【作者】 王洋；

【导师】 黄训江；

【作者基本信息】 东北大学 ， 技术经济及管理， 2020， 硕士

【摘要】 线上交易会产生大量的评论数据,评论数据能够为消费者和企业提供产品特征及情感倾向等多维度的信息,通过评论信息的挖掘,不仅能够提高消费者甄选商品的时间效率,影响消费者的购买决策,更能帮助企业从评论数据中获取消费者关注的产品特征及特征的情感倾向信息,从而对产品进行创新改进,进而在市场环境中获得竞争优势,本文以电商平台产品的线上评论为研究对象,利用TextRank算法从评论中识别受消费者关注的、利于商家进行创新改进的产品特征,针对传统TextRank算法中的点、边特征均权分配的缺点,对传统的TextRank算法进行改进,并利用真实的线上评论对改进后的方法进行验证。本文的主要工作如下:首先,本文依据人工建立的产品特征词库对分词结果进行修正,避免分词工具的错误识别及未登录词的遗漏,其次,针对TextRank算法中的边特征,本文利用Word2vec词向量模型对经结构化处理的文本数据进行训练,生成词向量,利用词向量间的余弦相似度构造相似度矩阵,并利用词汇之间的共现概率构造共现度矩阵,将二者进行线性融合,构造边特征矩阵,对传统图模型算法的边特征进行修正。最后,针对产品特征识别图模型的点特征,本文将粗粒度文本挖掘与细粒度文本挖掘相结合,计算候选产品特征词的语义相似度,同时,利用tf-idf算法计算词汇的频度特征,考虑词汇时间跨度及文本跨度计算候选产品特征词的跨度特征,利用序关系分析法对候选产品特征词的各个参数进行线性加权,得到各词汇的初始权重,对图模型算法中的点特征进行修正。本文针对在线评论产品创新特征识别问题,完善并改进了传统的TextRank算法,提出了融合图模型点特征及边特征的产品特征识别方法,并将其应用于华为Mate20X、iPad air、小米手环实例中,结果表明此识别方法在查全率等指标数据上均有较好的表现,实现了对中文产品评论中产品特征词的自动提取及重要性排序,同时根据本文的研究结果为企业的产品创新方向提供了可行性建议。本文提出的特征识别方法对于互联网上消费者关注度较高的产品具有较好的效果,能够为商家的产品创新战略提供指导性意见,并能为商家和消费者提供更加精准和简洁的产品信息。最后,本文仅从特征识别领域对产品创新特征的提取进行了研究,有关产品特征与消费者情感关联、引入第三方主观评判等的研究及特征权重分配问题尚需进一步探索。更多还原

【Abstract】 Online trading will produce a large number of reviews data,which can provide consumers and enterprises with multi-dimensional information such as product characteristics and emotional tendencies.Through the mining of review information,it can not only improve the time efficiency of consumers’ selection of goods,affect consumers’ purchase decisions,but also help enterprises to obtain the product characteristics and characteristics of consumers’ attention from the review data.In this paper,we use TextRank algorithm to identify the product features that are concerned by consumers and conducive to innovation and improvement of businesses.In view of the shortcomings of point and edge features in traditional TextRank algorithm,we use TextRank algorithm to assign equal rights to traditional TextRank algorithm is improved,and the improved method is verified by real online comments.The main work of this paper is as follows:First of all,this paper modifies the segmentation results according to the manually established product feature lexicon to avoid the wrong recognition of segmentation tools and the omission of unknown words.Secondly,for the edge features of TextRank algorithm,this paper uses Word2vec word vector model to train the structured text data,generate word vectors,and construct similarity moments using cosine similarity between word vectors Matrix,and use the co-occurrence probability between words to construct the co-occurrence matrix,the two are fused linearly,the edge feature matrix is constructed,and the edge feature of the traditional graph model algorithm is modified.Finally,according to the point features of the product feature recognition graph model,this paper combines coarse-grained text mining with fine-grained text mining to calculate the semantic similarity of candidate product feature words with the core word.At the same time,tf-idf algorithm is used to calculate the frequency characteristics of the words.Considering the time span and the text span,calculated the span characteristics of candidate product feature words.Using sequential relationship analysis,the parameters of feature words are weighted linearly to get the initial weight of each word,and the point features in the graph model algorithm are modified.In this paper,the traditional TextRank algorithm is improved,and the product feature recognition method based on the fusion of point feature and edge feature of graph model is proposed,which is applied to Huawei Mate20X,iPad air,MI bracelet.The results show that the recognition method has good performance in recall rate and other index data,and realizes the recognition of Chinese automatic extraction and importance ordering of product feature words in product reviews,at the same time,according to the research results of this paper,it provides feasible suggestions for the direction of product innovation of enterprises.The feature recognition method proposed in this paper has a good effect on the products with high attention of consumers on the Internet,can provide guidance for the product innovation strategy for businesses,and can provide more accurate and concise product information for businesses and consumers.Finally,this paper only studies the extraction of product innovation features from the field of feature recognition.The research on the relationship between product features and consumer emotion,the introduction of third-party subjective evaluation and the distribution of feature weight need to be further explored.更多还原