識別攻擊評論之研究-以旅館業為例

Identifying false review comments for Hostel Industry

林美玉(Mei-yu Lin)

指導教授 ： 許秉瑜

國立中央大學/管理學院/企業管理學系/碩士(2015年)

摘要

如今，身處網際網路盛行、資訊爆炸時代的消費者越來越敢於發表對商品的真實看法。網路評論成了企業所供一般資訊和推薦系統自動產生個人化建議之外重要訊息來源。各類網站累積了數以萬計消費者對不同商品或服務的各種評論。企業應重視這些促成消費個體間經驗交流和建立品牌口碑的網路評論可為企業帶來正面或負面的影響。亦有相關研究證實消費者對於來自其他顧客親身體驗的評論日益重視，其中負向評論對消費者做購買決策時的影響更為明顯。以致有些不肖業者利用此趨勢操控評論來不實宣揚自身商品優點或毀壞競爭對手品牌商譽，造成消費者和商業體系受到嚴重損害。
本研究以負向評論為研究目標，用TripAdvisor網站旅館真實評論與Amazon眾包平台集得虛假評論為分析對象，擷取3個關鍵屬性：模糊詞彙、虛假評論和真實評論重要詞彙。運用文字探勘技術，結合羅吉斯迴歸來模擬投入不同比率虛假評論的資料環境，建立一自動且準確預測虛假與真實評論之模型。本研究所發展出的模型結果顯示，虛假評論重要詞彙與模糊詞彙數量越多，評論為捏造的可能性越高。本研究測試資料在投入10%假評下呈現效果為精確度（Precision）=100％，準確度（Accuracy）=51.5％，召回率（Recall）=3％；而投入100％假評下，精確度（Precision）=64％，準確度（Accuracy）=64％，召回率（Recall）=64％，所有結果皆優於對照實驗。

關鍵字

謠言 ； 攻擊評論 ； 文字探勘 ； 羅吉斯迴歸

並列摘要

Nowadays, consumers are inclined to issue their opinions for merchandise in the era of web2.0. Therefore, besides the product information provided by companies the review comments provided by general public on the Internet have become another major information source for consumers. As a result, tens of thousands of review comments about different products or services are accumulated on various websites everyday. It has been found that to manipulate customer opinions, some dealers created the review comments in order to exaggerate the advantages of their own products or defame rival’s reputation.
This study strived to identify the negative fake review comments which were falsely created and aimed at attacking targeted products. The method created three word banks, namely, vagueness, and positive and negative attacks. The number of these words appearing in each review comments were calculated and applied to build logistic regression models.
The experiment was conducted with true hostel review comments taking from “TripAdvisor” and the comparison group “Fake reviews” on Amazon Mechanical Turk. In the case where the ratio of fake and true review comments are10% in the training data, the proposed method reached 100%, 51.5% and 3% of precision, accuracy and recall, respectively. When the raio is 50%, the method could reach 64%, 64%, 64% of precision, accuracy and recall respectively. The performance is better then the benchmark method which based on LIWC and SVM.

並列關鍵字

Rumors ； False reviews ； Text mining ； Logistic Regression

參考文獻

3. 張甜，“辨別虛假評論之研究－以旅館業為例＂，國立中央大學，碩士論文，2014。

1. Allport, G. W. & Postman, L., “An Analysis of Rumor”, Public Opinion Quarterly, (10), 1947: pp.501-517.

2. Allport, G. W. & Postman, L., “The Psychology of Rumor”, New York: Henry Holt, 1947.

3. Berry, M. J. A. & Linoff, G., “Data Mining Techniques: for Marking, Sales, and Customer Support”, New York: John Wiley & Sons Inc., 1997.

5. Chevalier, J. A. & Mayzlin, D., “The Effect of Word of Mouth on Sales: Online Book Reviews [J]”, Journal of Marketing Research, 43(3), 2006: pp.345-354.