【作者】 孙春华；

【导师】 刘业政；

【作者基本信息】 合肥工业大学 ， 管理科学与工程， 2012， 博士

【摘要】 在线评论是用户创造并发表在零售商或者第三方网站上的产品信息，由于包含了先前消费者的产品评价和使用体验等情感信息，而被消费者认为具有更高的可信度和感染力，成为供应商的产品信息的一个补充。先前研究显示，评论的情感表达包括有无情感信息、情感信息的倾向性等直接影响了消费者的产品态度和购买行为。由于大多数评论网站上的在线评论是自由文本格式的非结构化数据并且数量庞大，过去的研究多使用在线评分数据替代在线评论文本的情感倾向性。近年来，文本挖掘领域的技术发展为在线评论的情感信息抽取及倾向性分析提供了便利，应用文本挖掘技术，研究在线评论文本的情感表达对消费者感知和行为的影响将有助于企业理解消费者行为，制定有效的营销策略。论文以评论网站的在线评论和有用性投票数据为主要研究数据来源，以帮助企业及网站优化评论系统、制定更加有效的营销策略为目标，研究情感表达对评论有用性感知的影响。论文研究的主要思路是，首先分析在线评论的信息结构因素，包括在线评论的数量、长度、方向和类型对消费者感知和购买意图的影响，相关结论作为下文研究的理论框架，然后重点研究在线评论的情感表达对评论有用性感知的影响，包括情感词抽取及倾向性分析方法，产品特征词和情感词对评论有用性感知的影响，情感倾向性的方向、强度和混合程度对评论有用性感知的影响。论文的具体研究内容如下：（1）信息结构对消费者感知和行为的影响。论文界定了在线评论信息结构的四个维度，构建了在线评论信息结构对消费者感知和行为影响的模型。该模型揭示了信息结构因素对消费者感知和行为的作用差异以及信息结构因素影响消费者感知和行为的路径。论文采用实验研究方法，验证了评论数量、长度、类型对评论有用性感知的作用，评论数量、类型对评论可信度感知的作用，评论数量对产品流行性感知的作用，以及评论有用性感知、评论可信度感知和产品流行性感知对消费者购买意图的作用。（2）情感词抽取和倾向性分析方法。论文研究并提出了产品特征词和情感词的抽取方法、含强度的基准词选择和词汇倾向性判别方法、基于产品特征词关系识别的评论倾向性合成方法。论文针对评论特点，选择了基于统计的词汇抽取方法，实验结果证实了该方法的有效性。论文提出了一种含强度的基准词选择和词汇倾向性判别方法，该方法解决了以下两个问题：一是词汇情感强度的判别问题，二是依赖于上下文的情感词倾向性判别问题。论文提出了一种基于产品特征词关系识别的评论倾向性合成方法，该方法适用于包含多个产品特征词的评论，尤其是结构和功能相对复杂的产品的评论。（3）产品特征词和情感词对评论有用性感知的影响。论文基于产品质量感知的有关理论，将产品特征词划分为内在属性词和外在属性词，将情感词划分为评价词和情绪词，研究了在线评论中的内在属性词、外在属性词、评价词和情绪词对评论有用性感知的影响。论文以手机和音乐CD作为搜索型产品和体验型产品的代表，应用文本挖掘方法从手机评论和音乐CD评论中抽取产品特征词和情感词，对产品特征词和情感词进行类别判定，使用评论网站的有用性投票数据来检验产品特征词和情感词的作用。论文研究发现，对于搜索型产品而言，内在属性词和评价词对评论有用性的作用更大，对于体验型产品而言，外在属性词和情绪词对评论有用性的作用更大。（4）情感倾向性方向、强度和混合程度对评论有用性感知的影响。论文将有用性投票阶段分为注意阶段和评价阶段，研究了评论的情感倾向性方向、强度、混合程度在两个阶段中的作用。论文使用文本挖掘方法从在线评论中抽取情感词并计算情感词、上下文和评论的倾向性，通过模型建立和参数估计来检验评论倾向性对总投票和有用性的作用。手机评论集的数据分析结果表明，具有强烈情感倾向性的评论更容易引起消费者注意并得到有用或者无用的投票，表达了不同方向或者强度的观点的评论被认为更有用，总体而言负面评论更有用。论文研究结果可以帮助网络零售商优化评论系统，制定更有效的营销策略。更多还原

