A Fact-Aware Cascaded Framework for Dynamic-Granularity Timeline Summarization

In this paper, we propose a Fact-Aware Cascaded Framework for Dynamic-Granularity Timeline Summarization (DTELS), designed to address the limitations of LLM-generated timelines in social media contexts, which often lack satisfactory performance in Informativeness, Granular Consistency, and Factuality. By incorporating a fact-checking mechanism and employing a decoupled, modular task design, the framework sequentially executes data preprocessing, timeline initialization, atomic event extraction, node deduplication, dynamic granularity control, and low-quality node pruning. In the CCKS2025 Event Context Generation for Social Media Evaluation, this framework demonstrated balanced performance across all metrics, achieving the highest overall score and ranking first, validating its effectiveness and practical utility.