Design And Implementation Of Work Order Intelligent Processing System Based On BERT Model

With the expansion of the business field of the enterprise,the customer demands are becoming more and more complex,and the traditional work order system only relies on manual labor to handle the demands,which has problems of low processing quality and efficiency.The appeal information is mainly short text,which has the characteristics of sparse semantics,ambiguity and uncertainty.Traditional feature extraction methods cannot obtain the complete semantics of the text.In view of the above problems,combined with the BERT(Bidirectional Encoder Representation from Transformers)model to extract text features and fully consider the bidirectional context,which can better fit the characteristics of semantics,a work order dispatch model based on GSDMM(collapsed Gibbs Sampling algorithm for the Dirichlet multinomial Mixture Model)-BERT is designed to solve the problem of short text classification of work orders;a work order based on SBERT(Sentence-BERT)is designed.The single deduplication model solves the similarity problem of work order text repetition;a scheme recommendation model based on BM25(Best Matching 25)-SBERT is designed to solve the accuracy problem of fuzzy recommendation.The above method is applied to the work order intelligent processing system to improve the adaptability and processing capability of the system to external ambiguity and uncertain input,and to optimize the architecture of the work order system.Specific work includes:(1)Design a work order automatic dispatch model based on BERT and GSDMM.Aiming at the problem of feature dilution in the short text of the work order,the topic model GSDMM is used to expand the topic to generate the topic expansion vector,and then the BERT model is used to generate the global semantic vector,and the global semantic vector and the topic expansion vector are vector spliced to generate semantic and topic fusion vectors,and then go through the fully connected layer and the Softmax layer for classification.The assignment accuracy of the proposed method is verified by comparing the precision,recall and F1 value of the experiments.(2)Design a work order deduplication model based on SBERT.For the problem that the BERT model is inefficient for similar sentence pairs,the SBERT model that combines the Siamese network and the BERT model is used.Input the data into two identical BERT models respectively,obtain the semantic vector through the pooling layer,then calculate the cosine similarity,intercept the work orders with the similarity higher than the threshold,and optimize the work order creation process.(3)Design a historical solution recommendation model based on BM25 and SBERT.Aiming at the problem that the traditional recommendation does not integrate the title and content features,BM25 is used to calculate the weighted similarity of the title and solution content of the solution document,and several historical solutions with the highest similarity are screened out,and then the SBERT model is used for semantic similarity.Re-sorting and selecting TopK items to achieve the purpose of program recommendation and provide diversified references for users to deal with problems.(4)Design and implementation of work order intelligent processing system.Optimize the structure of the work order system by introducing natural language processing technology,develop the system based on Spring,Mybatis and Vue,realize the work order cycle management module,including the creation,dispatch and reception,processing and evaluation of work orders and other functions;realize the basic functions of the system Module,including user information management,solution recommendation and background management and other functional modules.Test each module of the system,and the test results show that the system functions meet the actual needs.