DSpace community: 資訊管理研究所

人工智慧輔助資訊引導雛型系統對高中學生媒體識讀與認知卸責之影響初探;Effects of an AI-Assisted Information-Guiding Prototype System on Media Literacy and Cognitive Offloading among Senior High School Students: An Exploratory Study

title: 人工智慧輔助資訊引導雛型系統對高中學生媒體識讀與認知卸責之影響初探;Effects of an AI-Assisted Information-Guiding Prototype System on Media Literacy and Cognitive Offloading among Senior High School Students: An Exploratory Study abstract: 本研究探討在生成式人工智慧 (Generative AI) 普及的環境下，高中學生面對數位資訊之媒體識讀表現、查證意願與互動行為。學習者面對高認知負荷的資訊判讀任務時，可能依賴外部工具、直接接受現成資訊，呈現低投入或與後設認知懶惰 (Metacognitive Laziness)、認知卸載 (Cognitive Offloading) 相關的行為特徵。為回應此教育困境，本研究建置結合人工智慧偵探情境與數位鷹架的 AI 輔助資訊引導雛型系統，作為高中媒體識讀教學與資訊判讀練習的輔助工具。本系統以鷹架理論 (Scaffolding Theory) 為基礎，透過媒體案件、熱區採證、分級提示與結案反思，引導學生觀察訊息、拆解線索、自主思考、主動求證與反思判斷。系統不直接揭示答案，而是依學生停留時間與錯誤點擊開放提示，學生主動點擊後，再由 AI 模組提供啟發式引導。另建構五階段思維層次分析架構 (被動接收、訊息拆解、自主思考、主動求證、初步內化跡象)，作為系統任務、歷程編碼與互動分析的操作性架構。本研究採單一組別前後測設計，進行短期介入，透過媒體識讀主觀態度量表、客觀能力測驗、互動日誌與典型個案資料進行評估。結果顯示，學生的媒體識讀主觀態度總分於介入後顯著提升；然而，客觀能力前後測未達顯著差異，顯示本研究尚未觀察到短期介入對客觀判讀能力或知識遷移的明顯影響。分析發現，部分學生呈現低投入或可能與認知卸載相關的特徵，如過快作答、反思不足或查證不充分；亦有學生展現訊息拆解、主動求證與較完整推理等高階行為。本研究系統與五階段架構可作為觀察高中學生數位資訊判讀歷程及設計媒體識讀教材的參考。惟結果限於短期介入的特定系統情境，不宜推論學生已形成穩定媒體識讀素養或認知卸載行為的長期改變。;This study examined senior high school students’ media-literacy performance， willingness to verify information， and interaction behaviors when encountering digital information in an environment where generative artificial intelligence (AI) has become widespread. When facing information-verification tasks with high cognitive demands， learners may rely on external tools or directly accept readily available information， displaying low engagement or behavioral characteristics associated with metacognitive laziness and cognitive offloading. To address this educational challenge， an AI-assisted information-guiding prototype system integrating an AI-detective scenario with digital scaffolding was developed to support senior high school media-literacy instruction and information-verification practice. Grounded in scaffolding theory， the system incorporated media cases， hotspot-based evidence collection， graduated prompts， and case-report reflection. These functions guided students to observe messages， deconstruct clues， think independently， actively verify information， and reflect on their judgments. Rather than directly revealing the answers， the system made prompts available according to students’ dwell time and erroneous clicks. After students actively selected the prompt function， the AI module provided heuristic guidance. In addition， a five-level framework for analyzing students’ thinking processes—comprising Passive Reception， Message Deconstruction， Independent Thinking， Active Verification， and Preliminary Signs of Internalization—was developed as an operational framework for system-task design， process coding， and interaction analysis. A one-group pretest–posttest design was adopted for a short-term intervention. The evaluation used a media-literacy attitude scale， an objective media-literacy assessment， interaction logs， and typical-case data. The results showed a significant improvement in students’ overall media-literacy attitude scores after the intervention. However， no significant pretest–posttest difference was found in objective performance， suggesting that the study did not observe clear short-term improvement in students’ information-evaluation ability or knowledge transfer. The analysis further showed that some students displayed low engagement or behavioral characteristics potentially related to cognitive offloading， such as responding too quickly， providing insufficient reflection， or conducting insufficient verification. Other students demonstrated higher-level behaviors， including message deconstruction， active verification， and more complete reasoning. Overall， the prototype system and the five-level framework may serve as an initial reference for examining senior high school students’ digital information-verification processes and designing media-literacy materials. However， the findings should be interpreted only within the specific system context of the short-term intervention and should not be taken as evidence that students had developed stable media-literacy competencies or experienced long-term changes in cognitive offloading behaviors.

