SSF-KW: Keyword-Guided Multi-Task Learning for Robust Extractive Summarization

The performance of extractive summarization models is often limited by their dependence on human references that may contain inaccuracies or subjective biases. Existing methods typically rely solely on sentence-level supervision, which lacks explicit grounding in the actual semantic content of the source document, thus limiting their robustness. We propose SSF-KW, a novel multi-task learning framework that enhances robustness by jointly optimizing keyword extraction and sentence selection. Our approach is designed to explicitly anchor salience decisions in the document’s intrinsic semantic structure, reducing reliance on potentially noisy labels. To this end, the model employs a shared BERT encoder to represent sentences, and identifies keywords through part-of-speech tagging, semantic similarity analysis, and fine-grained keyword signals with sentence-level representations via a transformer-based fusion module. The entire framework is optimized with a combined loss function that balances both tasks. Comprehensive evaluations on CNN/DailyMail, XSum, and WikiHow demonstrate that SSF-KW consistently outperforms baselines ROUGE-1 scores of 43.27, 25.43, and 30.03, respectively. Ablation studies confirm the contribution of each component, with the word-level module proving especially critical for capturing key concepts in procedural texts like WikiHow.