Design And Implementation Of A Telecom Customer Complaint Text Processing System Based On Semantic Understanding

As a complaint and feedback service provided by the telecommunication network to the public,with the improvement of the quality of digital life,people’s demand for communication and Internet access has become increasingly strong.The traditional telecommunication complaint service relies on manual time-consuming and laborious efforts to solve the complaint problems of each telecom customer,which has been unable to meet the current rapid and accurate feedback needs of the majority of telecom customers.The intelligent business processing system integrates advanced technologies such as natural language processing and knowledge management to better understand the complaint intention of telecom customers,return accurate natural language answers to users,or efficiently submit complaints to their processing departments for processing.However,most of the existing processing systems use large corpora as the source of knowledge,with redundant data and single dimension.At the same time,they are affected by synonyms and polysemy.In addition,the content of telecom customers’ complaints can not clearly reflect their demands,which will lead to wrong answers and incorrect system classification results.The context semantic fusion technology of bidirectional encoder representation from transformers(BERT)provides a good solution to the above problems.This thesis combines the BERT technology with the complaint handling system,uses the data set in the telecommunications field as the knowledge source of the complaint handling system,describes the information at the semantic and knowledge levels,introduces the semantic understanding and text classification technology,solves the problems of fuzzy telecom customers’ complaint intention and poor classification matching in the traditional complaint handling system,and realizes the improvement of the semantic understanding ability and text classification ability of the intelligent complaint handling system,Provide accurate and comprehensive complaint handling services.The main work of this thesis is as follows:(1)A semantic understanding model(BERT and Fusion Network for Semantic understanding,BFNet)based on BERT and Fusion network(Fusion Net)is proposed.This method is used to extract solutions in complaint handling.The semantic understanding method in this thesis is based on the output of the context semantic vector of the BERT model,and then the part of speech(POS)vector,named entity(NE)vector,word frequency vector and exact matching vector are spliced as the network input of the fusionnet interaction layer.After the problems and articles are treated with mutual attention and self attention for many times,the participation rights and of all problem word vectors are obtained according to the output layer,Calculate the probability of each word as "start" and "end",and finally extract the answer.(2)A text classification model BERT＿CNN based on BERT and convolutional neural network(CNN)is structured.This model provides effective support for the semantic understanding task in the system to screen more accurate candidate documents and realizes the classification function of complaint text.In this thesis,the text classification model obtains the vector representation of dependent word probability distribution and contact context information through BERT model,then takes this vector as the input of CNN network,passes through the convolution layer,pooling layer and full connection layer,and finally gives the classification according to the normalization result of softmax function.Experiments show that compared with Text CNN,Fast Text,BERT,BERT＿RNN and other models,this method can get better classification results.(3)A complaint handling system based on semantic understanding is implemented.Taking the telecom complaint field as the background,the FAQ database is constructed,and the complaint resolution strategy combining FAQ question answering and semantic understanding is designed to provide excellent solution suggestions for customer service.The system also provides a visual operation interface for text classification,FAQ database management and complaint handling platform.Finally,the practicability of the system is verified by the system function test,and the efficiency of the complaint handling platform combined with reading comprehension task and text classification task is verified by the performance test and blackand-white box test.