AI 輔助經營模式創新視覺化平台：動態知識累積與人機協作之應用研究;AI 輔助經營模式創新視覺化平台：動態知識累積與人機協作之應用研究

title: AI 輔助經營模式創新視覺化平台：動態知識累積與人機協作之應用研究;AI 輔助經營模式創新視覺化平台：動態知識累積與人機協作之應用研究 abstract: 近年來，隨著數位轉型與生成式人工智慧技術的發展，企業必須持續檢視並調整經營模式，以因應市場需求與競爭環境的變化。然而，現有經營模式分析工具多著重於靜態架構呈現，缺乏對元件間因果關係的視覺化分析，亦未能整合人機協作與知識持續累積機制，難以滿足企業於經營模式分析與決策規劃的實際需求。為解決上述問題，本研究建構一套 AI 輔助經營模式創新視覺化平台。平台以因果脈絡圖呈現經營模式元件及其關係，並整合大型語言模型 (Large Language Models， LLM )、檢索增強生成 (Retrieval-Augmented Generation， RAG )、人機協作 (Human-in-the-Loop， HITL ) 及知識動態更新機制，使用者可以檢視及修改 AI 建議，確認後的內容則作為知識庫後續更新的資料。研究採混合研究方法，透過使用者實驗、Likert 量表、操作紀錄分析及半結構式訪談進行初步評估。實驗結果顯示，本平台支援使用者進行經營模式分析與視覺化思考，並有助於促進思考擴展與因果關係建構。在團體情境中，平台亦能協助成員釐清觀點、整合討論內容，並有助於形成共同理解。此外，知識回寫結果初步顯示，平台具有跨使用者累積與再利用分析知識的可能性，但其知識品質與實際支援效果仍需後續研究進一步驗證。本研究成果可作為企業策略分析與團隊決策的輔助參考，並為 AI 輔助經營模式分析、人機協作及知識動態更新機制的設計與應用，提供初步的實證觀察。;As digital transformation and generative artificial intelligence continue to advance， firms must adapt their business models to changing market demands and competitive environments. Existing business model analysis tools mainly present static structures， provide limited support for visualizing causal relationships among business model components， and seldom integrate human–AI collaboration with mechanisms for continuous knowledge accumulation. This study develops an AI-assisted visual platform for business model innovation. The platform uses causal maps to visualize business model components and their relationships while integrating large language models (LLM)， retrieval-augmented generation (RAG)， human-in-the-loop (HITL)， and a dynamic knowledge updating mechanism. Users can inspect and revise AI-generated suggestions， and confirmed content is stored in the knowledge base for future reuse. A mixed-methods approach was adopted. The platform was evaluated using Likert-scale questionnaires， operation logs， and semi-structured interviews. The results indicate that the platform supported business model analysis by helping participants extend ideas and construct causal relationships. In group settings， participants reported that the visual representation facilitated communication， helped organize discussions， and promoted shared understanding. The knowledge write-back records showed overlaps between components contributed by earlier and later participants， suggesting the potential for cross-user knowledge accumulation and reuse， although direct retrieval from the knowledge base was not verified. Further validation through larger samples， source tracking， and expert evaluation is still required. The platform may serve as a decision-support tool for strategic analysis and team discussion. This study provides initial empirical insights into the design of AI-assisted business model analysis， human–AI collaboration， and dynamic knowledge updating mechanisms.

