Stance Detection in Persian Social Media Using a Deep Learning Approach

Given the rapid growth of social media usage and the increasing volume of textual content, identifying users' stances on various topics in these spaces has become increasingly important. However, the Persian language, due to its unique characteristics such as structural complexity and a wide vocabulary, still presents significant challenges for language processing models. In this study, a deep learning-based approach is proposed for stance detection in Persian tweets. The focus of this research is on the social media platform X, recognized as one of the most important and visited platforms worldwide. Initially, tweet data from the X platform (Twitter) was collected, covering various cultural, economic, social, and political topics. For data labeling, a hybrid machine-assisted and manual approach was used; in the first step, the large language model AYA238b was employed for initial labeling, followed by manual corrections by experts to improve labeling accuracy. The proposed model utilizes TookaBERT-base, an advanced model for Persian, for fine-tuning and classifying tweets into four categories: positive, opposing-ethical, opposing, and neutral. The evaluation results showed that the fine-tuned model outperformed traditional methods in detecting the stance of Persian tweets. This approach demonstrated better accuracy in identifying various stances in texts, particularly in distinguishing between opposing and neutral opinions, achieving an F1 score of 89%.

Stance Detection in Persian Social Media Using a Deep Learning Approach Keywords:

Stance Detection , User Profiling , Social Network Analysis , Persian Text Classification