The community model that shamed SpaCy into going native

What it does

Trains SpaCy v2 models on Chinese text so you get tokenization, dependency parsing, and NER out of the box. It wraps the OntoNotes 5.0 corpus into SpaCy’s pipeline format, then adds a convenience alias (zh) so downstream tools like Rasa NLU can find it.

The interesting bit

The README’s opening line is unusually honest: SpaCy’s official Chinese model now exists, it “referenced this project” and shares the same features, so this repo’s mission is “complete.” That’s a graceful exit for what was essentially a hostage negotiation with an open-source ecosystem — build the missing piece, prove demand, then watch the maintainers absorb it.

Key highlights

• Ships binary model files; install via pip then spacy link for alias convenience
• Includes Binder-hosted Jupyter demo for browser-based testing
• Trained on OntoNotes 5.0 (free for organizational users, LDC licensing hassle required)
• Provides accelerated download mirrors for China-based users
• Web demo in test.py serves dependency parse visualizations locally

Caveats

• Maintenance mode only: bug fixes, no new features; official SpaCy zh models are the future
• Several Doc attributes are broken or questionable for Chinese: pos_, is_stop, shape_, is_alpha, and vector quality are all flagged in the TODO
• Requires manual OntoNotes 5.0 procurement through LDC — no bundled data

Verdict

Use this if you’re pinned to an older SpaCy v2 setup or need the exact training pipeline documented in workflow.md. Otherwise, migrate to SpaCy’s official Chinese models and send a thank-you note to howl-anderson for making them exist.