شیرین دهقانی محمدابادی

With the proliferation of social networks such as Facebook, Instagram, and Twitter, a vast amount of information is generated and disseminated daily, which may contain suspicious and false content. This content is created for purposes such as attracting audiences, influencing people's beliefs and decisions, increasing revenue through clicks, and influencing significant events, such as political elections. Identifying this fake news through manual and heuristic methods is often time-consuming, costly, and exhausting. Consequently, the necessity of detecting fake news has become a fundamental requirement to help people navigate the confusion caused by this dubious and false content. Identifying fake news articles by understanding what other news sources report on the same subject can be the first valuable step, known as stance detection. In this research, initially, significant topics discussed on Twitter in Iran during the period from November ۱۵, ۲۰۲۲, to February ۱۳, ۲۰۲۳, have been identified, considering the literature of the subject. Subsequently, for each topic, a labeled dataset is presented. In the next step, utilizing a combination of machine learning and the XLM-t transformer, which is a BERT-based language model used as the embedding vector for words, the Stance Detection problem in Persian tweets has been addressed. The results are compared with the fasttext model, and the improvements are explained. Given that one of the main challenges in this domain is the lack of high-quality Persian datasets, data augmentation techniques have been employed to enhance the results. After implementing and comparing the results for each topic, XLM-t-based architectures demonstrated better performance. The best F۱ scores are ۷۸.۱% for government appointments, ۷۱.۷% for the Barjam (JCPOA), ۷۶.۱% for the Preservation Plan, ۸۸.۱% for the Coronavirus vaccine, ۷۳.۶% for the stock market, ۷۱.۴% for military service, and ۸۲.۲% for compulsory hijab.