Semuhi Sinanoglu with Lucan Way, Steven Levitsky & Yashar Talebirad (2025). “Tracking autocratization transparently: an event-level dataset (DERTT).” Working paper. Includes an original dataset.
An LLM-extracted, reproducible event-level dataset of democratic erosion: 76,792 unique incidents across 197 countries, 2000 to 2023.
We present a framework for extracting structured autocratization events from human rights reports using keyword filtering and LLM-based extraction. The pipeline processes 9,677 country-year reports from the U.S. State Department, Freedom House, and Amnesty International (2000-2023) using an open-weight model with conservative, precision-first prompts and a two-threshold semantic deduplication layer. Applied at scale, it produces DERTT, a dataset of 76,792 unique autocratization incidents across 197 countries and 6 event types, each traceable to specific sentences in the source reports.