Checklist OPEIA 2026 Virtual Phase · NLP task
Embedding Secrets
Recover names redacted from Spanish judicial documents using leaked document embeddings and a noisy candidate list.
The task
In a fictional Supreme Court data breach, the public rulings have every sensitive name replaced by 'xxxxxxxxxx', but the averaged FastText (cc.es.300.bin) embeddings of the original documents have leaked. There are 6,606 documents with 2–5 redactions each (23,131 in total) and 300 unique names.
Contestants receive the redacted text, the original and redacted document embeddings (300-d), 800 candidate name embeddings (300 noisy genuine ones and 500 decoys) and the normalised embedding of the redaction token, and must predict the hidden name at each redacted position. Names are compared lowercase and without accents.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
challenge_dataset.parquet,candidates.npy,e_bot.npy,sample_submission.csv; FastText cc.es.300.bin.- You submit
- CSV with Id '{
doc_id}_{token_position}' andpredicted_name(23,131 rows). - Scoring
- Score = 0.7 × bag score (per-document multiset overlap of names) + 0.3 × positional score (exact name at each position).
- Rules
- Solutions must be reproducible; the notebook used must also be submitted.
- Up to 5 submissions per day.
- Only the free version of ChatGPT may be used as an assistant (OPEIA page).
- Format
- OPEIA 2026, a single virtual phase on Kaggle (the OPEIA page gives 19 March – 5 April 2026; the public Kaggle competitions show a window of 23–24 April 2026).