科普閱讀能以生活化語言連結自然現象與科學概念，是培養國小學生科學素養、閱讀興趣與知識統整能力的重要途徑。然而，科普文本仍涉及抽象概念、因果推論與跨概念連結，學生若缺乏適切引導，容易停留於片段式理解。為回應此一教學需求，本研究建置一套以ORID焦點討論法為對話鷹架，並結合生成式人工智慧（Generative Artificial Intelligence， GenAI）與CoSci虛擬實驗室之國小科普閱讀學習同伴系統。系統透過GenAI聊天機器人檢索科普文本內容，並依據學生回應提供ORID四階段之結構化提問，引導學生進行客觀事實回顧、感受表達、概念詮釋與應用判斷。同時，系統於特定介入條件下整合CoSci虛擬實驗室，使學生能在聊書歷程中操作模擬實驗、觀察科學現象，並將實驗結果與文本概念相互連結，以提升其閱讀動機，並支持科學知識建構與閱讀理解。 為評估科普閱讀學習同伴系統及其結合虛擬實驗室之成效，本研究於臺灣北部某國小進行為期九週之準實驗研究，共154名國小六年級學生參與研究，最終納入128名具完整主要量化資料之學生作為分析樣本。研究採AB/BA平衡對抗設計，兩組學生皆使用以ORID焦點討論法為對話鷹架之GenAI科普閱讀學習同伴系統，差異在於不同實驗階段是否搭配CoSci虛擬實驗室。實驗期間，學生每週閱讀三篇科普文章，並與AI學習同伴進行結構化聊書互動；於結合CoSci之條件下，學生會操作虛擬實驗，觀察科學現象並連結文本概念。研究期間蒐集閱讀動機、科學知識與閱讀理解資料，並同步蒐集對話紀錄、ORID引導標籤、CoSci操作紀錄與系統日誌。本研究透過線性混合模型檢驗不同介入條件之成效，並運用K-means分群分析、滯後序列分析與認知網絡分析，探討不同閱讀歷程類型學生在學習成效與互動序列上的差異，最後輔以學生開放式問卷與教師訪談回饋補充量化結果之解釋。 研究結果顯示，在閱讀動機方面，結合CoSci虛擬實驗室之介入條件對學生整體閱讀動機具有顯著正向影響，主要表現在閱讀挑戰與閱讀認可兩個面向，顯示虛擬實驗室之融入有助於提升學生面對科普閱讀任務之挑戰意願與正向學習感受。在科學知識方面，學生表現隨實驗進行呈現提升趨勢，但CoSci虛擬實驗室之加入未產生顯著額外增益。在閱讀理解方面，學生表現主要受到施測時間影響，未因是否結合虛擬實驗室而呈現顯著差異。在對話歷程與互動結構方面，僅使用AI科普閱讀同伴時，ORID引導序列主要呈現由感受反應回到客觀事實，再推進至詮釋與決策之路徑；結合CoSci虛擬實驗室後，CoSci操作事件主要與客觀事實及現象觀察階段形成雙向轉換，顯示虛擬實驗室多嵌入於學生觀察現象與回顧文本事實之歷程中。認知網絡分析亦顯示，兩種介入條件下學生與AI聊書同伴之整體互動結構具有顯著差異。最後，依學生於系統中的平均對話次數與對話平均評分進行分群後，可辨識出不同閱讀歷程類型；其中，「高效互動群」與「密集互動型」學生在科學知識與閱讀理解表現上優於「有限互動型」學生，顯示互動次數並非唯一關鍵，學生回應品質、思考深度與互動穩定性可能更能反映其學習成效。 綜上所述，本研究顯示，以ORID為對話鷹架之GenAI科普閱讀學習同伴系統，能在國小科普閱讀教學中提供結構化聊書支持；而CoSci虛擬實驗室之整合，主要有助於提升學生閱讀動機，並改變學生與AI聊書同伴之對話歷程與互動結構。未來可依據學生互動品質與理解歷程，發展更具差異化與適應性的GenAI閱讀引導機制，並優化虛擬實驗與文本討論之銜接方式，以拓展其在科學閱讀與數位學習情境中的應用。;Popular science reading connects natural phenomena with scientific concepts through accessible language and serves as an important approach for cultivating elementary students’ scientific literacy， reading interest， and ability to integrate knowledge. However， popular science texts still involve abstract concepts， causal reasoning， and cross-concept connections. Without appropriate guidance， students may remain at a fragmented level of understanding. To address this instructional need， this study developed an elementary popular science reading learning companion system that uses the ORID Focused Conversation Method as a dialogue scaffold and integrates Generative Artificial Intelligence (GenAI) with the CoSci virtual laboratory. Through a GenAI chatbot， the system retrieves content from popular science texts and provides structured questions based on students’ responses across the four ORID stages: objective fact review， reflective response， interpretive meaning-making， and decisional application. Under specific intervention conditions， the system further integrates the CoSci virtual laboratory， allowing students to operate simulated experiments， observe scientific phenomena， and connect experimental results with textual concepts. The system aims to enhance reading motivation and support scientific knowledge construction and reading comprehension. To evaluate the effectiveness of the popular science reading learning companion system and its integration with the virtual laboratory， this study conducted a nine-week quasi-experimental study at an elementary school in northern Taiwan. A total of 154 sixth-grade students participated， and 128 students with complete primary quantitative data were included in the final analysis. An AB/BA counterbalanced design was adopted. Both groups used the GenAI popular science reading learning companion system scaffolded by the ORID Focused Conversation Method， while the difference between the two conditions was whether the CoSci virtual laboratory was incorporated during different experimental phases. During the experiment， students read three popular science articles each week and engaged in structured book-talk interactions with the AI learning companion. Under the CoSci-integrated condition， students operated virtual experiments， observed scientific phenomena， and