本論文包含兩個部分，旨在從實證與文獻探勘角度，探索社群媒體中內容特徵對使用者參與行為學術研究趨勢的影響。第一部分聚焦於Instagram平台，分析網紅業配文中的文字與視覺特徵如何影響貼文參與度;研究運用LIWC工具萃取語言特徵，並結合Show & Tell電腦視覺模型擷取圖像特徵，透過迴歸模型探討其對按讚率與留言率的影響，同時比較電腦模型與人工標註在視覺特徵辨識上的表現；結果顯示，在為維持品牌知名度的貼文中，品牌名稱與認知語言能提升參與度，過度個人化與情緒化則可能抑制互動；而在為宣傳品牌活動的貼文中，第一人稱用語與產品圖片則更能促進互動。另外，在此階段進一步探索潛在的關鍵視覺元素—凝視方向對參與度的影響；儘管結果顯示注視方向並不會獨自對參與度產生顯著影響，但它與第一人稱單數代詞的交互作用是值得注意的，尤其是在臉佔圖中比例大的商品類別。第二部分則應用主題模型技術，系統性回顧Scopus與Web of Science資料庫中有關直播與用戶生成影片的文獻，透過LDA、NMF、Top2Vec與BERTopic四種主題模型，分析並比較文獻聚焦的演變，結果顯示Top2Vec對於辨識細緻主題具有優勢，揭示直播和用戶生成影片兩個數位環境的多層性質。;The thesis comprises two sections aimed at investigating the impact of content characteristics on user engagement behavior and scholarly research trends in social media from both empirical and literature-mining perspectives. Section 1 focuses on the Instagram platform, examining how textual and visual features in influencer-sponsored posts affect post engagement. Linguistic features were extracted using the Linguistic Inquiry and Word Count (LIWC) tool, while visual features were derived via the Show & Tell computer vision model. Regression models were employed to assess the influence of these features on like and comment ratios. Additionally, the performance of automated models and manual annotation in identifying visual attributes was compared. The results indicate that brand name mention and cognitive language enhance engagement in posts that maintain brand awareness, whereas overly personalized or emotional language may suppress interaction. In contrast, posts that promote brand-specific events benefit more from first-person language and sponsored-product imagery in fostering user engagement. This stage also delves into the potential role of more nuanced visual elements—specifically, the influence of gaze direction on user engagement. Although the results indicate that gaze direction alone does not significantly affect engagement, its interaction with first-person singular pronouns is noteworthy, particularly within product categories characterized by a high facial prominence in images. Section 2 applies topic modeling techniques to systematically review the academic literature on live streaming and user-generated video content retrieved from the Scopus and Web of Science databases. Four topic modeling approaches—Latent Dirichlet Allocation (LDA), Non-negative Matrix Factorization (NMF), Top2Vec, and BERTopic—were employed to analyze and compare the evolution of scholarly focus in this domain. The findings reveal that Top2Vec offers superior granularity in topic identification, thereby uncovering the multifaceted nature of live streaming and user-generated video environments.
