Dynamic optimization of min-df in the GreedSum algorithm for enhanced extractive summarization

This study aims to improve extractive text summarization by dynamically optimizing the min-df (minimum document frequency) parameter in the GreedSum algorithm. To achieve this, three methods are proposed for dynamic tuning: a geometric approach, a percentile-based threshold, and a clustering-based adaptation strategy. These methods were applied to a large-scale dataset of 17,038 scientific articles from arXiv and PubMed. The experiments demonstrate a 2% improvement in ROUGE-1 F-measure over fixed min-df settings, achieving a peak ROUGE-1 F1-score of 45%. This performance surpasses several established extractive and hybrid baselines. Our contributions include a comparative evaluation of dynamic tuning strategies and the demonstration of their effectiveness for adaptive summarization. These findings confirm the potential of dynamic min-df optimization for generating accurate and efficient summaries, with future research focused on deep learning integration and real-time, multi-modal summarization.