RoLegalGEC: Legal Domain Grammatical Error Detection and Correction Dataset for Romanian

The harder problem here isn't building the dataset, it's the dual constraint the researchers faced: Romanian morphology is significantly more complex than English, and legal language adds a second layer of domain-specific phrasing that general-purpose grammar models consistently mishandle. Synthetic generation was necessary precisely because annotated legal corpora in Romanian are scarce to the point of being nearly nonexistent.

This work sits in the same broader territory as the MADE benchmark covered here in mid-April, where researchers built a living dataset for medical adverse event classification to fill a domain-specific training data gap in a high-stakes field. Both projects share the same structural argument: general-purpose models trained on broad corpora fail in specialized professional domains, and the fix starts with better-targeted data. The related coverage on LLM judge reliability is also worth noting, since automated evaluation of grammatical corrections in low-resource languages faces exactly the consistency problems that conformal prediction research is trying to address.

Watch whether any Romanian legal tech vendors or court administration bodies formally adopt RoLegalGEC as a benchmark within the next 12 months. Institutional uptake, not citation count, is the signal that this dataset escapes the research-only lane.