2025/08/27 by Shekari Shahrak, Azam, Mikaeilvand, Nasser, Mirabedini, Seyed Javad +2
paper · doi:10.57647/j.fomj.2025.0602.09
With the growth of information technology, the use of cyberspace is increasing daily, and individuals are constantly exchanging information through various platforms. Among them, social networks have gained increased attention if users spend hours each day on business or sharing information. As social media usage increases, the issue of spam becomes more prevalent. Companies and institutions worldwide spend billions of dollars combating spam; thus, researchers have focused on the fight against spam. In this context, we are currently witnessing significant advancements in natural language processing aided by large language models, which have demonstrated remarkable performance due to their high potential for providing innovative solutions. Thus, this study employs the RoBERTa transformer, which is one of the most advanced transformers in the text classification domain. TensorFlow was used to tune its hyperparameters to enhance the performance of this model. Furthermore, spam filtering is typically performed based on the message content or its non-content features. In this study, features are extracted based on both message aspects. Keyword extraction and vector representation of the text are performed after preprocessing the RoBERTa model. The resulting vector from tweet tokenization is fed into a Convolutional Neural Network for classification. Subsequently, the tweet's result is combined with non-content features and sent as input to a fuzzy system for classification. A Twitter dataset of approximately 12,000 lines was used for training and testing. The results obtained from the proposed method indicate that fine-tuning the transformer and classification with a neural network can increase the accuracy of spam detection to 99.82%.