Comparative Study of ML Techniques for Misinformation Detection in War Reporting

This study investigates machine learning techniques for detecting misinformation in news articles related to the Russian-Ukrainian war. The research compares traditional machine learning models using lexical features with fine-tuned transformers (BigBird and Longformer). The methodology involved preprocessing the corpus using a spaCy pipeline and extracting features such as character count, word count, and readability measures. Various models were trained, including SVM, Multinomial Naive-Bayes, KNN, Random Forest Classifier, and XGBoost, as well as the transformer models BigBird and Longformer. Results showed that the Random Forest Classifier achieved the highest accuracy (F1 = 0.800), outperforming more complex transformer models. The performance of the Random Forest Classifier and Multinomial Naive-Bayes models suggests that shallow text representations and basic text statistics can effectively capture linguistic patterns indicative of misinformation in this context. Limitations of the study include the small training dataset size and computational constraints in fine-tuning transformer models. This research contributes to the existing literature on developing robust tools for combating misinformation, particularly in the context of information warfare, and understanding topic-specific misinformation detection models.