BERT learns Chinese names, circa 2018

What it does

This repo is a thin orchestration layer around Google’s original TensorFlow BERT implementation. It adapts the BERT-Base, Chinese checkpoint to a named entity recognition task using BIO labels on the classic People’s Daily dataset, reporting an F1 of roughly 0.966 after 100 epochs. Think of it less as a framework and more as a working recipe from the early BERT era.

The interesting bit

Its value today is archaeological: it captures the exact moment practitioners started applying BERT’s Chinese weights to the canonical People’s Daily NER benchmark. The author is admirably candid, labeling the work an “attempt” and posting a redirect to a newer ALBERT-based successor.

Key highlights

• Fine-tunes BERT-Base, Chinese on the standard People’s Daily NER corpus using BIO tags
• README reports ~96.6% F1, precision, and recall after 100 epochs
• Requires manually integrating Google’s BERT-TF source and the Chinese checkpoint
• Author now maintains albert-chinese-ner as the recommended follow-up

Caveats

• The repo is essentially glue code; most heavy lifting is done by the original Google BERT implementation
• Setup is manual and not self-contained
• The author explicitly suggests using the newer ALBERT version instead

Verdict

Curious developers studying early Chinese BERT fine-tuning patterns or the People’s Daily baseline will find a concise reference here; everyone else should probably follow the author’s own advice and head to the ALBERT successor.