Sentence Extraction Framework with High Relevance and Divergence for Document Summarization

Document summarization task is to condense documents while retain their key information, playing a key role in processing large-scale textual data. However, large language models (LLMs) focus on modeling frequently occurring information but often overlook less frequent yet crucial details, leading to potential information loss. To address this limitation, we propose a novel solution, i.e., Sentence Extraction Framework with High Relevance and Divergence for Document Summarization (SERD). In the framework, a two-channel document attention module is designed for ensuring high relevance of the extracted sentences, and a coarse-fine granularity synergy MMR module is designed for ensuring high divergence. So SERD can capture low-frequency yet crucial information by balancing relevance and divergence. The extracted sentences of high quality are subsequently fed into generative models to generate summaries. SERD is a universal plug-in component for LLMs, overcoming their limitations. Extensive experiments on two public datasets verify that our solution outperforms the SOTA baselines and significantly enhances the summary generation capabilities of LLMs.