Abstract
The increasing number of messenger fraud cases requires early and precise threat detection at unprecedented levels. The research examines modern NLP-based approaches which detect deceptive messages in messaging applications. The research examines various NLP approaches which analyze text data from different messaging platforms through text classification and tonality analysis and anomaly detection and thematic modeling techniques. The paper examines model learning data types together with text pre-processing methods and essential text features and evaluates traditional methods (e.g., Bag of Words, TF-IDF) and modern neural networks. The researchers encounter multiple obstacles while working which include the complex nature of processing informal language and the presence of noisy data and the need to frequently update models to detect new fraudulent schemes. The research focuses on messenger platform fraud detection because it addresses the unique challenges of real-time message streams and informal language and multimodal communication. The review evaluates technical and contextual aspects by presenting suitable models and architectures for dynamic short-form content and identifying technologies that deliver low-latency responses. The research aims to assess existing methods while identifying optimal approaches and proposing new directions to boost the accuracy and reliability of messenger fraud detection systems.