【摘要】：隨著互聯(lián)網(wǎng)的發(fā)展,特別是電子商務(wù)的飛速發(fā)展,越來越多的消費(fèi)者青睞于網(wǎng)上購物,消費(fèi)者越來越容易針對自己購買的產(chǎn)品發(fā)表評論,這些產(chǎn)品評論信息為廠家以及潛在消費(fèi)者提供了寶貴的信息資源。由于存在某些利益關(guān)系,其中可能存在一些不實(shí)或虛假的內(nèi)容,這些虛假評論在一定程度上影響了評論信息的參考價(jià)值,從而誤導(dǎo)消費(fèi)者,因此檢測虛假評論尤為重要。最基本的評論信息是評論的內(nèi)容信息,對評論內(nèi)容信息進(jìn)行挖掘,利用評論內(nèi)容信息對虛假評論進(jìn)行檢測有著極其重要的意義；此外,對評論者行為進(jìn)行挖掘,通過發(fā)現(xiàn)異常的行為模式來識別虛假評論也有著重要的作用。本文以產(chǎn)品和服務(wù)評論為主,圍繞基于評論內(nèi)容的虛假評論檢測、基于評論者行為的虛假評論檢測、融合評論內(nèi)容及評論者行為這兩類特征來檢測虛假評論等關(guān)鍵問題開展研究,主要完成了以下研究工作： (1)提出了一種基于評論內(nèi)容的虛假評論檢測方法。該方法首先構(gòu)建基于情感依賴的評論主題-對立情感依賴模型(topic-opposite sentiment dependency model, TOSDM),利用該模型提取評論的主題信息以及主題對應(yīng)的情感信息；然后,結(jié)合評論的主題以及情感信息,分析并提取6維評論內(nèi)容特征；最后,利用這些評論內(nèi)容特征,采用有監(jiān)督學(xué)習(xí)的分類器對虛假評論進(jìn)行檢測。 (2)提出了一種基于評論者行為的虛假評論檢測方法。該方法首先根據(jù)評論數(shù)據(jù)選取10維反映評論者行為的特征,并對每維特征進(jìn)行歸一化處理；然后,根據(jù)每一條評論的特征構(gòu)建聚類矩陣,利用F統(tǒng)計(jì)量對K均值算法進(jìn)行改進(jìn),實(shí)現(xiàn)評論數(shù)據(jù)的自適應(yīng)聚類；最后,計(jì)算每個(gè)簇偏離整個(gè)評論數(shù)據(jù)集的程度,根據(jù)閾值確定異常簇,從而實(shí)現(xiàn)虛假評論檢測。 (3)提出了一種融合評論內(nèi)容及評論者行為的半監(jiān)督虛假評論檢測方法。該方法首先對評論的內(nèi)容特征以及評論者的行為特征進(jìn)行提取,然后借助Co-Training的半監(jiān)督學(xué)習(xí)思想,將這兩類特征看作相互獨(dú)立的視圖,利用這兩類獨(dú)立的特征分別建立分類器,挑選置信度高的未標(biāo)注樣本,最后使用這些挑選出的樣本更新訓(xùn)練模型,改善分類器效果。 (4)設(shè)計(jì)并實(shí)現(xiàn)了虛假評論檢測原型系統(tǒng),為進(jìn)一步研究虛假評論的檢測方法提供了便利。
[Abstract]:With the development of the Internet, especially the rapid development of electronic commerce, more and more consumers prefer to shop online, and it is more and more easy for consumers to comment on the products they buy. These product reviews provide valuable information resources for manufacturers and potential consumers. Due to the existence of some interest relations, there may be some false or false content, these false comments to a certain extent affect the reference value of comment information, thus misleading consumers, so it is particularly important to detect false comments. The most basic comment information is the content information of the comment. It is very important to mine the content information of the comment and detect the false comment by using the content information of the comment; in addition, it is very important to mine the behavior of the reviewer. Identifying false comments by discovering abnormal behavior patterns also plays an important role. This paper focuses on product and service reviews, focusing on the detection of false comments based on the content of comments, and the detection of false comments based on the behavior of reviewers. The key issues of detecting false comments such as comment content and reviewer behavior are studied. The main works are as follows: (1) A method of false comment detection based on comment content is proposed. In this method, a motif based on affective dependency is constructed, which is used by topic-opposite sentiment dependency model, TOSDM), to extract the subject information of comments and their corresponding emotional information, and then combines the subject and emotional information of comments. Finally, a supervised learning classifier is used to detect false comments. (2) A method of false comment detection based on reviewer's behavior is proposed. The method firstly selects 10 dimensions to reflect the behavior of the reviewer according to the comment data, and normalizes the feature of each dimension. Then, the clustering matrix is constructed according to the characteristics of each comment, and the K-means algorithm is improved by using F statistics. Finally, the degree of each cluster deviating from the whole comment data set is calculated, and the abnormal cluster is determined according to the threshold. Thus, the detection of false comments is realized. (3) A semi-supervised detection method of false comments is proposed, which combines the content of comments with the behavior of the reviewers. The method firstly extracts the content features of comments and the behavioral features of reviewers. Then, with the help of Co-Training 's semi-supervised learning idea, the two kinds of features are regarded as independent views, and the classifiers are constructed using the two independent features. The unlabeled samples with high confidence are selected. Finally, the training model is updated with these selected samples to improve the classifier effect. (4) A prototype system of false comment detection is designed and implemented. It is convenient to further study the detection method of false comment.
【學(xué)位授予單位】：昆明理工大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2014
【分類號】：TP393.08