connected their observations with textual concepts. Data on reading motivation， scientific knowledge， and reading comprehension were collected， along with dialogue records， ORID guidance labels， CoSci operation records， and system logs. Linear mixed models were used to examine the effects of different intervention conditions. K-means clustering， lag sequential analysis， and epistemic network analysis were further applied to explore differences in learning outcomes and interaction sequences among students with different reading process profiles. In addition， students’ open-ended questionnaire responses and teacher interview feedback were used to supplement the interpretation of the quantitative findings. The results showed that， in terms of reading motivation， the CoSci-integrated condition had a significant positive effect on students’ overall reading motivation， particularly in the dimensions of reading challenge and reading recognition. This finding suggests that integrating the virtual laboratory helped enhance students’ willingness to engage with challenging popular science reading tasks and strengthened their positive learning experiences. In terms of scientific knowledge， students’ performance showed an increasing trend as the experiment progressed; however， the integration of the CoSci virtual laboratory did not produce a significant additional effect. Regarding reading comprehension， students’ performance was mainly influenced by testing time， and no significant difference was found between conditions with and without the virtual laboratory. In terms of dialogue processes and interaction structures， when students used only the AI popular science reading companion， the ORID guidance sequence mainly followed a path from reflective responses back to objective facts， and then moved toward interpretation and decision-making. After the CoSci virtual laboratory was integrated， CoSci operation events mainly formed bidirectional transitions with the Objective stage， which involved factual review and phenomenon observation. This indicates that the virtual laboratory was primarily embedded in students’ processes of observing phenomena and reviewing textual facts. Epistemic network analysis also showed that the overall interaction structures between students and the AI book-talk companion differed significantly across the two intervention conditions. Finally， students were clustered based on their average number of dialogue turns and average dialogue quality scores in the system. Different reading process profiles were identified. Among them， students in the “Efficient Interaction Profile” and “Intensive Interaction Profile” outperformed those in the “Limited Interaction Profile” in scientific knowledge and reading comprehension. This finding indicates that the number of interactions was not the only key factor; rather， the quality of students’ responses， depth of thinking， and stability of interaction may better reflect their learning outcomes. In summary， this study demonstrates that a GenAI popular science reading learning companion system scaffolded by the ORID Focused Conversation Method can provide structured book-talk support in elementary popular science reading instruction. The integration of the CoSci virtual laboratory primarily helped improve students’ reading motivation and changed the dialogue processes and interaction structures between students and the AI book-talk companion. Future research may develop more differentiated and adaptive GenAI reading guidance mechanisms based on students’ interaction quality and comprehension processes. It may also further optimize the connection between virtual experiments and text-based discussion， thereby expanding the application of such systems in science reading and digital learning contexts.
