基于实体信息增强思维链的政策长文本摘要方法OA
A Long Policy Text Summarization Method Based on Entity Information Enhancement Chain of Thought
随着数字化治理的深入,政策公文的高效解读成为政务处理与社会治理的关键需求,文本摘要技术成为了快速提取核心内容的重要方法.然而政策公文篇幅冗长、要素复杂,针对传统人工摘要效率低下,通用大模型处理长文本政策时存在实体遗漏、逻辑断层等问题,该文提出一种基于实体信息增强思维链的政策公文长文本摘要生成框架.框架包含三大核心模块:首先通过实体抽取模块提取关键要素,引导大模型关注核心内容;然后通过关系建模模块构建语义关联,梳理公文内在逻辑;最后,通过提示生成模块生成最终的公文摘要,实现政策文本的关键信息精准提取和摘要生成.实验结果显示,该框架在 Rouge-1、Rouge-2、Rouge-L 和 BERTScore 指标上的 F1 值分别达到了 62.48%、33.02%、34.54%和75.34%,显著优于其他对比模型,相较于基础 Qwen2.5-7B 模型分别提升了4.38 百分点、7.1 百分点、10.9 百分点、5.37百分点.生成摘要示例表明,该框架所生成摘要在准确性、细节完整度和语言流畅性上均表征良好.
With the deepening of digital governance,the efficient interpretation of policy official documents has become the key requirement of government affairs and social governance,and text summarization technology has become an important method to quickly extract the core content.However,the length of policy official documents is long and the elements are complex.In view of the inefficiency of traditional manual summarization and the problems of entity omission and logical fault when dealing with long-text policies in general large model,a framework for generating long-text summaries of policy official documents based on entity information enhancement chain of thought is proposed,which consists of three core modules.Firstly,the key elements are extracted through entity ex-traction module to guide the large model to pay attention to the core content.Then,through the relationship modeling module,the semantic association is constructed,and the internal logic of official documents is sorted out.Finally,the final document summary is generated by the prompt generation module,and the key information of the policy text is accurately extracted and the summary is generated.The experimental results show that the F1 values of the framework on Rouge-1,Rouge-2,Rouge-L and BERTScore reach 62.48%,33.02%,34.54%and 75.34%,respectively,which are significantly better than that of other comparison models,and are improved by 4.38 percentage points,7.1 percentage points,10.9 percentage points,and 5.37 percentage points respectively compared with the basic Qwen2.5-7B model.The example of generating summaries shows that the summaries generated by the framework are well characterized in terms of accuracy,completeness of details and language fluency.
赵景欣;王志强;武永亮;董佳;唐松
河北省科学院 应用数学研究所,河北 石家庄 050081||河北省信息安全认证工程技术研究中心,河北 石家庄 050081河北省科学院 应用数学研究所,河北 石家庄 050081||河北省信息安全认证工程技术研究中心,河北 石家庄 050081石家庄铁道大学 信息科学与技术学院,河北 石家庄 050043河北省科学院 应用数学研究所,河北 石家庄 050081||河北省信息安全认证工程技术研究中心,河北 石家庄 050081河北省科学院 应用数学研究所,河北 石家庄 050081||河北省信息安全认证工程技术研究中心,河北 石家庄 050081
信息技术与安全科学
政策公文文本摘要大语言模型思维链实体抽取关系建模
policy official documentstext summarizationlarge language modelchain of thoughtentity extractionrelational modeling
《计算机技术与发展》 2026 (4)
121-129,9
河北省科学院科技计划项目(25606)河北省自然科学基金资助项目(F2024210005)
评论