Still water runs deep: A Generative Summarization Framework Guide by Dual Significance

Current generative summarization models are proficient in producing coherent summaries. However, these summaries often suffer from factual inconsistencies, presenting challenges in ensuring comprehensive coverage of key information while excluding irrelevant and redundant details. Previous studies have utilized auxiliary information extraction to guide models towards generating reliable summaries. Nonetheless, the extraction auxiliary method lacks flexibility and struggles with cross-domain summary generation, potentially leading to issues such as information loss or redundancy. In this paper, we propose a flexible saliency guided approach called SGGDS, which comprises a Summary Generation Guide by Dual Significance, leveraging dual guidance from article saliency information and salient sentences. We conduct experiments on three publicly available summarization datasets from different domains. Experimental results show that the proposed method achieves better performance than the baseline model, which proves the advantages of the proposed method in cross-domain summary generation.