Discord

Checklist IOAI 2025 GAITE Contest · Task 4

Combinatorial Word Segmentation

Split German compound words into their component words by predicting, for every letter, whether a component word ends there.

  • NLP
  • Sequence labelling (word segmentation)

The task

German often builds long compound words out of shorter ones, such as Fußball from Fuß and Ball. The task is to write a splitter that recovers the component words of each compound.

Each entry maps a compound to a 0/1 array with one value per letter, where 1 marks the last letter of a component word. The training file has 94,306 labelled compounds; the validation (11,788) and test (11,789) files list words with empty arrays for you to fill in.

Your submitted notebook must train the model and then predict both sets; a trained model alone is not accepted. The statement suggests embeddings with a deep learning model such as an LSTM.

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
JSON files mapping each compound word to a per-letter 0/1 array: train.json (94,306 entries), val.json (11,788) and test.json (11,789), the last two with empty arrays.
You submit
submission.ipynb that trains and predicts, producing a zip with submissionval.json and submissiontest.json in the same format as the training set.
Scoring
The average F1-score over compound words. The repo's metrics.py compares exact segment boundaries per word and scores a label array of the wrong length as 0. The Scientific Committee's highest score (0.95) and baseline score (0) on Leaderboard B were used for score normalisation.
Rules
  • Notebook must include both training and inference; submitting only a trained model is not allowed
  • Notebook must finish within 20 minutes on the test machine
  • At most 50 successful submissions; up to 2 can be selected and the higher Leaderboard B score counts
  • No self-made mounted datasets or .pth files may be submitted
  • Test machine has no internet access; pretrained models may be used only if they import offline
Format
GAITE Contest, Day 2 (6 August 2025); six-hour contest day

Details

Year
2025, Beijing, China
Round
GAITE Contest · Task 4
Language
English
License
CC BY 4.0, as stated by the source