【Abstract】 Online reviews are user-generated product information posted on retailer or third partywebsites. Including previous consumers’ subjective sentiment information such as productevaluation and user experience, online reviews are consider to be more credible and infectious,and become a complement to product information from suppliers. Previous studies indicated thatsentiment expression of online reviews has direct impact on consumer product attitude and purchasebehavior. Because online reviews from most websites are written in free-text format with very scantstructured metadata information and the volume of online reviews are very large. Most of previousstudies use online ratings instead of online reviews’ sentiment orientation. In recent years, text miningtechnologies facilitate the sentiment information extraction and sentiment orientation analysis of onlinereviews. Research on the impact of sentiment expression on consumer perception and behavior throughtext mining technologies can help the company to understand consumer behavior and make marketingstrategies more effectively.Using online reviews and helpfulness votes as the main research data, this paper investigatesthe impact of sentiment expression on consumer perception of reviews helpfulness. Our aim is tohelp the company to optimize the online reviews system and make marketing strategies moreeffectively. This pape first investigates the effects of information structure factors of onlinereviews including review number, review length, review direction and review type on consumerperception and behavior. The results of this research serve as theoretical framework. Then thispaper focuses on the impact of sentiment expression on consumer perceived reviews helpfulness.Specifically, this involves three parts: methods for sentiment information extraction and sentimentorientation analysis; the impact of product feature words and sentiment words on consumerperceived reviews helpfulness and the impact of the sentiment orientation including direction,intensity and admixture on consumer perceived reviews helpfulness. The detailed researches are asfollows:(1) The influence of information structure on consumer perception and behavior. This paperdefines four dimensions of information structure and proposes the model to describe the influenceof information structure factors of online reviews on consumer perception and behavior. In thismodel, the differential effects of information structure factors are analyzed and the paths ofinfluence are also revealed. We conduct a2×2×2×2between-subjects factorial design experimentto validate the hypotheses. The results show the effects of the number, the length and the type ofonline reviews on the perceived reviews helpfulness, the effects of the number and the type ofonline reviews on the perceived reviews credibility, the effect of the number of online reviews onthe perceived product popularity as well as the effects of the perceived reviews helpfulness, theperceived reviews credibility and the perceived product popularity on purchase intension. Theresults and implications of this research are discussed.(2) Methods for sentiment information extraction and sentiment orientation analysis. Thispaper investigates the method for extracting product feature words and sentiment words from online reviews, the method for the paradigm words selection with intensity information and wordsentiment orientation discrimination and the method for sentiment orientation combination basedon product features relationship recognition. This paper adopts the statistic-based method toextract the words. The method is proved to be effective in the experiment. This paper proposes amethod for the paradigm words selection with intensity information and word sentimentorientation discrimination. With this method, we can distinguish the direction and intensity ofsentiment words as well as differentiate the sentiment orientation of the words whose orientatindepending on the context. This paper proposes a method for sentiment orientation combinationbased on product features relationship recognition. The method can be used in sentiment analysisfor the reviews covering several product feature words, especially for the reviews of complex andmulti-function product.(3) The impact of product feature words and sentiment words on consumer perceivedreviews helpfulness. Based on theories of product quality perception, this paper dichotomizeproduct feature words into intrinsic attribute words and extrinsic attribute words and dichotomizesentiment words into evaluation words and emotion words. This paper investigate the impact ofintrinsic attribute words, extrinsic attribute words, evaluation words and emotion words onconsumer perception of reviews helpfulness. Mobile and music CD are selected as therepresentation of search product and experience product. Product feature words and sentimentwords are extracted and classified through text mining methods. The effects of product featurewords and sentiment words on consumer perception of reviews helpfulness are tested using thehelpfulness voting data. The study shows that intrinsic attribute words and evaluation words havemore effects on perceived helpfulness for search product while extrinsic attribute words andemotion words have more effects on perceived helpfulness for experience product.(4) The impact of the sentiment orientation on consumer perceived reviews helpfulness.This paper divides helpfulness voting process into two stages and then investigates the differentialeffects of sentiment orientation including direction, intensity and admixture during these twostages. This paper extracts sentiment words from online reviews and calculates sentimentorientation of the word, the context and the whole text using text mining methods. The effects ofsentiment orientation on total votes and helpfulness are tested through statistic model andparameters estimate. The results from mobile reviews dataset show that the reviews with higherintensity attract more attention and then get a large number of total votes; the reviews with higheradmixture are perceived to be more helpful; negative reviews are perceived to be more helpful.The results of this paper can help online sellers to optimize the online reviews system and makemarketing strategies more effectively.更多还原