图书情报工作

图书情报工作 >

2025 , Vol. 69 >Issue 15: 75 - 87

DOI: https://doi.org/10.13266/j.issn.0252-3116.2025.15.007

研究论文

融合大语言模型结构化信息和主题语法图的评论摘要生成方法

  • 张乐 ,
  • 陈岩松 ,
  • 张雷瀚
展开
  • 1 网络文化与数字传播北京市重点实验室 北京 100192;
    2 北京信息科技大学计算机学院 北京 100192;
    3 北京邮电大学经济管理学院 北京 100876
张乐，副教授，博士；陈岩松，硕士研究生；张雷瀚，副教授，博士，通信作者，E-mail: zhangleihan@gmail.com。

收稿日期: 2024-07-05

  修回日期: 2024-11-26

  网络出版日期: 2025-07-21

基金资助

本文系北京市教育委员会科学研究计划项目“融合舆情口碑的首都景区知识图谱构建方法研究”(项目编号：KM202311232001)和网络文化与数字传播北京市重点实验室开放课题“多模态网络语言的演化模式与影响力预测研究”(项目编号：ICDD202201)研究成果之一。

收起

A Review Summarization Generation Method Integrating Structured Information of Large Language Models and Topic Syntax Graph

  • Zhang Le ,
  • Chen Yansong ,
  • Zhang Leihan
Expand
  • 1 Beijing Key Laboratory of Internet Culture and Digital Dissemination Research, Beijing 100192;
    2 School of Computing, Beijing University of Information Technology, Beijing 100192;
    3 School of Economics and Management of Beijing University of Posts and Telecommunications, Beijing 100876
Zhang Le, associate professor, PhD; Chen Yansong, master candidate; Zhang Leihan, associate professor, PhD, corresponding author, E-mail: zhangleihan@gmail.com.

Received date: 2024-07-05

  Revised date: 2024-11-26

  Online published: 2025-07-21

Supported by

This work is supported by the general project of the Science and Technology Plan of the Beijing Municipal Education Commission, titled “Research on the Construction Method of the Knowledge Graph of the Capital Scenic Spots Integrating Public Opinion and Word-of-Mouth” (Grant No. KM202311232001), and the project of Beijing Key Laboratory of Network Culture and Digital Communication, titled “Research on the Evolution Pattern and Influence Prediction of Multimodal Network Language” (Grant No. ICDD202201).

Fold

摘要

[目的/意义] 针对现有评论摘要模型在生成过程中存在生成内容混乱以及属性词与情感词不对应的问题，提出一种融合大语言模型结构化信息和主题语法图的评论摘要生成方法。[方法/过程] 运用大语言模型GPT4.0抽取评论的结构化信息，利用句法依赖分析器提取句子的依存信息，通过语法图卷积神经网络获取语法特征向量，利用TextRank抽取主题信息，将主题信息和语法信息融合，通过交叉注意力机制分别获得包含主题语法信息引导语义信息的向量表示和包含结构化信息引导语义的向量表示，形成双通道信息，并获取融合注意力，替换指针生成网络的单通道注意力，最终生成评论摘要。[结论/结果] 相对于已有方法，提出的模型在ROUGE值评价指标上的效果相较于其他先进基线模型均有提升。生成的摘要更能符合评论文本原句的内容及顺序，能够缓解事实性错误的问题。

关键词： 大语言模型; 评论摘要; 注意力机制; 语法分析; 指针网络

本文引用格式

张乐 , 陈岩松 , 张雷瀚 . 融合大语言模型结构化信息和主题语法图的评论摘要生成方法[J]. 图书情报工作, 2025 , 69(15) : 75 -87 . DOI: 10.13266/j.issn.0252-3116.2025.15.007

Abstract

[Purpose/Significance] To address the issues of disorganized generated content and attribute words do not correspond to sentiment words in existing review summary models, a review summarization generation method is proposed that integrates the structured information of large language models and the topic syntax graph. [Method/Process] The large language models was used to extract the structured information of reviews, and the syntactic dependency parser was used to extract the dependency information. The grammatical feature vector was obtained by syntactic graph convolutional neural network, and the topic information was extracted by TextRank, then the topic information and grammatical information were fused. Through the cross-attention mechanism, a vector representation containing topic syntax information to guide semantic information and a vector representation containing structured information to guide semantic information were obtained respectively to form dual-channel information. By fusing the attention module, replacing the pointer to generate the single-channel attention of the network. Finally, the review summarizations were generated through a pointer generation network. [Result/Conclusion] Compared with the existing methods, the model proposed in this paper has improved performance in terms of the ROUGE evaluation metrics, surpassing other advanced baseline models. The generated summaries better adhere to the content and order of the original review sentences, alleviating the issue of factual errors.

