A Confound-Annotated Multi-Institution Curriculum Corpus with Parsed Prerequisite Logic
Description
This dataset is a machine-readable curriculum corpus assembled from the published catalogs of thirteen universities, centered on the Arabian Gulf (United Arab Emirates, Saudi Arabia, Qatar, Kuwait, Bahrain) and extended by the American University of Beirut and the California Institute of Technology as comparators. It comprises 32 catalog editions, 51,858 course records, and 39,709 prerequisite relations, released as two CSV tables (courses.csv, prereq_edges.csv) and a JSON summary (datasets.json), together with a data dictionary and a reference implementation of the analysis code. The dataset has three distinguishing properties. First, prerequisites are parsed into conjunctive-normal form, so the alternatives a catalog expresses with the word "or" are preserved as boolean structure rather than flattened into a list of mandatory codes. Second, one institution is covered by an eleven-edition panel spanning 2015-2016 to 2025-2026, which supports course-level longitudinal analysis without recourse to web archives. Third, every record is annotated with measurement-confound fields (notation drift, selective disclosure, subject-code renumbering, and prerequisite-operator ambiguity), each exposed as a filterable predicate so that diachronic and cross-institution analyses can control them rather than fall prey to them. The data support research on curricular structure, prerequisite-network complexity, academic advising and course planning, accreditation analysis, and the natural-language processing of course descriptions, and they cover a region that has been largely absent from existing curriculum datasets.
Files
Steps to reproduce
The corpus was built in four stages, all implemented in released code. (1) Harvesting: catalog content was obtained from each institution's published source (server-rendered HTML, Banner student-information pages, application-framework JSON, institutional bulletins, and archival catalog PDFs); documents whose format resisted deterministic parsing were read with large-language-model vision extraction and verified against the source. (2) Parsing and normalization: each prerequisite expression was read as a boolean expression and distributed into conjunctive-normal form, then serialized as prerequisite edges; malformed markup (omitted separators, run-on codes, nested conjunctions) was repaired under rules that preserve every course code. (3) Longitudinal reconciliation: course codes were resolved across editions into canonical lineages, checked against a coexistence constraint. (4) Validation: a thirty-three-assertion suite reconciles the parsed set against the catalog index and confirms identity across the JSON, CSV, and database representations, and a pre-registered sampled correctness audit reports per-field error rates. To reproduce the per-institution summary and the disjunctive-complexity results from the released files, run examples/load_and_summarize.py and code/complexity_analysis.py.
Institutions
- United Arab Emirates UniversityAbu Dhabi, Al Ain