【參考文獻(xiàn)】

相關(guān)期刊論文 前10條

1 譚文堂;朱洪;葛斌;李芳芳;肖衛(wèi)東;;垃圾評論自動過濾方法[J];國防科技大學(xué)學(xué)報(bào);2012年05期

2 曾雪強(qiáng),王明文,陳素芬;一種基于潛在語義結(jié)構(gòu)的文本分類模型[J];華南理工大學(xué)學(xué)報(bào)(自然科學(xué)版);2004年S1期

3 孫升蕓;田萱;;產(chǎn)品垃圾評論檢測研究綜述[J];計(jì)算機(jī)科學(xué);2011年S1期

4 邱云飛;王建坤;邵良杉;劉大有;;基于用戶行為的產(chǎn)品垃圾評論者檢測研究[J];計(jì)算機(jī)工程;2012年11期

5 魏小娟;李翠平;陳紅;;Co-Training——內(nèi)容和鏈接的Web Spam檢測方法[J];計(jì)算機(jī)科學(xué)與探索;2010年10期

6 解曉敏;李云;;最小最大模塊化網(wǎng)絡(luò)中基于聚類的數(shù)據(jù)劃分方法研究[J];南京大學(xué)學(xué)報(bào)(自然科學(xué)版);2012年02期

7 宋海霞;嚴(yán)馨;余正濤;石林賓;蘇斐;;基于自適應(yīng)聚類的虛假評論檢測[J];南京大學(xué)學(xué)報(bào)(自然科學(xué)版);2013年04期

8 張倩;瞿有利;;用于網(wǎng)絡(luò)評論分析的主題-對立情感挖掘模型[J];計(jì)算機(jī)科學(xué)與探索;2013年07期

9 周志華;;基于分歧的半監(jiān)督學(xué)習(xí)[J];自動化學(xué)報(bào);2013年11期

10 趙妍妍;秦兵;劉挺;;文本情感分析[J];軟件學(xué)報(bào);2010年08期

相關(guān)博士學(xué)位論文 前1條

1 李方濤;基于產(chǎn)品評論的情感分析研究[D];清華大學(xué);2011年

，

本文編號：2185124