Key words： large language models; review summarization; attention mechanism; syntax analysis; pointer network

参考文献

[1] 谷莹, 李贺, 祝琳琳. 融合主题聚类和语义图模型的产品评论自动摘要方法研究[J]. 图书情报工作, 2022, 66(13): 118-126. (GU Y, LI H, ZHU L L. Research on automatic summarization method of product reviews based on topic clustering and semantic graph model[J]. Library and information service, 2022, 66(13): 118-126.)
[2] 蒋亚芳, 严馨, 徐广义, 等. 融合多信息句子图模型的多文档摘要抽取[J]. 计算机工程与科学, 2020, 42(3): 535-542. (JIANG Y F, YAN X, XU G Y, et al. Multi-document summarization extraction based on multi-information sentence graph model[J]. Computer engineering & science, 2020, 42(3): 535-542.)
[3] 苏放, 王晓宇, 张治. 基于注意力机制的评论摘要生成[J]. 北京邮电大学学报, 2018, 41(3): 7-13. (SU F, WANG X Y, ZHANG Z. Review summarization generation based on attention mechanism[J]. Journal of Beijing University of Posts and Telecommunications, 2018, 41(3): 7-13.)
[4] TANG R X, CHUAN Y N, HU X, et al. The science of detecting LLM-generated text[J]. Communications of the ACM, 2024, 67(4): 50-59.
[5] 魏盼丽, 王红斌.融合关键信息与专家网络的生成式文本摘要[J]. 吉林大学学报(理学版), 2024, 62(4): 951-959. (WEI P L, WANG H B. Fusing key information and expert network for abstractive text summarization[J]. Journal of Jilin University(science edition), 2024, 62(4): 951-959.)
[6] 高永兵, 钟振华, 王宇, 等.基于混合方法的中文微博自动摘要技术研究[J]. 计算机工程与科学, 2016, 38(6): 1257-1261. (GAO Y B, ZHONG Z H, WANG Y, et al. Automatic summarization of Chinese micro-blog based on a hybrid method[J]. Computer engineering & science, 2016, 38(6): 1257-1261.)
[7] FATTAH M A. A hybrid machine learning model for multi-document summarization [J]. Applied intelligence, 2014, 40(4): 592-600.
[8] ERKAN G, RADEV D R. Lexrank: graph-based lexical centrality as salience in text summarization[J]. Journal of artificial intelligence research, 2004, 22(1): 457-479.
[9] MIHALCEA R, TARAU P. Textrank: bringing order into text[C]//Proceedings of the 2004 conference on empirical methods in natural language processing. Barcelona: ACL, 2004: 404-411.
[10] 林莉媛, 王中卿, 李寿山, 等.基于PageRank的中文多文档文本情感摘要[J]. 中文信息学报, 2014, 28(2): 85-90. (LIN L Y, WANG Z Q, LI S S, et al. Chinese multi-document opinion summarization via PageRank[J]. Journal of Chinese information processing, 2014, 28(2): 85-90.)
[11] GULCEHRE C, AHN S, NALLAPATI R, et al. Pointing the unknown words[C]//Proceedings of the 54th annual meeting of the Association for Computational Linguistics. Berlin: ACL, 2016: 140-149.
[12] SUTSKEVER I, VINYALS O, LE Q V. Sequence to sequence learning with neural networks[C]//Proceedings of 2014 annual conference on neural information processing systems. Montreal: Neural Information Processing Systems Foundation, 2014: 3104-3112.
[13] SEE A, LIU P J, MANNING C D. Get to the point: summarization with pointer-generator networks[C]//Proceedings of the 55th annual meeting of the Association for Computational Linguistics. Vancouver: ACL, 2017: 1073-1083.
[14] 余传明, 朱星宇, 龚雨田, 等. 基于序列到序列模型的抽象式中文文本摘要研究[J]. 图书情报工作, 2019, 63(11): 108-117. (YU C M, ZHU X Y, GONG Y T, et al. Research of abstractive Chinese text summarization based on Seq2seq model[J]. Library and information service, 2019, 63(11): 108-117.)
[15] TAN J, WAN X, XIAO J. Abstractive document summarization with a graph-based attentional neural model[C]//Proceedings of the 55th annual meeting of the Association for Computational Linguistics. Vancouver: ACL, 2017: 1171-1181.
[16] XIE N T, LI S J, REN H L, et al. Abstractive summarization improved by WordNet-based extractive sentences[C]//Proceedings of the 7th CCF international conference on natural language processing and Chinese computing. Cham: Springer, 2018: 404-415.
[17] 高玮军, 朱婧, 赵华洋, 等. 基于TRF-IM模型的个性化酒店评论摘要生成[J]. 计算机工程与应用, 2023, 59(2): 135-142. (GAO W J, ZHU J, ZHAO H Y, et al. Personalized product review summary generation based on TRF-IM model[J]. Computer engineering and applications, 2023, 59(2): 135-142.)
