Search for: All records

Recent research has explored the use of Large Language Models (LLMs) to develop qualitative codebooks, mainly for inductive work with large datasets, where manual review is impractical. Although these efforts show promise, they often neglect the theoretical grounding essential to many types of qualitative analysis. This paper investigates the potential of GPT-4o to support theory-informed codebook development across two educational contexts. In the first study, we employ a three-step approach—drawing on Winne & Hadwin's and Zimmerman's Self-Regulated Learning (SRL) theories, think-aloud data, and human refinement—to evaluate GPT-4o's ability to generate high-quality, theory-aligned codebooks. Results indicate that GPT-4o can effectively leverage its knowledge base to identify SRL constructs reflected in student problem-solving behavior. In the second study, we extend this approach to a STEM game-based learning context guided by Hidi & Renninger's four-phase model of Interest Development. We compare four prompting strategies: no theories provided, theories named, full references given, and full-text theory papers supplied. Human evaluations show that naming the theory without including full references produced the most practical and usable codebook, while supplying full papers to the prompt enhanced theoretical alignment but reduced applicability. These findings suggest that GPT-4o can be a valuable partner in theory-driven qualitative research when grounded in well-established frameworks, but that attention to prompt design is required. Our results show that widely available foundation models—trained on large-scale open web and licensed datasets—can effectively distill established educational theories to support qualitative research and codebook development. The code for our codebook development process and all the employed prompts and codebooks produced by GPT are available for replication purposes at: https://osf.io/g3z4x. 

This study explores the integration of Large Language Models (LLMs), specifically GPT-4o, with machine learning (ML) to automatically detect struggle behaviors in a science exploration game. We evaluate GPT-4o's ability to analyze text replays, which convert raw gameplay logs into human-readable sequences of player actions, against existing machine-learned struggle detectors. The LLM and ML approaches achieve comparable performance but have complementary strengths. GPT-4o is effective at identifying struggle in instances where the human-engineered features used in traditional ML models fail to capture meaningful behavioral patterns, while ML detectors perform better at identifying struggle in tasks that involve numerical reasoning (e.g., frequent pauses) and in more complex tasks where structured features provide clearer signals. 

This study explores the potential of the large language model GPT-4 as an automated tool for qualitative data analysis by educational researchers, exploring which techniques are most successful for different types of constructs. Specifically, we assess three different prompt engineering strategies — Zero-shot, Few-shot, and Few-shot with contextual information — as well as the use of embeddings. We do so in the context of qualitatively coding three distinct educational datasets: Algebra I semi-personalized tutoring session transcripts, student observations in a game-based learning environment, and debugging behaviours in an introductory programming course. We evaluated the performance of each approach based on its inter-rater agreement with human coders and explored how different methods vary in effectiveness depending on a construct’s degree of clarity, concreteness, objectivity, granularity, and specificity. Our findings suggest that while GPT-4 can code a broad range of constructs, no single method consistently outperforms the others, and the selection of a particular method should be tailored to the specific properties of the construct and context being analyzed. We also found that GPT-4 has the most difficulty with the same constructs than human coders find more difficult to reach inter-rater reliability on. 

In past work, time management interventions involving prompts, alerts, and planning tools have successfully nudged students in online courses, leading to higher engagement and improved performance. However, few studies have investigated the effectiveness of these interventions over time, understanding if the effectiveness maintains or changes based on dosage (i.e., how often an intervention is provided). In the current study, we conducted a randomized controlled trial to test if the effect of a time management intervention changes over repeated use. Students at an online computer science course were randomly assigned to receive interventions based on two schedules (i.e., high-dosage vs. low-dosage). We ran a two-way mixed ANOVA, comparing students' assignment start time and performance across several weeks. Unexpectedly, we did not find a significant main effect from the use of the intervention, nor was there an interaction effect between the use of the intervention and week of the course. 

Research into student affect detection has historically relied on ground truth measures of emotion that utilize one of three sources of data: (1) self-report data, (2) classroom observations, or (3) sensor data that is retrospectively labeled. Although a few studies have compared sensor- and observation-based ap-proaches to student affective modeling, less work has explored the relationship between self-report and classroom observa-tions. In this study, we use both recurring self-reports (SR) and classroom observation (BROMP) to measure student emotion during a study involving middle school students interacting with a game-based learning environment for microbiology educa-tion. We use supervised machine learning to develop two sets of affect detectors corresponding to SR and BROMP-based measures of student emotion, respectively. We compare the two sets of detectors in terms of their most relevant features, as well as correlations of their output with measures of student learning and interest. Results show that highly predictive features in the SR detectors are different from those selected for BROMP-based detectors. The associations with interest and motivation measures show that while SR detectors captured underlying motivations, the BROMP detectors seemed to capture more in-the-moment information about the student申fs experience. Evi-dence suggests that there is benefit of using both sources of data to model different components of student affect.