在課堂學習行為辨識中，過去研究多仰賴耗時的人工編碼或受限於特定訓練集的影 像辨識技術，難以彈性應對多變的實體課堂。在自我調節學習（SRL）的評估上，傳統 文獻多依賴主觀事後問卷，缺乏客觀行為數據佐證；而在課後複習方面，僅提供未經結 構化整理的完整錄影回放，常導致學習者資訊檢索效率低落。 為解決上述問題，本研究建構「AI-SRL 機制」，以三個模組輔助學習者落實自我調 節學習： 「行為觀察報告」模組對應 SRL 「預先思考階段」，此模組提供以下功能： (1) 運用視覺語言模型量化課堂行為並產出視覺化圖表，喚醒學習者行為自我覺察。(2) 系統自動產出個人化學習建議，引導學習者反思學習狀態。(3) 輔助學習者依據建議設 定次週學習目標，具體落實任務分析與策略擬定。 「多模態智慧筆記」模組對應 SRL 「表現與自我控制階段」，此模組提供以下功能：(1) 結構化圖文筆記：系統對齊授課內 容與板書生成知識節點，輔助課後快速複習。(2) 理解狀態動態標記：提供理解狀態標 籤與重點收藏功能，藉此檢視與監控知識吸收程度。(3) 筆記重點整理與反思：引導學 習者重新歸納學習重點並記錄疑惑，落實課後自我反思。 「課後自主學習測驗」模組對 應 SRL 「自我反思階段」，此模組提供以下功能：(1) 題型演練與 AI 即時評分：系統 提供多樣化題型，並透過人工智慧給予即時批改與具體修改建議。(2) 口說互動與語音 評分：提供語音互動練習與客觀量化評分，引導學習者針對發音弱點反覆修正。(3) 錯 題回顧：系統自動彙整錯題以引導再次練習，透過錯誤修正建立正確自我歸因，促使學 習者落實完整自我調節學習。 本研究以桃園市某補習班 23 名國中三年級英語學習者為研究對象，實驗共計 12 週。實驗分為四個階段進行：第一階段為無系統介入；第二階段提供「行為觀察報告」 模組；第三階段提供「行為觀察報告」與「多模態智慧筆記」模組；第四階段提供「行 為觀察報告」、「 多模態智慧筆記」與「課後自主學習測驗」模組。此外，本研究於各實 驗階段前後進行測驗，藉此提供不同模組對學習者學習行為與學習成效的影響。 研究結果顯示，AI-SRL機制可透過不同模組逐步促進SRL歷程，並帶來多層次學 習效益。首先，「多模態智慧筆記」導入後，筆試成績顯著優於僅提供行為回饋階段，顯 示結構化知識建構可有效提升課後複習效率與知識內化。其次，加入「課後自主學習測 驗」後，筆試與口說表現皆再度顯著提升，證實AI即時評分、錯誤回饋及語音互動能 有效促進形成性評量與持續修正。第三，在課堂學習行為方面，「行為觀察報告」導入 後，正向課堂行為即顯著增加，且於加入智慧筆記後持續提升，顯示客觀行為回饋能有 效提升學習者自我覺察與專注行為；第四階段正向行為雖趨於穩定，但飲食喝水行為顯 著增加，顯示學習者可能逐漸發展以短暫休息調節壓力之自主策略。此外，反思文本分 析顯示，加入AI自主測驗後，文本詞彙多樣性顯著提升，其中中等先備知識學習者進 步最為明顯，其表現已接近高先備知識組，顯示AI-SRL機制對中間能力群體具有顯著 支持效果。 II 本研究進一步證實，客觀量化之課堂學習行為具有預測學習成效的能力。逐步迴歸 分析顯示，「正向行為」、「目視書本／筆記」及「飲食/喝水」等課堂行為，以及系統中 的「行為觀察報告使用次數」與「知識節點查閱時間」，皆為顯著預測學習成效的重要指 標。此結果突破過去SRL研究主要依賴主觀量表之限制，首次證實多模態課堂行為資料 結合AI學習歷程紀錄，可作為預測學習表現的重要客觀證據。 綜合而言，本研究最大的學術貢獻在於提出並驗證一套整合多模態課堂行為辨識、 結構化智慧筆記、AI形成性評量與SRL理論之AI-SRL機制，不僅成功建立由「客觀 行為覺察—知識建構—形成性評量—自我反思」所構成的完整自我調節學習循環，更突 破傳統SRL研究長期依賴主觀自陳資料之限制，證實AI驅動的客觀行為分析能有效促 進課堂行為改變、提升筆試與口說學習成效，並建立可預測學習表現之客觀行為指標， 為AI支持自我調節學習、智慧教室與學習分析研究提供新的理論基礎與實證證據。;Regarding the recognition of classroom learning behaviors， previous studies mostly relied on time-consuming manual video coding or traditional image recognition technologies limited by specific training sets， making it difficult to flexibly identify diverse physical classroom behaviors. Secondly， in the evaluation of self-regulated learning (SRL)， past literature largely depended on post-event questionnaires or subjective memories for reflection. The lack of objective behavioral data often caused learners to fail in accurately grasping their true learning status. Lastly， regarding post-class review， since traditional notes struggle to completely restore the teaching context， previous research turned to providing full， unedited classroom video playbacks. However， such unstructured notes or video content compel learners to spend a significant amount of time and cognitive load on information retrieval. To address these issues， this study developed the "AI-SRL Mechanism，" comprising three modules to assist learners in implementing self-regulated learning. The "Learning Behavior Analytics Report" module utilizes multimodal vision-language models to objectively quantify classroom behaviors into visual charts， generating personalized learning suggestions to awaken learners′ behavioral self-awareness and assist them in setting learning goals. The "Smart Multimodal Notes" module aligns teacher