[18] 郭继峰, 费禹潇, 孙文博, 等. 一种融合主题的PGN-GAN文本摘要模型[J]. 小型微型计算机系统, 2023, 44(1): 199-203. (GUO J F, FEI Y X, SUN W B, et al. Text summarization model based on PGN-GAN with topic[J]. Journal of Chinese computer systems, 2023, 44(1): 199-203.)
[19] 章成志, 童甜甜, 周清清. 基于细粒度评论挖掘的书评自动摘要研究[J]. 情报学报, 2021, 40(2): 163-172. (ZHANG C Z, TONG T T, ZHOU Q Q. Automatic summarization of book reviews based on fine-grained review mining[J]. Journal of the China Society for Scientific and Technical Information, 2021, 40(2): 163-172.)
[20] 任淑霞, 张靖, 赵宗现, 等. 结合主题和位置信息的两阶段文本摘要模型[J]. 智能计算机与应用, 2023, 13(9): 158-163. (REN S X, ZHANG J, ZHAO Z X, et al. A two-stage text summarization model combining topic and position information[J]. Intelligent computer and applications, 2023, 13(9): 158-163.)
[21] 汪凯, 刘明童, 张玉洁, 等. 基于图神经网络的汉语依存分析和语义组合计算联合模型[J]. 中文信息学报, 2022, 36(7): 24-32. (WANG K, LIU M T, ZHANG Y J, et al. Joint learning Chinese dependency parsing and semantic composition based on graph neural network[J]. Journal of Chinese information processing, 2022, 36(7): 24-32.)
[22] LI R, CHEN H, FENG F, et al. Dual graph convolutional networks for aspect-based sentiment analysis[C]//Proceedings of the 59th annual meeting of the Association for Computational Linguistics and the 11th international joint conference on natural language processing. ACL, 2021: 6319-6329.
[23] PARSA K Z, MOJTABA V, MEHDI R Z, et al. Sorted LLAMA: unlocking the potential of intermediate layers of large language models for dynamic inference[C]//Proceedings of the Association for Computational Linguistics. Malta: ACL, 2024: 2129-2145.
[24] LIU Q J, CHEN N, TETSUYA S, et al. ONCE: boosting content-based recommendation with both open and closed-source large language mod-els[C]//Proceedings of the 17th ACM international conference on Web search and data mining. New York: WSDM, 2024, 10: 452-461.
[25] SHI Y W, REN P, WANG J, et al. Leveraging GPT-4 for food effect summarization to enhance product-specific guidance development via iterative prompting[J]. National Center for Biotechnology Information, 2023, 148: 104533.
[26] DOZAT T, MANNING C D. Deep biaffine attention for neural dependency parsing[C]//Proceedings of the 5th international conference on learning representations. Toulon: OpenReview.net, 2017.
[27] 王震宇, 朱学芳, 杨睿. 基于多模态大语言模型的关系抽取研究[J/OL]. 数据分析与知识发现, [2025-01-21]. http://kns.cnki.net/kcms/detail/10.1478.G2.20240117.1110.024.html. (WANG Z Y, ZHU X F, YANG R. Research on relation extraction based on multimodal large language model[J/OL]. Data analysis and knowledge discovery, [2025-01-21]. http://kns.cnki.net/kcms/detail/10.1478.G2.20240117.1110.024.html.)
[28] VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]// Proceedings of the 31st international conference on neural information processing systems. California: Curran, 2017: 6000-6010.
[29] 李岱峰, 林凯欣, 李栩婷. 基于提示学习与T5 PEGASUS的图书宣传自动摘要生成器[J]. 数据分析与知识发现, 2023, 7(3): 121-130. (LI D F, LIN K X, LI X T. Automated book summary generator based on prompt learning and T5 PEGASUS[J]. Data analysis and knowledge discovery, 2023, 7(3): 5998-6008.)
[30] 李红莲, 陈浩天, 张乐, 等. 融合情感-主题双通道信息的评论摘要生成模型[J]. 数据分析与知识发现, 2024, 8(6): 30-43. (LI H L, CHEN H T, ZHANG L, et al. A review summary generation model with emotion-topic information[J]. Data analysis and knowledge discovery, 2024, 8(6): 30-43.)
[31] LIU Y, LAPATA M. Text summarization with pretrained encoders[C]//Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing. Hong Kong: ACL, 2019: 3728-3738.
[32] 李健智, 王红玲, 王中卿. 基于图卷积网络的专利摘要自动生成研究[J]. 计算机科学, 2022, 49(S1): 172-177. (LI J Z, WANG H L, WANG Z Q. Automatic generation of patent summarization based on graph convolution network[J]. Computer science, 2022, 49(S1): 172-177.)
Options
文章导航

/

〈 〉