| 摘要: | 在現代的社會中網路購物逐漸的普及，普遍群眾都有網路購物的經驗，而消費者在網路購物中可能會遇到許多問題，進而提出疑問甚至是產生客訴，隨著網際網路和社群媒體的發展，消費者開始在論壇、聊天室、Blog和社群媒體上分享產品和服務的使用體驗和意見，研究指出線上評論對消費者的決策有重大影響，並對產品銷售產生積極影響，了解消費者對網路購物平台的評價和意見對於企業改善服務品質和了解消費者需求至關重要，在Web 2.0時代，使用者生成內容就成為了重要資源，提供了消費者真實體驗和資訊交流的管道，也是一般企業在客戶服務管道無法取得足的資訊，因此主動搜集社群媒體上的資訊並進行系統化分類可以節省時間且蒐集更多資訊，利用社群媒體數據分析的方式，將網路論壇上的評論資料轉化為有關網路購物的有意義的知識，本研究抓取Dcard網路論壇之網路購物看板之資料作為研究中使用的資料集，將評論內容透過文字清洗、取得文件特徵，將非結構化資料轉變為可以使用的資料，研究中透過詞頻-逆文本頻率(Term Frequency-Inverted Document Frequency, TD-IDF)、Sentence-BERT (SBERT)以及fastText取得文件特徵，以及搜集台灣前五大網路購物平台之客服專區提供的問題，統計六種網路購物問題，再對評論內容進行分類的標記，最後實驗使用Orange3資料探勘工具，將三種文件特徵透過單一文件特徵以及組合文件特徵兩種方式實驗，透過工具內的五種監督式學習模型支援向量機(Support Vector Machine, SVM)、隨機森林(Random Forest, RF)、邏輯斯回歸(Logistic Regression, LR)、K-近鄰演算法(K-Nearest Neighbors, KNN)以及樸素貝葉斯(Naïve Bayes, NB)等五種較常使用在分類任務研究的演算法來對網路購物評論進行分類，研究結論指出，在使用單一文件特徵時以fastText方法使用LR模型表現最佳，而在使用組合文件特徵依然是LR模型表現最好，而研究發現在文件特徵部分組合中有混和fastText方法能夠有效提升模型效能及分類準確度。;In modern society, online shopping has gradually become widespread, and the general public has experience with it. Consumers may encounter various problems while shopping online, leading to questions and even complaints. With the development of the Internet and social media, consumers have started sharing their product and service experiences and opinions on forums, chat rooms, blogs, and social media platforms. Studies have shown that online reviews have a significant impact on consumer decision-making and positively affect product sales. Understanding consumer evaluations and opinions on online shopping platforms is crucial for businesses to improve service quality and understand consumer needs. In the era of Web 2.0, user-generated content has become an important resource, providing a channel for consumers to share their real experiences and exchange information, which is often not accessible through traditional customer service channels. Proactively collecting and systematically categorizing information from social media can save time and gather more information. By analyzing social media data, online forum reviews can be transformed into meaningful knowledge about online shopping. This study uses data from the Dcard online forum′s online shopping board as the dataset, processing review content through text cleaning and extracting document features, transforming unstructured data into usable information. Term Frequency-Inverted Document Frequency (TD-IDF), Sentence-BERT (SBERT), and fastText are used to obtain document features, along with collecting questions provided by the customer service sections of the top five online shopping platforms in Taiwan. Six types of online shopping problems are identified, and review content is categorized accordingly. Orange3 data mining tool is used in the experiment, applying both single document features and combined document features approaches. Five supervised learning models, Support Vector Machine (SVM), Random Forest (RF), Logistic Regression (LR), K-Nearest Neighbors (KNN), and Naïve Bayes (NB) are used for classification tasks. The study concludes that when using single document features, the best performance is achieved with the fastText method and the LR model, while the LR model still performs best when using combined document features. The study finds that incorporating the fastText method in the feature combination can effectively improve model performance and classification accuracy. |
