Fraudulent News Headline Detection With Attention Mechanism

Online social media platforms and instant messaging services are ideal places for news dissemination,while most of them are facing a serious problem,the spread of fraudulent news headlines.For fraudulent news headline detection,the previous method is basically laborious manual review,but under the extreme condition that total number of news headlines goes as high as a million,the manual review on news headline truthfulness may not be practically feasible.In addition,the processing performances of regular Machine Learning models in large-scale text data are unsatisfactory.However,at present,Long Short-Term Memory(LSTM)neural network and Attention Mechanism in Deep Learning field have powerful processing capabilities for news headline text data.On the basis of LSTM neural network with its own Attention Mechanism and the specialized Multi-Head Attention Mechanism implemented by Google Transformer,the Deep Learning classification models proposed in this thesis can efficiently fit the contextual information of news headlines,so as to detect fraudulent news headlines more quickly and accurately.The main research contents and contributions of this thesis are as follows:1.Fraudulent news headlines and text data preprocessing: Introduce fraudulent news headlines,analyze their security threats,and compare their typical examples.Perform text data preprocessing on the original news headline data,and convert them into the data form which is suitable for Deep Learning classification models.2.LSTM model with its own Attention Mechanism for fraudulent news headline detection: Implement word embedding layer,employ LSTM neural network with its own Attention Mechanism,build brand-new Deep Learning classification models.Compared with regular Machine Learning classification models,optimize the classification performance of news headlines,and detect fraudulent news headlines more quickly and accurately(accuracy: 85.6551%).3.Mini-Transformer model for fraudulent news headline detection: Implement word embedding layer,based on the Transformer machine translation model launched by Google Inc.,avoid any recurrent computational unit or convolution computation,make the best of the Multi-Head Attention Mechanism and fully connected layer,and build a state-of-the-art Mini-Transformer text classification model in order to further improve the detection performance of fraudulent news headlines(accuracy: 86.5692%).