Building the Romanian NLP API that should already exist

If you've tried to do anything programmatic with Romanian text, you've probably hit the same wall I did.

There's no clean API for it. You end up scraping DEXonline, wrestling with incomplete library support, or calling a general-purpose LLM and hoping it gets the grammar right. None of that is good enough for production.

The specific gap

Given an arbitrary Romanian sentence, return for each token: its lemma, part of speech, grammatical case, number, gender, person, and tense.

This is what spaCy does for English, French, and German in a pip install. LexicRo is that as a hosted, versioned HTTP contract for Romanian: callable without shipping a model, stamping the version it answered with, and telling you per token whether the answer came from a lexicon or a prediction.

Romanian NLP tooling lags well behind English. The academic resources are there — DEXonline (313k+ lemmas), RoLEX (330k morphosyntactic entries), the Universal Dependencies Romanian Treebank — they're just not packaged in a way developers can actually use.

But why not just use ChatGPT?

Fair question. The short answer: for production, an LLM is not infrastructure — it's an oracle. The accusative form of câine is always câinele, regardless of the LLM's mood that day. LexicRo returns structured linguistic data, and every response carries the model_version that produced it. The weights, lexicon and tagset are versioned together as one unit, so you can pin against a version and know exactly what you are pinned to. Beyond that: LLM costs at scale are unpredictable; a dedicated endpoint does one thing and does it the same way every time; and structured output from an LLM requires prompt engineering, validation, and retry logic. LexicRo returns clean JSON, every time.

What I'm building

LexicRo — an open-core, hosted API platform covering the endpoints Romanian developers actually need:

POST /analyze
→ lemma, POS, case, gender, number, person, tense per token, plus 
  where each answer came from

GET /conjugate/{verb}
→ conjugation table — seven moods, including perfect simplu
  and viitor I. For a verb the conjugator does not recognise,
  returns a predicted paradigm and marks it as predicted.

GET /inflect/{word}
GET /lookup/{word}
POST /difficulty
→ these were withdrawn on 2026-08-17 and are not currently served.

Technical approach

Not starting from scratch — the data and models are there:

• Base model: bert-base-romanian-cased-v1 fine-tuned for morphological tagging
• Conjugation: verbecc Romanian XML templates
• Lexical: DEXonline database dump + RoLEX dataset
• Infrastructure: FastAPI, Docker, full OpenAPI spec

Licence and access

• Code: MIT
• Model weights: licensing terms for the model weights are still being worked out
• Free tier: 1,000 req/day, no credit card, all endpoints

Phase 1 ships first

The conjugation and lexical lookup endpoints are the straightforward part — wrapping verbecc and DEXonline cleanly. That's what ships first (~3 months). The morphological analyser (the hard part, requiring fine-tuned BERT) follows in phase 2.

What I'm looking for

I'm in pre-development and genuinely looking for:

1. Feedback on the endpoint design — does this cover what you'd actually need?
2. Early users working with Romanian text at any scale
3. Academic connections
4. Anyone who's built adjacent to this — what did you learn?

Update, 2026-08-30: the plan described above changed. /lookup, /inflect and /difficulty were withdrawn on 2026-08-17. /analyze — described here as Phase 2 work — shipped and is live. There is a demo, no key required: https://demo.lexicro.com

Links: lexicro.com · demo.lexicro.com · github.com/LexicRo · [email protected]

Romanian deserves the same NLP infrastructure as French or German. Building it in public — feedback welcome.