10th U6+ Consortium of African Universities International Conference

Social network platforms have evolved and their usage is now spanning across numerous applications. This popularity is occasioned by the recent increase in number of internet users who utilize these networking platforms for mutual interactions. In spite of these successes, the surging social interactions have attracted adversaries, who now target these networks with spam contents. These unwanted or unsolicited contents and messages have the potential of irritating users and compromising the overall quality of these social networking platforms. Due to the devastating effects that spam has on the credibility of the platform and users, the recent past has witnessed concerted efforts in the development of spam detection techniques. These spam detection mechanisms are based on diverse learning models such as deep learning, conventional machine learning, and large language models. These models have exhibited varied success capabilities in spam detection. Nevertheless, majority of these spam detection models still face numerous challenges that must be resolved. For instance, some of these models have challenges in the extraction of relevant features especially when dealing with short social media comments. In addition, current deep learning and large language models face serious challenges concerning contextual understanding of messages as well as the need for huge computational resources. In this paper, we propose a Lexical Semantics-Aware recurrent neural network (LSA-RNN) that facilitates optimal lexical understanding of spam messages in short-text tasks such as SMS or email. Extensive performance evaluations are performed using metrics such as accuracy, precision, recall, F1 score, area under curve receiver operating characteristics (AUC-ROC) based on three different datasets. The results indicate that the proposed spam detection model attained the most ideal performance compared with other base models as well as the state-of-the-art models in this domain.