Checklist ROAI 2024 National Stage · NLP task
Grammar Is Everything: Grammatical Error Correction
Romanian title: Gramatica este totul. Corectarea erorilor gramaticale
Build and publish an encoder–decoder grammatical error correction model for Romanian, training only on provided clean Romanian Wikipedia text.
The task
Grammatical error correction (GEC) turns a possibly erroneous sentence into its corrected version. For Romanian, errors include missing diacritics, one- or two-letter misspellings, wrong prepositions, gender/number/case/person disagreements, cacophonies and punctuation mistakes; the notebook gives three example sentence pairs.
Contestants must build the best GEC model for Romanian under these rules: an encoder–decoder model (e.g. based on mBART or mT5), only 'base' variants, no GEC data or GEC-fine-tuned models from others, and training only on the provided Romanian Wikipedia texts, which may be altered freely to create training pairs. Evaluation is sentence by sentence on a diverse set of sentences.
Abridged and translated by SOTA from the official Romanian materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- The Hugging Face dataset Olimpiada-AI/
ro_wiki(access on request) and a baseline notebook that corrupts sentences trivially, fine-tunes google/mt5-base and reports BLEU, METEOR and ROUGE. - You submit
- A model pushed to the Hugging Face Hub and a technical report of at most two pages (Romanian or English); the notebook ends with a section for a report of at most 300 words.
- Scoring
- The final evaluation metric is not stated in the notebook; the baseline computes BLEU, METEOR and ROUGE on a validation split.
- Rules
- Encoder–decoder architectures only, 'base' size where available
- No third-party GEC data or GEC-fine-tuned models
- Only the provided Wikipedia texts may be used for training
- Google Colab with Hugging Face read/write tokens stored as Colab secrets
- Format
- ROAI 2024 national stage, 12 July 2024; 4-hour test in which each student chose either the NLP or the CV problem.