Research And Application Of Fraud Phone Recognition Method Based On NLP

In recent years,domestic telecommunications network fraud activities have been rampant,and the incidence of cases has increased year by year,posing a great threat to social stability.Telephone fraud is a common form of telecommunications fraud,which refers to the criminal activities that fraudsters defraud money of victims through the phone.It occupies a relatively high proportion especially among the elderly.How to manage telephone fraud effectively is a major challenge for operators.At present,natural language processing（NLP）technology is in the stage of rapid development.NLP tasks can be divided into four categories including sequence labeling tasks,classification tasks,sentence relationship judgment,and generative tasks.Text classification is a classic problem in classification tasks,which is widely used in spam filtering,news classification,part-of-speech tagging,emotion analysis and other fields.Inspired by this,the thesis applies text classification tasks to fraud phone identification at the terminal.Firstly,we convert the voice on phone into text,then use text classification technology to classify phone texts automatically to determine whether it is a fraudulent call or not.The main works of this thesis are as follows:⑴ Building phone text datasets.There are no public telephone text datasets at home and abroad currently,because telephone calls involve personal privacy.The data collected in this thesis comes from various phone fraud cases that users have encountered on websites such as Baidu,Know,Micro Blog,and Little Red Book.Detailed introductions about datasets are in section 3.2.⑵ Two kinds of fraud phone classification models based on attention mechanism are proposed.The first is a classification model based on Bi LSTM-Attention.That is,attention mechanism is introduced after the bidirectional long short term memory（Bi LSTM）neural network model,which solves the problem that a single neural network cannot extract key features effectively.The second is the classification model that integrates the self-attention mechanism.That is,self-attention mechanism is introduced after the convolutional neural networks（CNN）model.Self-attention mechanism can extract the internal dependencies of the sequence and reduce the reliance on external information.Comparative experiments are carried out on phone text datasets to verify the performance of the two proposed models.The experimental results show that the proposed models have improved in terms of accuracy and F1 value and so on.⑶ A fraud phone fusion classification model based on improved Transformer and CNN is proposed.The output vector after word embedding model is input to Encoder of Transformer and Text CNN model respectively in order to give full play to the advantages of Transformer in extracting long-distance features and CNN in extracting local features.Then we splice with the two output vectors and input to the softmax layers to classify finally.The performance of the model is verified on the phone text datasets.