Novice programmers frequently struggle with compilation errors, and providing precise and targeted feedback in the IDE still remains a challenge. Although prior studies have established a foundation for error taxonomies, existing frameworks have been constrained by limited category granularity or the reliance on expert annotation. We address these limitations by introducing a novel hybrid labeling methodology based on the large-scale IBM Code Net dataset. Our approach utilizes Large Language Models (LLMs) to significantly reduce human annotation efforts while providing comprehensive categories that enable our classifier to distinguish between unintentional mistakes and knowledge gaps. This granular classification is crucial for downstream tasks, allowing feedback to focus on the root cause of compile errors. Subsequently, we evaluated several pretrained language models to identify the most effective classifier architecture. Ultimately, this work lays the foundation for intelligent tutoring systems capable of delivering precise, targeted, and real-time interaction with novice programmers.

Novice programmers often get stuck while programming and debugging. They would turn to AI tools like ChatGPT for help when instructors or teaching assistants (TA) are unavailable. However, conventional AI tools give away answers too easily, which undermines the learning process. This project aims to solve the lack of guided support for novice programmers by developing a learning-centered intelligent tutoring system. This system mimics human TAs by asking guiding questions to build problem solving skills and deeper understanding. Integrated into an online IDE, this AI tutor leverages error categories to deliver context-sensitive feedback in real time to keep students engaged. To ensure accessibility and control, the system uses smaller, locally deployable models rather than relying solely on proprietary services. By avoiding overreliance on AI-generated solutions, the system aims to sustain engagement, reduce frustration, and foster independent learning.

The multi-modal approach for conducting education research is gaining traction among the engineering and computing education disciplines. This panel aims to introduce this novel and emergent approach to answering important research questions related to engineering and computing education. Additionally, the panelists will share their personal experiences on how they have incorporated a multi-modal approach for their research. The panelists will also share the best practices and common mistakes researchers could make using this methodology.

For novice students, learning programming is hard, hence, inducing a variety of emotions. According to literature, two of the most commonly occurring emotions that students experience while learning programming are frustration and confusion. Although these emotions are momentary, they may have long-term effects on students' motivation, performance, and retention in computing. In this study, we aim to discuss challenges that students report when they feel frustration and confusion while working on the rainfall problem. This problem has been used extensively in literature to understand students' problem-solving skills. However, little is understood about how students react emotionally when they work on this problem. We recruited twenty-eight students who took CS1 during Fall 2022. They worked on the problem for twenty minutes while we collected their biometrics, clickstream, and keystroke data. A retrospective think-aloud interview was conducted soon after the task, where participants elaborated on their emotional experiences while watching the video replay of their programming task. We analyzed interview data using qualitative content analysis and triangulated these findings with biometric, clickstream, and keystroke data. Some of the challenges that trigger confusion and frustration are getting unexpected output, inability to resolve compilation and logical issues, forgetting syntax, and conceptual misunderstanding. Moreover, we found an alignment between the different data sources to provide a near real-time view of emotional experiences. Instructors may use the findings of this study to design interventions that support problem-solving, such as problem-based teaching, using interactive programming tools, and tracing for debugging.

This overarching study aims to establish the validity of the Programming Emotions Questionnaire (PEQ). The PEQ is an instrument that could be used for assessing students' emotions in a computer programming class, lab, or test. It is derived from the achievement emotions questionnaire (AEQ) that is grounded in the control-value theory of achievement emotions. Since AEQ has numerous psychometric challenges (e.g., some items are not atomic), we make significant changes to revise it, hence, we will be validating the questionnaire from scratch. In this poster we discuss the preliminary findings of our validation study, specifically, findings from the content validity phase. For content validity, we had discussions with psychometric experts and computer science experts. We also present findings of a pilot study of the target population, that is, CS1 students. The result of content validity suggests that there is a need to develop the PEQ that significantly modifies the AEQ. The immediate next steps include conducting steps to ensure construct and criterion validity, and its reliability.