Design And Implementation Of A Multi-Task Human-Machine Dialogue System

Natural language understanding(NLU)is an important component of task-oriented dialogue system,including domain recognition,intent recognition and semantic labeling.As the leading task of natural language understanding,do-main recognition is the basis of the follow-up task,which affects the result and quality of dialogue,and is an important task of natural language understanding.In lots of real-world applications,domain recognition needs to identify the domain of the current sentence,and then the next intent recognition and slot filling task identify and analyze the sentence intention and slot value in this domain.In real dialogues,because of the large number of domains and no asso-ciated reference information,it is often more difficult to classify a single sen-tence for domain recognition.In addition,the data imbalance between the fields in the multi-domain dialogue task will also seriously affect the results of domain classification.Therefore,effectively identifying and classifying sentences has become a key concern in dialogue.Most of the traditional researches focus on the classification results of single sentences,and rarely use the effective infor-mation above the dialogue to assist the classification.To address the problem,this thesis has fully investigated related researches and does some works taking the real demands into consideration.More specif-ically,contents in this thesis include:A domain recognition model based on multi task learning is proposed.The model introduces the above information for the domain recognition of the cur-rent sentence,and constructs a domain recognizer based on the current sentence and the above information.With the change of the above information window,multiple different domain recognizers can be constructed,and these recogniz-ers can be jointly trained.The experimental results show that the model has a better domain recognition effect on the evaluation of dialog data sets.At the same time,the problem of unbalanced training data in different fields is an-alyzed and explored,and various methods of processing unbalanced data are realized,including oversampling method,cost sensitive learning method,etc.,and the optimal performance is explored.A multi service after-sale customer service human-machine dialogue sys-tem is realized.Through in-depth exchanges with enterprise product and cus-tomer service departments,the system requirements analysis was completed,and the system was developed on the basis of integrating the above-mentioned technology into the system outline design and detailed design.The functional evaluation of the system shows that the system can successfully identify the domain of the input sentence and carry out subsequent dialogue,and return the correct results.It has achieved better dialog performance and experience in en-terprise practical applications.