Checklist NOAI China 2025 Round 2 (China Stage, national final) · Task 1
Compound Word Segmentation
English title: 组合词分割问题
Build a character-level segmenter that splits German compound words into their constituent words.
The task
German compounds are formed by joining shorter words (for example 'Fußballspieler' = 'Fuß' + 'ball' + 'spieler'). The contestant must segment each compound word into its constituent words.
Each record maps a compound word to a 0-1 array with one entry per letter, where 1 marks the last letter of a constituent word and 0 marks a beginning or middle letter. The arrays are empty in the validation and test files and must be filled in. The statement recommends an embedding plus deep-learning model; the reference solving time is under one hour.
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
train.jsonwith 94,306 segmented compound words (downloadable);val.json(11,788) andtest.json(11,789) with empty label arrays, readable only through encrypted environment variables (see the baseline).- You submit
submission.ipynbcontaining the full training and inference process (a trained model alone is not accepted); it outputs a zip withsubmissionval.jsonandsubmissiontest.jsonin the training-set format.- Scoring
- Mean per-word F1 over segments: each word's labels are converted to a set of (start, end) intervals; TP = exact interval matches; precision = TP/|P|, recall = TP/|G| (0 if the denominator is 0), F1 = 2PR/(P+R) (0 if P+R = 0); the score is the mean F1 over all words. Leaderboard A uses
val.json; thetest.jsonscore (shown after the contest) is final. - Rules
- GPU training plus testing time must not exceed 10 minutes (connection and queueing time excluded); the statement notes that a suitable model trained for 8-32 epochs on a Tesla T4 fits within this limit.
- Format
- NOAI 2025 China Stage final, 2 June 2025: four problems in 6 hours on the Bohrium platform with A/B leaderboards; contestants had unrestricted access to a ChatGPT-4o-level LLM. Republished on Bohrium as the 'NOAI2026 teaching test' / 'APOAI2026 Mock Competition' (practice window 5 Dec 2025 - 20 Jun 2026). Question 1; reference solving time within 1 hour.