【作者】 张欣；

【导师】 周德宇； 夏雪；

【作者基本信息】 东南大学 ， 软件工程（专业学位）， 2024， 硕士

【摘要】 随着在线电子商务平台的迅猛发展和用户交互的日益频繁,用户评论数据呈现爆炸式增长。面向电商平台用户评论的情感分析作为洞察用户反馈、优化商品服务的关键手段,在电子商务领域的重要性日益凸显。然而,传统的面向电商平台用户评论的情感分析往往基于静态假设,即假定用户的偏好在不同时间段保持不变。这种假设与现实中用户评论文本表述及其情感随时间变化的流式特性相悖。更进一步的,电商平台的用户评论普遍存在稀疏性问题,在按时间段划分的流式场景中,这种稀疏性更加突出。因此,本文针对过去方法“基于静态假设”和“流式用户评论稀疏性”的两个问题,聚焦于面向电商平台流式用户评论的情感分析,提出新的解决方案,并集成到系统中。本文涉及的面向电商平台流式用户评论的情感分析旨在分析过去时间段内用户对多个商品的多样化评论,并尝试预测用户在未来某一时间段的评论情感倾向。这一框架突破了单条评论的局限,将情感分析拓展至整个时间跨度内的评论,从而更全面地揭示用户评论文本表述及其情感随时间变化的动态趋势。本文的主要工作如下:(1)针对传统的静态建模方法无法捕捉用户动态偏好的问题,本文设计了基于双通道动态图神经网络的面向流式用户评论的情感分析方法。通过设立针对用户和商品两个维度的专用通道,该方法实现了对流式用户评论及评分动态变化特性的多维度捕捉。具体而言,它首先利用双通道文本编码器从评论文本中提取当前的局部和全局上下文信息,分别作为用户与商品的相关信息。然后,将用户和商品作为动态图中的节点,评论作为连接节点的边,通过深入分析用户和商品之间的交互关系及其随时间的变化,更准确地揭示用户情感的动态变化过程。实验结果表明,该方法在多个真实流式用户评论数据集上明显提升了情感分析的准确性。(2)针对流式用户评论的数据稀疏问题,本文提出了一种基于大语言模型数据增强的面向流式用户评论的情感分析方法。该方法首先根据流式用户评论稀疏的原因对用户进行分类,然后利用大语言模型的图理解能力,为不同类别的用户设计特定的方案,以选择性理解评论中的关键图元素,如局部-全局图结构、二阶关系及商品属性等。通过这种方法,成功生成了高质量的流式用户评论数据,有效弥补了实际数据的不足。实验结果显示,利用这些合成数据,情感分析的性能得到了明显提升,特别是在处理极度稀疏的流式用户评论数据时,性能提升更为明显。最后,基于上述研究,本文设计并实现了一个面向电商平台流式用户评论的情感分析系统。该系统简化了流式用户评论情感分析流程,并支持流式用户评论合成功能。该系统经过设计与测试验证,表现出稳定的性能与较高的准确性,为电子商务平台提供了有效的用户情感分析工具。更多还原

【Abstract】 With the rapid development of online e-commerce platforms and increasingly frequent user interactions,user review data has exploded in growth.Sentiment analysis of user reviews on e-commerce platforms,as a crucial means to gain insights into user feedback and optimize product services,has become increasingly important in the field of e-commerce.However,traditional sentiment analysis methods for e-commerce platform user reviews often rely on static assump-tions,assuming that user preferences remain unchanged over time.This assumption contradicts the streaming nature of user review texts and their emotions,which change over time in reality.Furthermore,the sparsity of user reviews is a common issue on e-commerce platforms,which becomes even more prominent in streaming scenarios segmented by time periods.Therefore,addressing the two issues of”static assumption-based”and”sparsity of stream-ing user reviews”in past methods,this paper focuses on sentiment analysis of streaming user reviews on e-commerce platforms,proposes new solutions,and integrates them into a system.The sentiment analysis of streaming user reviews on e-commerce platforms discussed in this paper aims to analyze users’diverse reviews of multiple products over past time periods and attempt to predict users’sentiment orientations in a future time period.This framework breaks the limitations of analyzing individual reviews,extending sentiment analysis to cover reviews across the entire time span,thereby revealing more comprehensively the dynamic trends of users’review text expressions and their emotions over time.The main work of this paper is as follows:(1)To address the problem that traditional static modeling methods cannot capture users’dynamic preferences,this paper designs a sentiment analysis method for streaming user reviews based on a dual-channel dynamic graph neural network.By establishing dedicated channels for both user and product dimensions,this method achieves a multi-dimensional capture of the dynamic changes in streaming user reviews and ratings.Specifically,it first utilizes a dual-channel text encoder to extract current local and global contextual information from review texts,serving as relevant information for users and products,respectively.Then,users and products are treated as nodes in a dynamic graph,with reviews serving as edges connecting the nodes.By deeply analyzing the interaction relationships between users and products and their changes over time,this method more accurately reveals the dynamic changes in user sentiment.Experimental results show that this method significantly improves the accuracy of sentiment analysis on multiple real-world streaming user review datasets.(2)To address the data sparsity issue in streaming user reviews,this paper proposes a sentiment analysis method for streaming user reviews based on large language model data aug-mentation.This method first classifies users based on the reasons for sparsity in streaming user reviews.Then,it utilizes the graph understanding capabilities of large language models to design specific solutions for different user categories,selectively understanding key graphi-cal elements in reviews,such as local-global graph structures,second-order relationships,and product attributes.Through this method,high-quality streaming user review data is success-fully generated,effectively compensating for the lack of real data.Experimental results show that using these synthetic data,the performance of sentiment analysis is significantly improved,especially when dealing with extremely sparse streaming user review data.Finally,based on the above research,this paper designs and implements a sentiment anal-ysis system for streaming user reviews on e-commerce platforms.This system simplifies the sentiment analysis process for streaming user reviews and supports the synthesis of streaming user reviews.After design and testing,the system demonstrates stable performance and high accuracy,providing an effective user sentiment analysis tool for e-commerce platforms.更多还原