隨著電子商務與社群媒體的普及，越來越多消費者在購物時仰賴網路評論作為參考依據。在實際購買產品或服務前，許多人會先閱讀他人分享的經驗，以減少資訊落差與降低錯誤決策的風險。然而，虛假評論問題日益嚴重，不僅誤導消費者的判斷，也進一步破壞市場的公平性與平台的可信度。特別是在推薦系統與生成式人工智慧技術迅速進步的情況下，假評論的數量、傳播速度與擬真程度都有顯著提升，使得傳統的偵測方法越來越難以應對這些新型態的挑戰。過去的研究多聚焦於語言特徵的分析，對評論者的行為模式與模型的可解釋性則關注較少，導致其在實務應用上的彈性與拓展性受限。近年來，研究逐漸朝向整合多重特徵與深度學習技術的方向發展，然而，由於深度學習模型本身缺乏透明性，使得管理者與實務使用者在信任模型預測結果方面仍存疑慮，進而影響其在實務上的採用意願與可行性。 基於此，本研究致力於整合來自不同平台的多來源文本與結構化資料，並結合自注意力機制，提出一個具備高泛化能力與解釋性的多模態假評論偵測模型，以更有效地因應目前多樣且複雜的虛假評論問題。本研究資料來源包括 Yelp、Amazon，以及使用 ChatGPT 所生成的虛假評論，以模擬生成式 AI 帶來的挑戰與風險。整體研究分為三個階段：第一階段針對文字特徵與各類機器學習模型進行測試，選出最佳基礎模型組合；第二階段則將評論者行為特徵與商品屬性等結構化資料納入，建構多模態模型，以強化其整體偵測效能與跨平台適應能力；第三階段進一步導入自注意力機制，強化模型對關鍵特徵的辨識與預測解釋能力，並驗證其在實務應用中所展現之潛在價值與貢獻。 ;With the widespread adoption of e-commerce and social media, online reviews have become an important reference for consumer purchasing decisions. However, the emergence and growing severity of fake reviews not only mislead consumers but also undermine market fairness and platform trust. In particular, with the rapid development of recommendation systems and generative artificial intelligence, the volume, spread, and realism of fake reviews have greatly increased, making traditional detection methods increasingly inadequate. The literature indicates that early detection methods mainly focused on linguistic features, with limited attention to reviewer behavior and model interpretability. Recent studies have shifted towards multi-feature integration and deep learning. However, the lack of interpretability inherent in deep learning models poses challenges for managerial trust, thereby reducing their practical adoption in real-world settings. In response, this study aims to develop a multimodal fake review detection model with high generalizability and interpretability by integrating review data from various platforms and multiple data types and incorporating a self-attention mechanism. The data sources include Yelp, Amazon, and synthetic fake reviews generated by ChatGPT to balance the dataset. The experiment is structured in three phases: first, evaluating various machine learning algorithms using textual features to establish a performance benchmark; second, incorporating reviewer behavioral data and product-related attributes to construct a multimodal framework aimed at boosting detection accuracy and cross-domain generalization; and third, implementing a self-attention mechanism to strengthen the model’s focus on critical features and enhance interpretability.
