How spraakbanken/multiged-2023 works
MultiGED-2023 is a shared-task data and evaluation repository for token-level grammatical error detection in Czech, English, German, Italian, and Swedish. Participants obtain annotated corpus splits, develop systems that label tokens correct or incorrect, and evaluate hypothesis TSVs against references with the included scorer; German M2 annotations can also be converted to a tabular format. The README documents the task, datasets, and official CodaLab evaluation. The sampled scripts show conversion and scoring behavior; dataset internals and participant systems are not part of this repository.
Main entry points: eval.py Token scorer · german German Falko-MERLIN · czech Czech GECCC · english English corpora
Components 10
Language corpora
- Czech GECCC
czech - English corpora
english - German Falko-MERLIN
german - Italian MERLIN
italian - Swedish SweLL
swedish
Task utilities
- M2 converter
m2_to_conll_conversion-script.py - Token scorer
eval.py
Other components
- Participant
- Hypothesis TSV
- CodaLab evaluation
Connections 15
- Participant → Czech GECCC: gets data
README.md - Participant → English corpora: gets data
README.md - Participant → German Falko-MERLIN: gets data
README.md - Participant → Italian MERLIN: gets data
README.md - Participant → Swedish SweLL: gets data
README.md - German Falko-MERLIN → M2 converter: M2 input
m2_to_conll_conversion-script.py - M2 converter → Participant: writes tabular output
m2_to_conll_conversion-script.py - Participant → Hypothesis TSV: produces predictions
- Hypothesis TSV → Token scorer: hypothesis input
eval.py - Czech GECCC → Token scorer: reference input
- English corpora → Token scorer: reference input
- German Falko-MERLIN → Token scorer: reference input
- Italian MERLIN → Token scorer: reference input
- Swedish SweLL → Token scorer: reference input
- Participant → CodaLab evaluation: submits system
README.md