lecture audio with blackboard images to generate structured knowledge nodes， providing dynamic comprehension status marking and guiding learners to summarize key points， thereby reducing cognitive load during post-class review. The "Self-Regulated Learning Assessment" module provides diverse practice questions， real-time AI grading， speaking interaction with voice scoring， and an incorrect question review mechanism， prompting learners to iteratively correct errors and establish correct self-attribution. The subjects of this study were 23 ninth-grade English as a Foreign Language (EFL) learners from a cram school in Taiwan， and the experiment lasted for 12 weeks. The experiment was conducted in four phases: the first phase involved no system intervention; the second phase provided the "Learning Behavior Analytics Report" module; the third phase added the "Smart Multimodal Notes" module; the fourth phase added the "Self-Regulated Learning Assessment" module. This study conducted tests before and after each phase and collected behavioral data to evaluate the impact of the mechanism. The research data showed that this mechanism could assist learners in improving written and speaking test achievements， driving positive changes in classroom behaviors， and reflecting these improvements in their reflective text performance. Following the order of module introduction: first， the visual "Learning Behavior Analytics Report" assisted learners in generating behavioral self-awareness， thereby increasing positive behaviors in the physical classroom. Secondly， after adding the "Smart Multimodal Notes" module， the structured IV knowledge assisted learners in repeated reviewing， which helped improve written test achievement; long-term reading of the multimodal notes also assisted learners in absorbing subject vocabulary， driving a significant increase in "textual lexical diversity" (MTLD). Next， after adding the "Self-Regulated Learning Assessment" module， real-time AI grading and error correction further drove significant growth in both written and speaking test achievements. Text analysis also revealed that learners with moderate prior knowledge showed improvements in lexical diversity during this phase， presenting no significant difference from the high-score group， indicating that the system mechanism provided a supportive effect for the mid-score learners. Furthermore， stepwise regression analysis confirmed that physical classroom behaviors such as "looking at books/notes" and "eating/drinking water" as a micro-break strategy， along with online system operations like the "number of accesses to the behavior observation report" and the "total time spent viewing knowledge nodes，" all possessed significant predictive power for learning achievement. Overall， this study demonstrated that through objective behavioral data feedback， structured notes， and real-time AI grading， learners could be effectively assisted in implementing a self-regulated learning cycle， thereby driving positive changes in classroom behaviors and significantly improving written and speaking test achievements.
