基于意见词的隐性产品特征提取方法研究及应用

发布时间：2018-04-13 15:47

本文选题：意见词 + 隐性产品特征　；参考：《东华大学》2016年硕士论文

【摘要】：目前,电子商务发展日益迅速,电商平台上每天都会有大量包括购买记录、产品评论等在内的数据信息,其中充分分析评论信息以得出用户对产品的情感倾向,将会给商家以及其他用户带来一定的参考价值。但是,仅仅知道用户的情感倾向,却无法得知用户是对产品哪一个特征进行的评论将会导致商家不知如何改进产品、其他用户无法对比选择。因此,为了使评论分析更加细粒化,基于意见词的特征挖掘研究是很有必要的。产品特征可分为显性产品特征和隐性产品特征,目前,大多数研究都关注显性产品特征的提取,而隐性产品特征的提取研究关注度较少。基于以上背景,本文以隐性产品特征提取为研究目标,以评论集为研究对象。本文的工作内容可概括如下:(1)本文针对现有的提取有效词的方法只考虑词频一个方面,考虑不够全面的问题,提出了综合加权的方法建立意见词和上下文词库。本文提出了一种词库建立的算法,该算法综合考虑了四个影响词的有效性的方面,并加权它们的权值。实验表明,此方法可以提高意见词和上下文词库建立的准确性。(2)本文针对现有的基于上下文的算法只考虑同一个评论句的上下文具有一定的片面性这个问题,提出了主题-意见词联合模型(JTO),通过在LDA模型中加入上下文层级来得到意见词在全部评论集中的上下文概率分布。实验表明,这种方法在隐性产品特征提取的准确性上比现有的基于上下文的隐性产品特征提取方法有更好的表现。(3)本文针对评论句中上下文信息不一定可信的问题,提出了考虑上下文权重的提取方法,以评估上下文信息的可信度,本方法的内容包括考虑词之间的距离改进共现矩阵、基于两个概率分布和余弦相似度计算上下文权重。实验表明,本方法在召回率和准确率方面都有了一定的提升。(4)本文基于上述两个模型实现了基于意见词的隐性产品特征提取原型系统,并将该系统应用于荣华饼家项目。
[Abstract]:At present, with the rapid development of electronic commerce, there are a large number of data information including purchase records, product reviews and so on on the e-commerce platform every day, in which the comments are fully analyzed in order to get the emotional tendency of the users towards the products.Will bring certain reference value to merchants and other users.However, just knowing the emotional tendency of the user, but not knowing which feature the user is commenting on, will lead to the merchants do not know how to improve the product, other users can not compare the choice.Therefore, in order to make comment analysis more fine-grained, it is necessary to study feature mining based on comment words.Product features can be divided into dominant product features and hidden product features. At present, most researches focus on the extraction of dominant product features, but less attention is paid on the extraction of hidden product features.Based on the above background, this paper takes recessive product feature extraction as the research object and comments set as the research object.The work of this paper can be summarized as follows: (1) in view of the existing methods of extracting valid words, only one aspect of word frequency is considered, and the problem is not comprehensive enough, a comprehensive weighted method is proposed to establish the word base of opinion and context.In this paper, a lexicon building algorithm is proposed, which synthetically considers four aspects that affect the validity of words and weights their weights.The experimental results show that this method can improve the accuracy of the construction of comment words and context lexicon.) aiming at the problem that the existing context-based algorithms only consider the one-sidedness of the context of the same comment sentence,In this paper, a topic-opinion word association model is proposed, and the context probability distribution of the comment word in the whole comment set is obtained by adding context level to the LDA model.Experimental results show that this method has better performance than the existing context-based implicit product feature extraction methods.) this paper aims at the problem that the contextual information in comment sentences is not always credible.In order to evaluate the credibility of context information, a method of extracting context weight is proposed. The content of this method includes considering the distance between words to improve co-occurrence matrix and calculating context weight based on two probability distributions and cosine similarity.Experiments show that this method has a certain improvement in recall rate and accuracy.) based on the above two models, this paper implements a prototype system of recessive product feature extraction based on opinion words, and applies the system to Rong Hua pie house project.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1

【参考文献】