Fake News and Offensive Content Detection in Malayalam Using Machine Learning, Deep Learning and Transformer based method with XAI

Digital platforms have become one of the most powerful sources to spread the news and other information with much ease compared to traditional methods. However, this also made it easier to spread fake news and offensive contents too. A necessary strategy is required to detect fake news and offensive content for maintaining the integrity of information and fostering a safer online environment. The present study focuses on advancing the detection of fake news and offensive content in Malayalam language, integrating transformer models with explainability techniques for rendering transparency along with ensuring high accuracy. We used multilingual transformer models such as mBERT, IndicBERT, XLM-RoBERTa and MuRIL for classification tasks. Additionally comprehensive experiments with hyperparameter tuning are conducted on deep learning methods (CNN, LSTM, BiLSTM, GRU, BiGRU, CNN-LSTM, CNN-GRU) and traditional ML methods (SVM, LR, RF, DT, XGBoost, CATBoost, AdaBoost). The approach integrates TF-IDF and FastText embeddings for the representations, which is compared with BERT, IndicBERT, XLNeT, mBERT, XLM-RoBERTa and MuRIL. All models were evaluated using 10-fold cross-validation to ensure robustness and generalization of results. It is observed from the experiments that the XLM-RoBERTa model gave better result with F1 Score of 99.17% for the Fake news dataset and mBERT model gave better result of F1 Score 95.95% for offensive dataset. Furthermore, the present work utilizes explainable AI (XAI) approaches using LIME, Anchor, and Occlusion, to gain insights in the models decision making process. For the first time, an occlusion-based XAI method is proposed for fake news and offensive detection tasks. Also, the present work is first in its kind to use Anchor and Occlusion based XAI for fake news and offensive content detection in Malayalam. It was evident from the results that all three explainability methods occlusion, LIME, and Anchor effectively captured the key words relevant for detecting both fake news and offensive content, with a high degree of overlap in the identified terms across models. This consistency demonstrates their strong adaptability and suitability for handling the complex linguistic and code-mixed characteristics of Malayalam text.