Research
Selected Publications
To view all publications, see Google Scholar.
Cognitive Agent Compilation for Explicit Problem Solver Modeling
Moon, H., Rosé, C., & Stamper, J. (2026). Artificial Intelligence in Education (AIED 2026), Blue Sky.
A position paper presented in the Blue Sky track at AIED 2026, on a more precise way to identify knowledge components. LLMs have gotten better at identifying knowledge components, and the paper proposes how to break them down reliably into finer-grained units.
Using Large Multimodal Models to Extract Knowledge Components for Knowledge Tracing from Multimedia Question Information
Moon, H., Davis, R. L., Neshaei, S. P., & Dillenbourg, P. (2025). Educational Data Mining (EDM 2025).
Presented at Educational Data Mining 2025, this paper shows how large multimodal models (LLMs that can also take images and other kinds of input) can efficiently extract knowledge components from learning content. Labeling learning content used to be a labor-intensive bottleneck, but around 2024 advances in LLMs made it possible to extract these labels automatically at good quality.
Evaluating the Knowledge Dependency of Questions
Moon, H.*, Yang, Y.*, Yu, H., Lee, S., Jeong, M., Park, J., Shin, J., Kim, M., & Choi, S. (2022). Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP 2022), pp. 10512–10526.
Published at EMNLP 2022, this paper proposes a way to measure question quality, starting from the idea that an educational question should be judged by whether it can tell who has a given piece of knowledge. At the time, generated educational questions were mostly evaluated by semantic similarity to human-written references. I proposed a different criterion: a learner without the knowledge should get the question wrong, and a learner with it should get it right.
Dialogue Summaries as Dialogue States (DS2), Template-Guided Summarization for Few-shot Dialogue State Tracking
Shin, J.*, Yu, H.*, Moon, H.*, Madotto, A., & Park, J. (2022). Findings of the Association for Computational Linguistics: ACL 2022, pp. 3824–3846.
Published in Findings of ACL 2022, this was my first full conference paper, as a co-first author. It worked with the LLMs of that time, which were far more rigid than today's instruction-tuned ChatGPT. The paper shows that a model fine-tuned for dialogue summarization can be adapted to identifying specific fields, by guiding its summaries with suitable templates.
* Equal contribution