| PARIS 2024 ON X: COMPARING TWITTER-ROBERTA AND GPT-4 FOR REAL-TIME SENTIMENT AND ENGAGEMENT INSIGHTS |
| Paper ID : 1753-SPORTCONGRESS |
| Authors |
|
مهدی غفوری یزدی *1, حسین رفیعی زاده2, پوریا تجاسب3 1Assistant Professor (Department of Sport Management, Faculty of Sport Sciences and Health) Employment 2Master of Science in Network Science and Technology (Department of Data Science and Technology) 3PhD Candidate (Department of Industrial Engineering & Management Systems) Employment |
| Abstract |
| Introduction: This study examines public sentiment during the Paris 2024 Olympics on X (Twitter) using a dual-model pipeline that compares a Twitter-native transformer (Twitter-RoBERTa-base (Barbieri et al., 2020)) with a general-purpose LLM (GPT-4 (OpenAI, 2023)), aiming to extract fast, actionable insights for organizers, sponsors, and media. Methods: A corpus of 30,300 tweets was collected (26 Jul–14 Aug 2024), applied a three-stage cleaning pipeline, and manually annotated a 1,500-tweet gold standard (Fleiss’ κ=0.81) . zero- and few-shot GPT-4 (OpenAI, 2023) were evaluated against Twitter-RoBERTa using accuracy and macro-F1, then profiled sentiment over time, by country and by sport, and compared engagement (likes, retweets) across sentiment classes. Results: Twitter-RoBERTa outperformed GPT-4 (OpenAI, 2023) on short, slang-heavy posts (accuracy 0.81; macro-F1 0.80 vs 0.74 and 0.72). Few-shot prompts improved GPT-4 (OpenAI, 2023) by about six macro-F1 points. Overall sentiment was 45.6% positive, 45.5% neutral, and 8.9% negative. Although fewer, negative posts drew disproportionate engagement. Sentiment peaks aligned with medal moments (Aug 7–9), while the open-air opening ceremony produced mixed reactions. Country-level, the USA and India skewed positive and Russia more negative. By sport, tennis, swimming, and basketball were most positive, while boxing attracted the highest negative share. Tweets about women were majority positive (~57%). Conclusion: For real-time event monitoring, domain-adapted transformers deliver stronger classification on short social posts, while GPT-4 (OpenAI, 2023) adds value on longer, nuanced commentary. Practically, organizers should watch early negativity around ceremonies and logistics, brands can time activations to medal peaks and women’s sport storylines, and broadcasters can counter controversy with balanced context. The dual-model setup offers a compact, replicable template for mega-event sentiment tracking. |
| Keywords |
| Social media sentiment analysis; Olympics; RoBERTa; GPT-4; Engagement |
| Status: Abstract Accepted (Oral Presentation) |
