基于特征的商品在线评论情感倾向性分析

发布时间：2019-03-28 06:24

【摘要】：随着互联网技术和电子商务的快速发展,我们已经进入了“全民网购”的时代。消费者对商品的在线评论为其他消费者、企业产品反馈提供了重要的资源。因此,如何高效、自动化的剖析在线评论中消费者对产品及其相关特征所持有的态度成为情感倾向性分析领域的热点课题。然而,由于中文自然语言本身的多样性和复杂性,加之网络语言的非规范性,让商品在线评论的分析和研究变得更加困难。本文针对目前商品在线评论的情感分析领域中存在的难题,研究了特征级文本情感倾向性分析的理论方法及实现算法,根据商品在线评论的文本特点,提出了基于句型结构、词性规律搭配的在线评论的特征情感三元组提取方法,并根据特征情感三元组的数据,用神经网络算法进行商品在线评论情感判别。主要研究工作如下:1.研究了基于特征的文本情感倾向性分析的相关理论,对篇章级、句子级、实体特征级三种不同粒度的文本情感分析方法进行比较,得出商品在线评论的文本情感分析中实体特征级的方法能够提供商品的更详细的情感倾向,优于其他两种粒度的情感分析方法。2.提出了一种基于句型结构、词性规律搭配的商品在线评论的特征情感三元组提取方法:首先收集领域依赖特征词集和网络流行情感词集,将网上获取的评论数据经过预处理,得到的主观子句提取句型模式,根据词性规律并判断子句的句型模式是否与我们提出的6种在线评论的基本句型模式相同,若相同,则提取对应的特征情感三元组;最后,将提取出的特征情感三元组进行去噪处理。该方法完美融合了商品在线评论的领域依赖性、程度副词对情感极性的影响等因素,有效的提取了特征情感三元组,并提高了商品在线评论中产品特征及其相关情感的识别能力。3.研究了基于人工神经网络的商品在线评论文本情感倾向性分析。首先建立BP神经网络及RBF神经网络模型进行训练并仿真;针对神经网络算法收敛速度慢、易陷入局部最优等缺点,本文基于全局寻优的思想对神经网络算法进行改进。改进神经网络算法是将神经网络模型中的各参数值进行全局优化,将隐含层中的权值整合,利用全局寻优的特点,确定神经网络中各参数最合理的值,从而提高神经网络算法的性能。分别对改进的神经网络模型进行训练并仿真;最后对本文中的四种神经网络从不同的维度进行实验比较。仿真实验表明,PSO-BP算法在处理商品在线评论情感倾向性分析问题中有更高的准确率,但收敛速度更慢;基于PSO-RBF神经网络在处理情感分析问题时有较高的正确率,且收敛速度比PSO-BP网络更快。
[Abstract]:With the rapid development of Internet technology and e-commerce, we have entered the era of "people-wide online shopping". Online reviews of goods by consumers provide important resources for feedback from other consumers and enterprises. Therefore, how to analyze the consumers' attitude towards the product and its related characteristics efficiently and automatically has become a hot topic in the field of emotional orientation analysis. However, due to the diversity and complexity of Chinese natural language itself, and the non-standardization of network language, it is more difficult to analyze and study the online comment of goods. Aiming at the difficult problems existing in the field of emotional analysis of commodity online comment, this paper studies the theoretical method and realization algorithm of emotional tendency analysis of feature-level text, and puts forward a sentence structure based on the text characteristics of commodity online comment. The feature emotion triple is extracted from the online comment based on the part of speech rule. According to the data of the feature emotion triple, a neural network algorithm is used to judge the online comment emotion of the commodity. The main research work is as follows: 1. This paper studies the related theory of feature-based text affective tendency analysis, and compares three kinds of text emotion analysis methods with different granularity, such as text level, sentence level and entity feature level. It is concluded that the method of entity feature level in the text emotion analysis of commodity online comment can provide the more detailed emotion tendency of commodity, which is better than the other two kinds of granularity emotion analysis method. 2. In this paper, a method of extracting feature emotion triple of online comments based on sentence structure and part of speech law is proposed. Firstly, domain dependent feature word set and network popular emotion word set are collected, and the online comment data are preprocessed. According to the rule of part of speech and judging whether the sentence pattern of the clause is the same as the basic pattern of the six online comments, if the pattern is the same, the corresponding feature emotion triple is extracted. Finally, the extracted feature emotion triples are de-noised. This method combines the domain dependence of commodity online comment, the influence of degree adverbs on emotional polarity and so on, and extracts the characteristic emotional triple effectively. It also improves the recognition ability of product features and related emotions in online reviews of goods. 3. This paper studies the emotional orientation analysis of commodity online comments based on artificial neural network (Ann). Firstly, the BP neural network and RBF neural network model are established for training and simulation. Aiming at the shortcomings of slow convergence rate and easy to fall into local optimization, this paper improves the neural network algorithm based on the idea of global optimization. The improved neural network algorithm is to globally optimize the parameters of the neural network model, integrate the weights in the hidden layer, and make use of the characteristics of the global optimization to determine the most reasonable values of the parameters in the neural network. Thus, the performance of neural network algorithm is improved. The improved neural network model is trained and simulated respectively. Finally, four kinds of neural networks in this paper are compared with each other from different dimensions. The simulation results show that the PSO-BP algorithm has higher accuracy in dealing with the emotional tendency analysis of online comments, but the convergence speed is slower. The PSO-RBF-based neural network has a higher accuracy rate when dealing with emotional analysis problems, and the convergence speed is faster than that of the PSO-BP network.
【学位授予单位】：上海师范大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1

【参考文献】