What it does

This is a fork of the pre-1.0 Rasa NLU library with Chinese language support bolted on. It takes Chinese text, tokenizes it with Jieba, extracts named entities and intent using MITIE (and optionally scikit-learn for classification), then returns structured JSON you can act on. The README shows a medical query — “我发烧了该吃什么药？” — returning “disease: 发烧” and intent “medical”.

The interesting bit

The heavy lifting isn’t the fork itself; it’s the pre-trained word feature extractor built from Chinese Wikipedia and Baidu Baike. Training that from scratch takes 2–3 days on MITIE’s wordrep tool, so the project offers a downloadable model. That corpus choice — general encyclopedia text — is a pragmatic bet that works for broad domains but may drift on narrow ones.

Key highlights

• Two pipeline presets: MITIE+Jieba, or MITIE+Jieba+sklearn (the latter recommended)
• Supports custom Jieba user dictionaries for domain-specific tokenization
• Pre-trained Chinese word features available via the author’s blog
• Standard Rasa NLU server interface: train, serve, curl against localhost:5000/parse
• 1,532 stars suggests it filled a real gap in 2017-era Chinese chatbot tooling

Caveats

• README explicitly points to the official Rasa NLU docs for “newest instructions,” implying this fork may lag behind upstream
• No version compatibility notes; Rasa has since rearchitected entirely (Rasa Open Source 2.x/3.x)
• Confidence scores in the example are modest (0.54 top intent), with four other intents getting non-trivial probability mass

Verdict

Worth studying if you’re maintaining a legacy Chinese bot built on 2017-era Rasa, or if you need a fully self-hosted NLU without cloud APIs. Skip it if you’re starting fresh — modern Rasa, spaCy, or dedicated Chinese NLP frameworks (HanLP, LTP) have superseded this approach.