基于Transformer和TextRank的微博观点摘要方法  

Microblog Opinion Summarization Method Based on Transformer and TextRank

作　　者：孙旭 沈彬 严馨[1,2] 张金鹏[3,4] 徐广义 SUN Xu;SHEN Bin;YAN Xin;ZHANG Jinpeng;XU Guangyi(School of Information Engineering and Automation,Kunming University of Science and Technology,Kunming Yunnan 650500,China;Key Laboratory of Artificial Intelligence in Yunnan Province(Kunming University of Science and Technology),Kunming Yunnan 650500,China;School of Computer Science and Engineering,Yunnan University,Kunming Yunnan 650091,China;School of Information,Yunnan University of Finance and Economics,Kunming Yunnan 650221,China;Yunnan Nantian Electronic Information Industry Co.,Ltd.,Kunming Yunnan 650040,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500 [2]云南省人工智能重点实验室(昆明理工大学),云南昆明650500 [3]云南大学信息学院,云南昆明650091 [4]云南财经大学信息学院,云南昆明650221 [5]云南南天电子信息产业股份有限公司,云南昆明650040

出　　处：《广西师范大学学报（自然科学版）》2023年第4期96-108,共13页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金(U21B2027,61972186);云南省重点研发计划(202103AA080015);云南省基础研究计划(202001AS070014);云南省科技人才与平台计划(202105AC160018)。

摘　　要：针对已有研究没有考虑微博文本之间情感关联的问题,本文提出基于Transformer和TextRank的微博观点摘要方法。首先通过Transformer中的编码器和量化空间部分对文本的字向量进行编码和量化;然后根据量化结果实现语义聚类来划分微博文本集的观点类别,并选取重要的类别进行摘要抽取;之后将情感特征向量和微博文本的特征向量进行拼接;接着在每个类别中使用融入情感特征的TextRank算法,将抽取出的权重最高的微博文本作为摘要文本;最后将所有类别下最具有代表性的摘要文本相结合,得到最终的微博观点摘要。实验结果表明:在加入情感极性影响因子后,相比于基线方法,本文方法的各项ROUGE值均有明显地提升,Rouge-1、Rouge-2和Rouge-SU4的F-measure值最高达到0.4937、0.2555、0.2706,证明本文方法对于微博观点摘要抽取任务是有效的。The association of sentiment among microblog texts has not been considered by previous research.A microblog opinion summarization method based on Transformer and TextRank is proposed in this paper.Firstly,the word vectors of the texts are encoded and quantified by encoder and quantization space of Transformer.Then according to the quantization results,the opinion categories of microblog textset are divided by semanteme clustering,and the important categories are selected for summary extraction.Then the sentiment feature vector and the microblog text feature vector are concatenated.Then TextRank algorithm with sentiment features is used in every category,and the microblog text with the highest weight is extracted as the summary text.Finally,the most representative summary texts in all categories are combined to obtain the final microblog opinion summarizations.The experimental results show that,after adding the sentiment polarity influence factor,the ROUGE values of the proposed method has significantly improved compared with the baseline method.The maximum F-measure values of Rouge-1,Rouge-2 and Rouge-SU4 can top out at 0.4937,0.2555,0.2706 respectively.It proves that the proposed method is effective for the task of extracting microblog opinion summarizations.

关 键 词：情感特征 观点摘要 语义聚类 摘要抽取 TRANSFORMER TextRank 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]