Shared Task Data

The full data will be published in Lindat after the shared task. During the task, the data is temporarily available at the following links:

Training and Development Data

Within the .tar.gz package, the datasets are organized similarly to UD releases, that is, CoNLL-U files are in subfolders that represent treebanks rather than languages. Each folder contains the README and LICENSE files (all treebanks come with some type of open license but some of them are restricted for non-commercial use). Besides gold-standard CoNLL-U files, there are also raw text files; this is the form which the participating systems should expect as input. Some folders do not contain data because the corresponding UD treebank contains only test data and no training data. Participants can expect that data from such treebanks will be part of the evaluation; while they cannot use the actual data from those treebanks to train their systems, they may use the information from the README files (e.g. to learn what genres the treebank contains). There is always at least one other treebank for the same language that has training data.

See the task description for more details on what data can be used in the shared task.

Gold and Blind Data by Language

Besides being included in the train-dev package described above, the development data is also available in the form in which the test data will be provided. Participants can use this data to test their system in similar conditions as those in which the test phase will run. Here the blind package contains simply one text file per language, and the system submission should mimic the structure of the gold package.

Change Log

  • 2021-04-05 22:14 UTC+2: The first train-dev package published.
  • 2021-04-07 21:02 UTC+2: Updated annotation of Arabic, Czech, Lithuanian, and Slovak.
  • 2021-04-12 12:58 UTC+2: Added more English data (from the GUM treebank).