面向文本摘要的基于分段权重与语义覆盖率的主动学习方法OA

In recent years,with the increasing demand for high-quality annotated data for text summarization tasks,active learn-ing,as a method to reduce annotation costs,has gradually become a research hotspot.Especially in scenarios where labeling re-sources are limited,active learning can significantly enhance annotation efficiency through intelligent sample selection.How-ever,current active learning methods primarily rely on either diversity sampling or uncertainty sampling alone,resulting in insuf-ficiently stable performance in summarization tasks.To address these issues,this paper proposes an active learning method for text summarization based on segment weighting and semantic coverage.Specifically,segment weights and semantic coverage are introduced into diversity sampling,adjusting the sampling weights based on the number of iterations.Additionally,the degree of sample discretization in the semantic space is calculated using information entropy,combined with uncertainty sampling to im-prove the effectiveness of active learning.Experimental results on three summarization models and three benchmark datasets indi-cate that the proposed method outperforms existing active learning methods in the ROUGE-1,ROUGE-2,and ROUGE-L met-rics.Compared to the current optimal method,DUAL,the proposed method improves the ROUGE-1,ROUGE-2,and ROUGE-L metrics by 0.86 percentage points,0.42 percentage points,and 0.23 percentage points,respectively,on the PEGASUSLARGE model and the BillSum dataset.These results validate the superiority of the proposed method in text summarization tasks,provid-ing a feasible solution for enhancing model performance under conditions of limited annotated data.