Research On Task-oriented Dialogue Based On Knowledge Fusion And Its Application In Instrument Field

In recent years,the progress of deep learning and natural language processing technology has promoted the development of man-machine dialogue.All kinds of dialogue products have entered people’s lives,among which Task-oriented dialogue products that help people complete all kinds of tasks have broad application space and are very popular.In Task-oriented dialogue,the completion of the target task depends on the system’s use of background knowledge.Most of the existing research on Task-oriented dialogue is based on a single form of knowledge,and the mixed multiple types of knowledge is still less.Among them,the use of unstructured knowledge includes sentence similarity calculation and single topic decoding.The latter has higher accuracy in knowledge retrieval.However,in a round of dialogue based on unstructured knowledge,single topic decoding can not meet the multiple needs put forward by users at the same time.In addition,due to people’s conversational habits,the omission of entity information exists in natural language,which will hinder the use of knowledge in the model.The tasks in the field of instruments are complex and diverse.The integration of intelligent technology can improve the efficiency of task completion.At present,Task-oriented dialogue is rarely used in this field.Based on the above problems,this thesis studies and puts forward the corresponding solutions.Firstly,aiming at the problem that the existing unstructured knowledge utilization methods can not solve the multiple needs of users in one round of dialogue,this thesis proposes a multi-topic decoding scheme.By designing topic slots and decoding the preset topic slots in each round of dialogue,multiple user’s needs are mapped to multiple topics.At the model level,the model is constructed by using GRU and attention mechanism,and multiple topics are decoded through the interaction between context and topic slot.Finally,the model uses a specific retrieval method to retrieve the unstructured knowledge corresponding to multiple topics and participate in the response generation.Secondly,aiming at the problem that the omission of entity information in the user input discourse leads to the deviation of the positioning of entity name in the model,which eventually leads to the model retrieving wrong knowledge and generating wrong response,this thesis proposes an ellipsis recovery solution,constructs the ellipsis recovery module by using GRU,replication network and pre-training model BERT,and integrates it into the Task-oriented dialogue model.Through training,we learn the correlation characteristics between the dialogue with Ellipsis and the complete dialogue form,and have the ability to recover the omitted entity name.Finally,this thesis analyzes the task requirements of electronic component selection in the process of circuit design in the instrument field,and finds that the task is similar to the completion idea of the target task in the current task-based dialogue system.The utilization of Task-based dialogue system can improve work efficiency.This thesis constructs the knowledge base and data set related to components.Based on the current research work,constructs a prototype dialogue system for completing the dialogue task of component selection,which can recommend components to users through multiple rounds of dialogue.