Analyzing public sentiment in Iranian presidential elections on Twitter using large language models

This study investigates the dynamics of public sentiment surrounding the 2024 Iranian presidential election by analyzing Persian-language tweets. We introduce the IranElectionTweet dataset, a comprehensive collection of 111,386 election-related tweets enriched with textual content, user metadata, and engagement indicators. Due to the sensitive political context and privacy considerations, the full dataset is not publicly released; instead, we provide a manually annotated subset of 500 tweets (Tweet IDs and dates) for benchmarking, along with reconstruction instructions and analysis code. To conduct sentiment analysis, we fine-tuned GPT-4 on a publicly available Persian sentiment dataset, adapting it to the linguistic and cultural nuances of Persian political discourse. In parallel, we evaluated three cutting-edge large language models, Claude Sonnet 3.7, DeepSeek-V3, and Grok-4, using a few-shot learning framework due to the unavailability of fine-tuning access at the time of experimentation. All models were benchmarked on a manually annotated subset of 500 tweets. DeepSeek-V3 attained the highest weighted F1-score and overall accuracy, indicating stronger performance on the majority classes and was selected as the primary model for sentiment classification. The final sentiment analysis was applied to the full dataset, capturing hourly and daily variations in sentiment and candidate mentions throughout the election period. The results reveal distinct patterns in public opinion corresponding to key political events, offering valuable insights into the real-time evolution of electoral sentiment on social media. This research highlights the effectiveness of advanced multilingual language models in low-resource settings and contributes to the broader understanding of political behavior in digital environments.