LLM interest-extraction pipeline
ICALT 2026 study
The pipeline behind Study 2 of my dissertation, which extracted students’ interests, with explanations, from their browsing history and social-media posts.
Grounding: checking what models assert about people
A unit-level check of every claim a model makes about a student — is it grounded in the student’s own material, or inferred?
In two studies, I checked every claim a model made about a student, one unit at a time: is it grounded in the student’s own material, or inferred?
Study 2 extractor: 384 explanations · 1,130 units
- 625 grounded
- 505 inferred (44.7%)
Study 1 tutor (dissertation Appendix E) · 57 units
- 23 grounded (40.4%)
- 34 inferred (59.6%)
“Inferred” means the student’s material doesn’t show it, not that it is false: an inferred claim may still be true. The check compares claims with what the student provided; it does not test truth.
Only grounded units count toward Depth and Richness in my interest-identification work.
The same convention on this site
The legend at the bottom of every page uses the same idea: a solid underline for something shown in the source, a dotted underline for an inference. The planned Ask and Life & You pages will mark their claims this way.
Publications
- (2026). Identifying Student Interests for Interest-Based Learning: A Comparison of Interest Extraction from Browsing History and Social Media Posts. IEEE ICALT 2026. Full paper.