Licensing & credits

nlpgraph’s code is MIT. It is a clean-room “modern NLP in TypeScript” build — not a port of Stanford CoreNLP (GPL-3.0). It drives a pre-trained biaffine ONNX model and adds the pure-TS layer around it.

Licensing at a glance

• This library’s code: MIT.
• The vendored model (ghotriw/...-en-ewt) is published by its author under MIT and was trained only on UD English-EWT (CC BY-SA 4.0) — commercial use is permitted with attribution; the share-alike term may attach to the model artifact, not to this MIT code.
• ⚠️ Avoid the ...-en-ewt-gum variant for commercial use: GUM is CC BY-NC-SA 4.0 (non-commercial). We deliberately default to the EWT-only model.

Not legal advice. Whether trained weights inherit their training data’s license is legally unsettled; consult counsel for high-stakes commercial deployment.

Credits

This project stands on the shoulders of others, with gratitude:

• Parser model: ghotriw/deberta-v3-small-biaffine-dep-pos-en-ewt — all parsing accuracy is due to this model and its author.
• Encoder: microsoft/deberta-v3-small (MIT).
• Training data: Universal Dependencies English-EWT (CC BY-SA 4.0).
• Method: Dozat & Manning, Deep Biaffine Attention for Neural Dependency Parsing (ICLR 2017); architecture inspiration from spaCy and Stanza.
• Runtime: ONNX Runtime (MIT) and Transformers.js (Apache-2.0) for sub-word tokenization.
• Tokenizer design: rule ordering modeled on spaCy’s tokenizer and sentence rules on pySBD; the regulatory recognizers are our own.
• Docs & demo hosting: Astro Starlight with Pagefind search, on Cloudflare Pages + R2, auto-deployed from GitHub Actions.

Not affiliated with Stanford CoreNLP.