| Abstract | The rapid dissemination of information on social media platforms such as Twitter has heightened the risk of misinformation, making efficient fact-checking more important than ever. While methods for detecting false information automatically exist, they come with several limitations, most often still requiring human experts for clarification. To support manual fact-checking, check-worthiness can be performed to identify, which posts require manual review. Traditional approaches to check-worthiness estimation have primarily focused on analyzing the text of tweets in isolation, overlooking the rich context provided by user interactions and discussions. This work investigates how incorporating user comments and interaction metrics can enhance check-worthiness detection for tweets. Using an augmented CheckThat! Lab 2022 dataset, we employ Transformer models for textual analysis and Graph Neural Networks (GNNs) for modeling comment structures. Our results show that including comments yields a moderate, consistent boost in detection accuracy, with BERT-based models achieving the best F1-scores. Interaction metrics alone provide little value. While GNNs do not outperform BERT in this setting, they show promise for more complex conversational structures. We highlight challenges in data collection and model interpretability, and emphasize the importance of qualitative discussion signals for efficient, automated fact-checking. Future work should explore deeper graph structures, richer context integration, and improved explainability. |
|---|
