Rumor Detection in Bangla: A Comparative Study of ML and DL Models Integrated with PMI-SO

The swift growth of online platforms has transformed the processes of information dissemination and absorption globally. The swift distribution of information facilitates the proliferation of misinformation and rumors, which present considerable social, political, and economic risks. Rumors can provoke public panic, damage reputations, influence electoral results, and potentially incite violence, making their detection crucial. This paper introduces an approach for detecting rumors on internet platforms by employing ML and DL techniques. A dataset containing 1,070 non-rumor instances and 1,863 rumor instances was collected using Selenium automation and was subsequently preprocessed for experimentation. The preprocessing steps involved removing tags, URLs, punctuation, and other extraneous elements, as well as normalizing text and tokenizing to improve consistency and optimize model performance. Pointwise Mutual Information with Semantic Orientation (PMI-SO) was incorporated with Keras embedding to derive significant semantic and contextual features. This integrated strategy enables the model to distinguish between positive and negative sentiments and identify the relationships between words in the text, thereby improving categorisation accuracy. Better features made it easier to train and evaluate a variety of ML and DL models. The machine learning model that performed best was XGBoost, achieving an accuracy rate of 84.72 %. A combination of DL models, including CNN and LSTM networks, showed better performance with an accuracy of 91.82%.