AAI_2025_Capstone_Chronicles_Combined

10

Figure 3 Prevalence of clustering techniques in recent educational data science publications (adapted from Le Quy et al., 2023)

.

each conversation. These extracted attributes are then fed into the DEC model for clustering. This technique effectively steers the unsupervised model by focusing its analysis on a specific dimension of the data, allowing for more targeted insights. For example, while a standard clustering of raw tutoring conversations might typically yield clusters based on mathematical subjects, clustering on an extracted "mood" facet would instead produce groups based on the emotional state of the student. The prompt used for the facet extraction directly controls the content resulting summaries, which dictates how conversations are grouped together based on semantic similarity in the embedding space (Tamkin et al., 2024). This capability is particularly valuable in an educational setting, as it allows teachers to gather highly targeted and nuanced insights from their students’ interactions with the tutoring chatbot. Grouping by various extracted facets can empower teachers to gain a deeper, more tailored understanding of their students’ conversations with the tutor. A full example of facet extraction can be found in Appendix B.

234