基于知识基因增强的BERT科技文献自动综述研究  

Research on Automatic Summary of BERT Scientific and Technological Literature Based on Knowledge Gene Enhancement

作　　者：赵梦梦 白如江[1] 张玉洁 刘明月 邢莹 Zhao Mengmeng;Bai Rujiang;Zhang Yujie;Liu Mingyue;Xing Ying(Institute of Information Management,Shandong University of Technology,Zibo 255000;School of Information Management,Nanjing University,Nanjing 210023)

机构地区：[1]山东理工大学信息管理研究院,淄博255000 [2]南京大学信息管理学院,南京210023

出　　处：《图书情报工作》2022年第23期125-136,共12页Library and Information Service

基　　金：国家社会科学基金项目"多源数据融合驱动的智慧情报感知研究"(项目编号:21BTQ071)研究成果之一。

摘　　要：[目的/意义]揭示文献核心创新点,自动生成多篇文献综述,帮助科研人员快速掌握文献核心内容,提高科研效率。[方法/过程]提出一种基于知识基因增强的BERT科技文献自动综述研究方法,分为3个步骤,首先综合考虑主题相似度、论文发表时长和被引次数,提出一种核心文献推荐指数,选取文献综述候选文献;然后对文献综述候选文献中的代表科技文献核心观点的知识基因进行抽取;最后提出一种基于知识基因注意力增强的BERT科技文献自动综述模型,将知识基因融入到注意力机制中,判断语句显著度并进行排序抽取,以获取更多的语义信息。[结果/结论]经过多组实验,与单纯的BERT相比,本文模型在ROUGE-1分别提高了14.28%;ROUGE-2分别提高了12.13%;ROUGE-L分别提高了17.69%。在ROUGE-1与ROUGE-2测评中基于知识基因增强的BERT科技文献自动综述模型效果均优于TextRank模型。基于知识基因注意力增强的BERT科技文献自动综述能够深入文本内容,挖掘文献核心内容,生成简明扼要的文献综述。[Purpose/Significance]To reveal the core innovation points of literature,automatically generate multiple literature reviews,help researchers quickly grasp the core content of literature,and improve the efficiency of scientific research.[Method/Process]This paper proposed an automatic review research method for BERT scientific and technological literature based on knowledge gene attention enhancement,which was divided into three steps.Firstly,a core literature recommendation index was proposed to select literature review candidated by comprehensively considering topic similarity,publication time and citation times.Then the knowledge genes representing the core viewpoints of scientific and technological literature were extracted from literature review candidates.Finally,an automatic BERT scientific literature review model based on knowledge gene attention enhancement was proposed.Knowledge genes were integrated into the attention mechanism to judge the significance of sentences and sort and extract them to obtain more semantic information.[Result/Conclusion]After several sets of experiments,compared with BERT alone,Rouge-1 of the proposed model is improved by 14.28%,respectively.Rouge-2 is increased by 12.13%;Rouge-l is increased by 17.69%respectively.In the evaluation of Rouge-1 and Rouge-2,the automatic review model of BERT scientific and technological literature based on knowledge gene enhancement is better than TextRank model.Automatic review of BERT scientific and technological literature based on enhanced knowledge gene attention can dig into the text content,explore the core content of literature,and generate concise literature review.

关 键 词：知识基因 文献综述 自动摘要 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]