Detecting Propaganda Techniques in News Articles

The binary propaganda classification task in the news articles is described in the paper. The classification experiment was conducted for word and text fragments. The classification was done using the data of the SemEval 2020 competition that is in open access. The dataset consists of the manually labeled news articles. The classes in the presented data are strongly unbalanced. The problem was solved with classical methods of machine learning and more complex methods of deep learning. Logistic Regression and the Support Vector Classification models were trained with linguistic and syntactic features were constructed in the work. The neural network was also developed and trained. The architecture of the neural network includes the BERT tokenizer, the natural language sequence model BERT, and a layer of bidirectional long short-term memory (BiLSTM). The paper discusses methods of preprocessing and augmentation of the text data for balancing the presented classes, such as replacing some words with their synonyms, deleting and inserting tokens in a sentence, switching words. The work also describes the loss function used while training neural networks with unbalanced data. The paper presents the results of binary classification for models of different types. The results show that a more complex model architecture improves metrics in the token classification problem. The metric values for the fragment classification problem are comparable for simple machine learning models and a neural network based on BERT and BiLSTM. In the future, the work done can be the basis for a deeper study of the methods for classifying the propaganda. The propaganda detection problem is relevant and in demand in the present time. Algorithms of propaganda detection, existing at the moment, have not yet shown excellent results for the classification of words and phrases in a text.