基於共同基金持股序列之下一籃推薦模型研究：DNNTSP 與外部特徵融合之實證分析;Next-Basket Recommendation for Mutual Fund Holdings Prediction: An Empirical Analysis of DNNTSP with External Feature Fusion

title: 基於共同基金持股序列之下一籃推薦模型研究：DNNTSP 與外部特徵融合之實證分析;Next-Basket Recommendation for Mutual Fund Holdings Prediction: An Empirical Analysis of DNNTSP with External Feature Fusion abstract: 在現代金融市場中，基金持股資料提供了觀察機構投資人投資組合配置與跨期調整行為的重要依據。由於共同基金持股並非單一標的選擇，而是由多個證券組成之投資組合，因此本研究將基金下一期持股組合預測問題轉換為下一籃推薦(Next-Basket Recommendation， NBR)任務，並探討深度學習模型在基金持股序列預測上的應用效果。本研究使用 CRSP 共同基金持股資料，建構 2014 年至 2024 年之基金季度持股資料集。研究中將基金投資組合代碼視為使用者，將證券識別碼視為項目，並將同一基金於同一季度之持股組合依時間順序建立為基金持股序列。經前處理後，本研究資料包含17，203 檔基金、56，214 個非BOND持股標的，以及 457，970 個基金季度持股組合。模型架構採用 DNNTSP 作為核心，並設計四種模型版本進行比較，分別為 ID-only、ID + Stock Feature、ID + User/Fund Feature，以及 ID + Stock + User/Fund Feature。股票層級特徵包含中期動能、短期反轉、市值規模、流動性不足程度與波動度；基金層級特徵則包含基金 NAV 變化率、基金波動度、基金規模、費用率與週轉率等資訊。實證結果顯示，四種 DNNTSP 模型在 Precision@K 與 Recall@K 上表現相當接近。相較於 ID-only 模型，加入股票層級特徵或基金層級特徵後僅帶來小幅變化，而同時加入兩類特徵後並未進一步提升整體表現。進一步分析發現，基金持股組合具有高度跨期延續性，前一期持股基準模型在測試集上取得相當高的表現，顯示上一期持股組合本身即為預測下一期持股組合的重要訊號。此外，部分持股紀錄無法對應至股票特徵資料表、基金層級特徵粒度較粗，以及歷史持股序列與項目嵌入向量已能捕捉主要共現關係，皆可能限制外部特徵對整體預測表現的邊際貢獻。本研究結果顯示，基金持股組合預測是一個高度受持股延續性主導的下一籃推薦任務。外部股票層級特徵與基金層級特徵雖可提供額外資訊，但在整體 Precision@K 與 Recall@K 評估下，其改善幅度有限。本研究結果可作為後續金融下一籃推薦研究之參考，並指出未來可進一步針對新增持股、高換手率基金、低持股延續性樣本，或更完整的特徵覆蓋率進行延伸分析。;Mutual fund holdings data provide a direct view of institutional investors’ portfolio allocation and adjustment behavior. Since fund holdings consist of multiple securities rather than a single asset， this study formulates next-period fund holdings prediction as a Next-Basket Recommendation (NBR) task. Using CRSP mutual fund holdings data from 2014 to 2024， this study constructs quarterly holdings sequences by treating funds as users， CUSIPs as items， and each fund-quarter portfolio as a basket. After preprocessing， the dataset contains 17，203 funds， 56，214 non-BOND CUSIPs， and 457，970 fund-quarter baskets. The study applies DNNTSP and compares four model variants: ID-only， ID + Stock Feature， ID + User/Fund Feature， and ID + Stock + User/Fund Feature. Stock-level features include momentum， reversal， size， illiquidity， and volatility， while fund-level features include NAV change， volatility， fund size， expense ratio， and turnover ratio. The empirical results show that all four DNNTSP variants achieve similar Precision@K and Recall@K performance. Adding stock-level or fund-level features leads only to marginal changes， and combining both feature types does not further improve overall performance. Further analysis shows that fund holdings exhibit strong persistence across periods. The last-quarter baseline performs strongly on the test set， indicating that prior holdings are already a dominant predictive signal. Overall， this study finds that fund holdings prediction is highly driven by holdings persistence. Although external stock and fund features provide additional information， their marginal contribution under overall Precision@K and Recall@K evaluation is limited. These findings provide an empirical reference for future financial NBR research， especially for analyses focusing on new holdings， high-turnover funds， low-persistence samples， or improved feature coverage.

