Conference Papers

In recent years, using AI to create an engaging vocabulary learning experience has been a prominent topic. Studies have shown AI can provide educational affordances for enhancing vocabulary learning. However, the appropriation of these affordances varies depending on teachers’ use. This paper presents a case study on six teachers appropriated an AI-powered vocabulary learning system, particularly focusing on the affordances of monitoring and regulation, and engaging co-construction enhanced by AI-enabled automatic feedback and recommendations. By examining teachers’ beliefs and knowledge of self-regulation and collaborative learning, the study details how they appropriated the affordances in their classes. The study provides suggestions for the design of AI in education and teachers’ professional development during the implementation of AI-supported learning system.

Peer feedback is a pedagogical strategy for peer learning. Despite recent indications of Large Language Models (LLMs) ' potential for content analysis, there is limited empirical exploration of their application in supporting the peer feedback process. This study enhances the analytical approach to peer feedback activities by employing state-of-the-art LLMs for automated peer feedback feature detection. This research critically compares three models—GPT-3.5 Turbo, Gemini 1.0 Pro, and Claude 3 Sonnet—to evaluate their effectiveness in automated peer feedback feature detection. The study involved 69 engineering students from a Singapore university participating in peer feedback activities on the online platform Miro. A total of 535 peer feedback instances were collected and human-coded for eleven features, resulting in a dataset of 5,885 labeled samples. These features included various cognitive and affective dimensions, elaboration, and specificity. The results indicate that GPT-3.5 Turbo is the most effective model, offering the best combination of performance and cost-effectiveness. Gemini 1.0 Pro also presents a viable option with its higher throughput and larger context window, making it particularly suitable for educational contexts with smaller sample sizes. Conversely, Claude 3 Sonnet, despite its larger context window, is less competitive due to higher costs and lower performance, and its lack of support for training and fine-tuning with researchers' data weakens its learning capabilities. This research contributes to the fields of AI in education and peer feedback by exploring the use of LLMs for automated analysis. It highlights the feasibility of employing and fine-tuning existing LLMs to support pedagogical design and evaluations from a process-oriented perspective.

This study explores the differences in students’ verbal behavior sequences between high-performing (HP) and low-performing (LP) groups in a computer-supported collaborative learning (CSCL) environment. Employing quantitative content analysis and Lag Sequential Analysis (LSA), this study analyzed the verbal interactions of these two groups. The findings reveal that HP groups frequently engaged in cycles of negotiation, clarity-seeking, and task coordination, leading to effective collaboration and problem-solving. In contrast, LP groups exhibited fragmented problem-solving approaches and frequent off-task behaviors. These insights highlight the importance of structured support and focused task management in enhancing collaborative learning outcomes. These findings suggest that educators should foster learning environments that promote continuous critical evaluation and seamless coordination to improve group performance.