融合知识和语义信息的双编码器自动摘要模型  

Dual-Encoder Automatic Summarization Model Incorporating Knowledge and Semantic Information

作　　者：贾莉 马廷淮 桑晨扬[1] 潘倩 JIA Li;MATinghuai;SANG Chenyang;PAN Qian(School of Computer Science,Nanjing University of Information Science&Technology,Nanjing 210044,China;School of Computer Engineering,Jiangsu Ocean University,Lianyungang,Jiangsu 222005,China;School of Software,Nanjing University of Information Science&Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学计算机学院、网络空间安全学院,南京210044 [2]江苏海洋大学计算机工程学院,江苏连云港222005 [3]南京信息工程大学软件学院,南京210044

出　　处：《计算机工程与应用》2025年第7期213-221,共9页Computer Engineering and Applications

基　　金：国家自然科学基金(62372243,62102187);江苏省自然科学基金(BK20210639)。

摘　　要：为了解决自动文本摘要任务存在的文本语义信息不能充分编码、生成的摘要语义冗余、原始语义信息丢失等语义问题,提出了一种融合知识和文本语义信息的双编码器自动摘要模型(dual-encoder automatic summarization model incorporating knowledge and semantic information,KSDASum)。该方法采用双编码器对原文语义信息进行充分编码,文本编码器获取全文的语义信息,图结构编码器维护全文上下文结构信息。解码器部分采用基于Transformer结构和指针网络,更好地捕捉文本和结构信息进行交互,并利用指针网络的优势提高生成摘要的准确性。同时,训练过程中采用强化学习中自我批判的策略梯度优化模型能力。该方法在CNN/Daily Mail和XSum公开数据集上与GSUM生成式摘要方法相比,在评价指标上均获得最优的结果,证明了所提模型能够有效地利用知识和语义信息,提升了生成文本摘要的能力。To solve the semantic problems of automatic text summarization tasks,such as text semantic information cannot be adequately encoded,the generated summaries are semantically redundant,and the original semantic information is lost,it proposes a dual encoder automatic summarization model that incorporates knowledge and text semantic information.The method employs a dual encoder to fully encode the semantic information of the original text,where the text encoder obtains the semantic information of the text,and the graph structure encoder maintains the contextual structure information of the text.The decoder part adopts Transformer-based structure and pointer network to better capture text and structural information for interaction and take advantage of the pointer network to improve the accuracy of generated summaries.Meanwhile,the training process employs a gradient of self-critical strategies in reinforcement learning to optimize the model capabilities.Compared with the GSUM generative summary methods on CNN/Daily Mail and XSum public datasets,the proposed method achieves optimal results on evaluation indexes,which proves that the proposed model can effectively utilize knowledge and semantic information.This method improves the ability to generate text summaries.

关 键 词：知识图谱编码器 图注意力机制 指针网络 增强训练 自动摘要 

分 类 号：TP393[自动化与计算机技术—计算机应用技术] G250[自动化与计算机技术—计算机科学与技术]