一种用于方面情感四元组预测的最优意见树生成方法  

An optimal opinion tree generation method for aspect sentiment quad prediction

作　　者：李博男 杨振宇[2] LI Bonan;YANG Zhenyu(China Radio and Television Shandong Network Company Limited,Jinan 250101,China;Qilu University of Technology(Shandong Academy of Sciences),Jinan 250013,China)

机构地区：[1]中国广电山东网络有限公司,山东济南250101 [2]齐鲁工业大学(山东省科学院),山东济南250013

出　　处：《山东理工大学学报（自然科学版）》2025年第1期21-27,共7页Journal of Shandong University of Technology:Natural Science Edition

摘　　要：使用对比学习来增强树结构表示是将树结构应用于方面级情感分析(ABSA)的主流方法,但该方法本质上依赖于数据增强,而对原始文本应用数据增强可能会删除与下游预测相关的语义信息且难以获得最优结构。为解决上述问题,提出一种用于方面情感四元组预测(ASQP)的结构熵约束的最优意见树生成方法,该方法通过文本编码器和直接生成正样本的结构编码器实现。结构编码器以意见树嵌入为输入,利用结构熵最小化原理提取意见树中固有的本质信息,并通过表示学习将该信息注入到文本表示中。在2个常见数据集上进行的实验验证了该方法的优越性。Enhancing tree structure representations using contrast learning is a mainstream approach to applying tree structures to aspect based sentiment analysis(ABSA),but the method inherently relies on data enhancement,which may remove semantic information relevant to downstream predictions and makes it difficult to obtain optimal structures.To address the above issues,an optimal opinion tree generation method with structural entropy constraints for aspect sentiment quad prediction(ASQP)is proposed,which is implemented by a text encoder and a structural encoder that directly generates positive samples.The structural encoder takes the opinion tree embeddings as input,extracts the intrinsic essential information inherent in the opinion tree using the structural entropy minimization principle,and injects this information into the textual representation through representation learning.Experiments on two common datasets validate the superiority of the method.

关 键 词：方面情感四元组预测 意见树 结构熵 贪心算法 方面级情感分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]