Binary text classification plays a crucial role in fields like incident reporting and safety analysis, where accurately differentiating between accidental and non-accidental events is essential. This research assesses both supervised and semi-supervised learning techniques on a Bengali news article dataset for binary classification. Supervised learning, which depends exclusively on labeled data, typically exhibits strong classification performance but can be hindered by limited annotated samples. Semi-supervised learning uses both labeled and unlabeled data to enhance model results, particularly when there are fewer labeled examples. We seek to determine each method’s relative efficacy in incident classification by examining its benefits and drawbacks. Semi-supervised approaches show significant benefits in situations when labeled data is scarce, emphasizing the significance of choosing a suitable learning approach based on the data at hand. We see that our supervised model which is the combination of Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) and Generative Adversarial Networks (GANs) achieve the highest accuracy 95% that outperforms the previous results.