PromptSED: An evolving topic-enhanced prompting framework for incremental social event detection

The recent success of Transformer-based pre-trained language models (PLMs) offers a novel perspective on social event detection (SED), which seeks to identify clusters of social messages corresponding to real-world events. However, employing PLMs for SED tasks necessitates careful consideration of the inherent characteristics of social data, including the short text format with limited information, as well as the continuous updates as new events emerge over time. To address these challenges, this work introduces PromptSED, an evolving topic-enhanced prompt learning framework for SED. PromptSED is a novel paradigm that dynamically tracks topics in social streams and selectively injects them into PLMs to facilitate incremental SED. The framework is distinguished by three key innovative designs: (1) Selective Topic Injection: A mechanism that converts evolving topic-related information into prompts for PLMs; (2) Noise-Tolerant Optimization: A strategy that enhances the framework's resilience to noise while improving its ability to differentiate between events; (3) Training-Free Topic Tracking: A method that enables event detection in an incremental social stream without requiring additional training or manual labeling. Experiments conducted on two publicly available event datasets demonstrate that PromptSED achieves significant performance improvements for SED. Further analysis validates the contributions of each design component, emphasizing the framework's overall effectiveness. Additionally, inspired by the remarkable performance of recent decoder-only large language models across diverse tasks, we evaluate their applicability to SED, providing a comparative analysis against the proposed framework.