph

Currently,the scale of data sets increases quickly,and people hope to obtain necessary information anytime and anywhere in a natural and friendly manner.Thus,a dialog system,an effective human-machine interaction system,has been developed.Dialog systems can be classified into two types: task-oriented and non-task-oriented dialog systems according to their applications.In this paper,we focused on task-oriented dialog systems.As internal movement mechanism of a task-oriented dialog system,research on the task-oriented dialog model has increased owing to the development of artificial intelligence and natural language processing.Although the performance of the dialog model has dramatically improved,many problems have not yet been solved.For example,as a critical component of natural language understanding,intent detection and slot filling modules are significantly related.Existing research does not make full use of incidence relations and shared resources between the two modules.Additionally,the text that a user inputs into the task-oriented dialog model is limited,which leads to the shortage of textual features.The role of a dialog response generation module transforms the interim output into natural language style that the user can understand.Some existing models may output sentences that contain grammatical errors or semantic ambiguities,and some sentences are perfunctory and useless.To address these problems,this paper proposes a new dialog model based on multitask learning and knowledge graph.The proposed model aims to improve the performance of natural language understanding and dialog response generation module.Hence,the main work and innovations of this paper are as follows:(1)We made full use of the incidence relations and shared resources between intent detection and slot filling modules and introduced knowledge to assist the completion of these modules.Furthermore,we propose a joint model for intent detection and slot filling based on multitask learning with knowledge base.Our model employs external knowledge and high-quality relationship information between the intents and slots in three parts.First,this model obtains sharedparameters and features between the intent detection and slot filling modules based on Long Short Term Memory and Convolutional Neural Networks.Second,we introduced a knowledge base into the model to improve its performance.Finally,we built a weighted loss function to optimize the joint model.(2)The presented response generation model depends on inadequate information to generate a response,and the model is inclined toward generate general reactions that lack information.For this problem,we propose a dialog response generation model based on knowledge graph.This model simulates prior knowledge of an individual brain through the knowledge graph and implements the model using double attention mechanism and knowledge bias probability.To solve the problem of model output that generate repeated information,we introduced coverage mechanisms.(3)This model combines multitask learning and knowledge graph to implement a task-oriented dialog model.Thus,we constructed a dialog model on the Cam Rest676 dataset to execute a specific restaurant ordering task.The result of this experiment verified the availability and effectiveness of the proposed methods.