【作者】 夏雨；

【导师】 姚寒冰；

【作者基本信息】 武汉理工大学 ， 计算机应用技术， 2016， 硕士

【摘要】 随着WEB 2.0技术的快速发展,几乎所有类型的网络平台上都会有评论信息的功能模块,越来越多的消费者会在这些网络平台上发表产品评论信息。这些产品评论中蕴藏着丰富的有价值的信息,一方面可以帮助消费者做出正确的购买决策,另一方面可以帮助企业了解消费者的切实需求,逐步完善产品质量。由于在线产品评论规模庞大,人工难以阅读完所有的评论。因此,对在线产品评论进行自动分析,挖掘出评论文本中的意见要素,并以清晰、直观的方式对结果进行展示,具有重要的研究价值。本文从细粒度层面,即基于产品特征,对在线产品评论进行意见挖掘,主要的工作内容总结如下:1)针对现有的评论要素提取查全率低、具有领域依赖性的问题,对大量评论语料的词法结构进行分析后,本文提出了基于副词的评论要素的迭代提取方法。首先基于跨领域的种子副词实现副词与意见词的迭代提取,然后基于意见词实现特征词与意见词的迭代提取。另外,针对提取过程中可能产生噪声的原因,本文使用了三种方法对候选产品特征进行修剪过滤。2)由于不同的用户可能使用不同的词语来表示同一个产品特征,因此需要将各个特征的同义词进行合并。在现有的同义词识别的研究中,基于词汇字面计算相似度或者基于语义词典计算相似度的方法都存在固有的缺陷。鉴于此,本文充分利用已有的词典资源,并结合特征词汇所在的上下文信息制定相似度计算规则,在一定程度上互补,提高了同义产品特征词识别的查全率。3)针对已有的极性判断方法无法判断部分意见词极性的缺点,本文提出了基于聚类算法的情感极性判断方法。通过对大量评论语料的统计,总结影响特征意见词对情感极性的因素,以此制定规则来初始化评价单元(特征词,意见词,否定词个数,修饰副词)之间的相似度。然后基于改进的k-中心点聚类算法对评价单元进行聚类,实现同极性的评价单元聚在同一簇中,最后根据每个簇中的种子情感词的极性来确定整个簇的极性。最后,基于上述的工作成果,本文设计并实现了在线产品评论意见挖掘原型系统。该系统能够有效地挖掘出评论中蕴藏的有价值的信息,展示消费者们对产品各个特征的情感倾向情况。更多还原

【Abstract】 With the rapid development of the WEB 2.0 technology,almost all types of network platform have the module of reviews,more and more consumers tend to deliver their opinions about the product at the network platform.Comments on these products are rich in valuable information,on the one hand,it can help consumers to make the right buying decision,but on the other hand,it can help enterprises to understand the consumers’ real needs,gradually improve the quality of their products.Due to the large scale of online product comments,it is hard to read all the comments manually.Therefore,analyzing the online product reviews automatically,mining the opinion elements from the comment text,and showing the results in a clear and intuitive way have important research value.This thesis uses the fine-grained opinion mining technology,namely,based on product features to analysis online comments.The main content of this thesis is summarized as follows:1)Regarding problems of domain dependency and low recall,and according to analyze the lexical structure of a large number of comments corpora,this thesis proposes the adverbs based opinion elements extracting approach.First of all,realizing the iterative extraction of adverbs and opinion words based on the interdisciplinary seed adverbs,then realizing the iterative extraction of feature words and opinion words based on the opinion words.In addition,regarding the possible causes of noise in the process of extraction,this thesis uses three pruning methods to remove invalid feature words.2)Due to different users may use different words to express the same product features,so we need to merge the feature words that having the same meaning.Regarding both the semantic dictionary based similarity computing method and the literal based similarity computing method have their deficiencies,this thesis makes full use of existing dictionary resources,and the contextual constraint relations between feature words are mined to compute similarity,the two methods can complement each other in some degree and enhance the recall of recognizing the feature words that having the same meaning.3)Regarding the problem that the existing methods could not predict the orientation of some opinion words,this thesis proposes the clustering algorithm based predicting orientation approach.Through statistics of a large number of comments corpora,this thesis summarizes some factors that affecting the orientation of feature-opinion pairs,in order to make the rules to initialize the similarity between the assessment unit(features,opinion,number of negatives,modification adverbs).Then using the improved k-medoids algorithm to classify the assessment unit,so that the units that have the same orientation are in the same cluster,finally,predicting the orientation of the whole cluster according to the orientation of the seed opinion words.Finally,based on the work above,this thesis designs the online product reviews mining system.The system can mine the valuable information from the online product reviews,and show consumers’ emotional orientation to each product feature.更多还原