Research On Classification Of Complaint Texts Based On Improved SVM

Customer complaints about the company's products and services are increasing,and timely feedback to the complaining users is the key to the company's reputation.The automatic classification of complaint text can help enterprises improve the efficiency of handling complaints,improve user satisfaction and avoid losses of customers.At present,the complaint classification process in enterprises is still dominated by manual identification,which is not only inefficient and costly,but also greatly affected by people's experience and judgment.How to accurately and timely classify customer complaints has become an urgent problem to be solved.The emergence of natural language processing technology makes automatic text classification possible.Text modeling methods and classification methods develop rapidly and are gradually applied in real life.Therefore,the classification of complaint text has important theoretical and practical value.Firstly this thesis introduces the relevant content of text preprocessing,text presentation method,classification algorithm and ensemble learning,and then analyzes the generation,influence and existing characteristics of complaint text.Based on the characteristic of complaint texts,BTM is able to extend the complaint short text based on internal corpus,while Doc2 vec can obtain corpus information that cannot be obtained by the Topic Model.The paper chooses the method of combining the BTM and Doc2 vec to present the complaint text.The feature vector of the complaint texts contains word co-occurrence information,syntax and semantic information,it also reduces the dimension of text feature in this way,the model can be repeatedly updated,the iterative stronger,with the increase of new data set,the presentation of the text becomes more and more accurate.In terms of classifier,a kernel function combining linear kernel and polynomial kernel is proposed to improve SVM.And using soft interval to avoid overfitting problem.Finally,Bagging method is introduced to conduct ensemble classification to further improve the accuracy of the classification model.In the practical stage,real enterprise complaint texts were selected to carry out kernel function comparison experiment,ensemble classification comparison experiment and parameter influence experiment,and the algorithm proposed in this paper was verified.The empirical results showed the effectiveness of the method.