Fake news has become one of the global risks that endanger people's production and life. With the development of electronic information and artificial intelligence, the technology of automatic detection of fake news has also appeared. Inspired by these issues, the main goal of this article is to develop a multi-labelling multi-class detector fake news test algorithm. We use the dataset from the MediaEval 2022 fake news detection challenge, which consists of 1913 texts from Twitter. We propose a Bert-based multi-label multi-category fake news detection model to solve the text classification task in the MediaEval 2022 fake news detection challenge. We first convert the 9 categories of tweet tags into 27 categories, and then input them into the model for classification. The Accuracy Score of our model is 0.495, the Micro F1 Score is 0.926, and the Macro F1 Score is 0.538. 1 Currently news contains multiple modal data, such as text, images, and videos. We can detect the authenticity of news and classify it only through text. Ma Jing and others applied deep learning technology to fake news detection for the first time [4]. This method inputs each sentence of the news into RNN [5], and uses the hidden layer vector of the cyclic neural network to represent the news information. Then input the hidden layer information into the classifier to get the classification result. FENG et al. Using convolutional neural network modeling articles for the first time [6], the various posts of the news event were mapped to vector space, and then each POST vector was spliced into a matrix. Embedded vector into the classifier to get the classification results. Ma Jing and others applied Multi-Task Thought to fake News Test [7]. This method is to combine fake news testing tasks and position classification tasks into a multi-tasking model, and use RNN as Backbone to train two tasks. Ma Jing and others applied the idea of fighting training into the test of fake news [8]. This method uses the generator to transform the rumor into non -rumor, expand the training data, and then enter the generated news and original news to the judgment Fake news testing in the device. VAIBHAV and others transform fake news detection issues into graph classification problems [9]. This method models news articles as a sentence with a sentence with a sentence, with the similarity between sentences, and using GCN [10] fusion diagram. The information between the middle nodes obtains the node embedding vector, embeds the figure pool by the node vector, and enters into the classifier to obtain the detection result.
The MediaEval Fake News Detection Challenge has been running for several years [
We use the BERT pre-trained model to obtain text features, and use the fully connected layer as a classifier. After we get the text features, this news representation is then passed through a fully connected neural network for classification. passes its results through a feed-forward network, and then hands it off to the next encoder. A detailed description of the textual feature extractor is shown in Figure 1. 3.2
In order to facilitate data processing, we convert the labels according to the following rules, "0,0,1" means label 1, "0,1,0" means label 2, and "1,0,0" means label 3. The original 9-category labels are converted to 27 categories. The text representation and labels are obtained from this, and the fully connected layer is used as a classifier for classification.
We divide the training set and test set according to the ratio of 8:2. The labels of the data set are converted according to the rules before being input into the model, and the labels are restored after the model output is obtained.
We use the classic pre- trained BERT model and full-connected layer mode for multi-label text classification, and no preprocessing operation of the data set text is not performed. In order to meet the input and output of the model, we divided the label from 9 into 27 categories, and the final Accuracy Score was 0.495, and the F1 score was 0.538. The final result failed to reach the expected goal. After merging 27 tags into 9 categories, each type of score is large, the highest ones reach 0.7 or even higher, and the lowest is less than 0.3. Therefore, we believe that the classification of a single model and multi-labelling multiclass into more labels is inappropriate for this challenge. Especially decomposing the number of labels into more quantities may cause the performance of the model to decrease.
Thanks to the organizers of the MediaEval2022, especially to those organizers for FakeNews Detection. This work was supported in part by the Scientific Research Leader Studio of Jinan (Grant No. 2021GXRC081), and in part by the Joint Project for Smart Computing of Shandong Natural Science Foundation (Grant No. ZR2021LZH010, ZR2020LZH015, and ZR2022LZH012).