MMRCL: Medoid-Guided Multi-Label Contrastive Learning with Swapping Reconstruction for Video Classification;MMRCL: Medoid-Guided Multi-Label Contrastive Learning with Swapping Reconstruction for Video Classification

title: MMRCL: Medoid-Guided Multi-Label Contrastive Learning with Swapping Reconstruction for Video Classification;MMRCL: Medoid-Guided Multi-Label Contrastive Learning with Swapping Reconstruction for Video Classification abstract: 隨著影片資料在社群影音、線上教育、智慧監控與智慧城市等應用中的快速成長，影片表徵學習已成為影片理解中的重要研究議題。相較於傳統單一類別影片分類，多標籤影片分類更能描述真實影片中同時包含物件、場景、動作、事件與屬性等多層次語意的特性。近年來，對比學習（Contrastive Learning， CL）已廣泛應用於表徵學習，其目標是拉近語意相似樣本並推遠語意不同樣本。然而，多標籤影片之間常存在部分語意重疊，使傳統正樣本與負樣本的二元關係難以完整描述細緻的語意關聯。為解決上述問題，本研究提出 Medoid-Guided Multi-Label Contrastive Learning with Swapping Reconstruction（MMRCL），一個針對多標籤影片表徵學習所設計的結構化對比學習框架。MMRCL 由 V2P、Swapping、V2V 與 P2M 四個模組組成，分別用於全域語意對齊、共同與差異語意分解、分支層級關係排序，以及原型與代表性樣本校準。本研究於 HVU mini 與 Charades 兩個多標籤影片資料集上進行實驗，結果顯示 MMRCL 相較於多數基準方法取得較佳表現。進一步的表示空間分析、消融實驗、參數敏感度分析與案例研究亦驗證 MMRCL 能學習更符合多標籤語意關係的影片表示，並展現各模組的互補效果。;With the rapid growth of video data in social media， online education， smart surveillance， and smart city applications， video representation learning has become an important research topic in video understanding. Compared with conventional single-label video classification， multi-label video classification better describes real-world videos containing objects， scenes， actions， events， and attributes simultaneously. Recently， Contrastive Learning (CL) has been widely used in representation learning to pull semantically similar samples closer and push dissimilar samples apart. However， multi-label videos often contain partial semantic overlap， making binary positive and negative relationships insufficient for modeling fine-grained semantic relations. To address this issue， this study proposes Medoid-Guided Multi-Label Contrastive Learning with Swapping Reconstruction (MMRCL)， a structured contrastive learning framework for multi-label video representation learning. MMRCL consists of four modules， V2P， Swapping， V2V， and P2M， which respectively perform global semantic alignment， shared and distinct semantic decomposition， branch-level relational ranking， and prototype-to-medoid calibration. Experiments on HVU mini and Charades show that MMRCL outperforms several baseline methods. Further representation space analysis， ablation studies， sensitivity analysis， and case studies verify that MMRCL learns video representations that better reflect multi-label semantic relationships and demonstrate the complementary effects of the proposed modules.