Digital platforms are a major facilitator for the rapid spread of false information, which is one of the main threats to the information system. This narrative review presents state-of-the-art methodologies for the detection of fake news based on Natural Language Processing (NLP). The review surveys existing work and classifies existing methods into knowledge-based, feature-based, modality-based, hybrid and transformer-based models. It emphasizes the rising importance of multimodal frameworks that integrate text and visual data as well as pre-trained transformer models (like BERT and AraBERT) that lead to better contextual information. In these studies, the evaluation metrics used are accuracy, precision, recall, and F1-score. Nonetheless, key challenges remain in the form of dataset constraints, ethical reservations, computational inefficiencies, and cross-domain generalization. The results imply that any research going forward needs to focus on better scalable AI models, real-time detection capabilities, and multimodal misinformation monitoring. Addressing these difficulties, makes fake news detection systems based on NLP more reliable, adaptable, and ethical in different information systems.