popularity-accuracy-correlation-popqa-0.4-0.8
IN premise — summaries/2026/08/24/mallen-2023-when-not-to-trust-sA-appendix.md
Created 2026-08-25T02:58:12+00:00
On POPQA, log(entity popularity) and accuracy show positive correlations of 0.4–0.8 for most of the 16 relation types
Summary
The model's accuracy on fact questions is strongly tied to how well-known the entity is, and this pattern holds broadly across most of the question types tested rather than being limited to a few categories. In practical terms, the system can expect reliable performance on famous entities but should treat results on obscure ones with much more skepticism, since the model's knowledge is unevenly distributed across the entity space.