IC-1463ChatGPT achieves 34.9% average F1 on zero-shot NER across 43 datasets spanning 9 domains

Wenxuan Zhou, Sheng Zhang, Yu Gu, Muhao Chen, Hoifung Poon

SourceUniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition

The paper evaluates ChatGPT (gpt-3.5-turbo-0301) in a zero-shot setting on the UniversalNER benchmark, which comprises 43 NER datasets across 9 domains (general, biomedical, clinical, STEM, programming, social media, law, finance, transportation). Using strict entity-level micro-F1, ChatGPT attains 34.9% average F1. Performance varies widely by domain, from 51.0% in general to 3.8% in transportation. The paper uses the prompting template from Ye et al. (2023) for NER.

Evidence
correlational
Key metric
34.9% average F1 (zero-shot, strict entity-level micro-F1, 43 datasets, 9 domains); domain range 3.8% (transport) to 51.0% (general)
Caveat
Evaluation uses up to 200,000 sampled passage-query pairs per dataset due to size constraints; strict F1 requires exact entity type and boundary match, which the authors note may underestimate zero-shot capability.
Model
ChatGPT
Related work
Ye et al. 2023 (GPT-3/GPT-3.5 capability analysis) [builds-on]
Related findings
IC-1464, IC-1465
Extraction
automatic-extraction