sbx/superlim-2.1

Dataset Card for Superlim-2

Dataset Summary

SuperLim 2.0 is a continuation of SuperLim 1.0, which aims for a standardized suite for evaluation and analysis of Swedish natural language understanding systems. The projects is inspired by the GLUE/SuperGLUE projects from which the name is derived: "lim" is the Swedish translation of "glue".

Since Superlim 2.0 is a collection of datasets, we refer for information about dataset structure, creation, social impact etc. to the specific data cards or documentation sheets in the official GitHub repository: https://github.com/spraakbanken/SuperLim-2/

Supported Tasks and Leaderboards

See our leaderboard: https://lab.kb.se/leaderboard/

Languages

Swedish

Dataset Structure

Data Instances

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Data Fields

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Data Splits

Most datasets have a train, dev and test split. However, there are a few (supersim, sweanalogy and swesat-synonyms) who only have a train and test split. The diagnostic tasks swediagnostics and swewinogender only have a test split, but they could be evaluated on models trained on swenli since they are also NLI-based.

Dataset Creation

Curation Rationale

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Source Data

Initial Data Collection and Normalization

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Who are the source language producers?

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Annotations

Annotation process

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Who are the annotators?

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Personal and Sensitive Information

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Considerations for Using the Data

Social Impact of Dataset

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Discussion of Biases

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Other Known Limitations

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Dataset Curators

See individual datasets: https://github.com/spraakbanken/SuperLim-2/

Licensing Information

All datasets constituting Superlim are available under Creative Commons licenses (CC BY 4.0, 8144 CC BY-SA 4.0, respectively).

Citation Information

To cite as a whole, use the standard reference. If you use or reference individual resources, cite the references specific for these resources:

Standard reference:

Superlim: A Swedish Language Understanding Evaluation Benchmark (Berdicevskis et al., EMNLP 2023)

@inproceedings{berdicevskis-etal-2023-superlim,
    title = "Superlim: A {S}wedish Language Understanding Evaluation Benchmark",
    author = {Berdicevskis, Aleksandrs  and
      Bouma, Gerlof  and
      Kurtz, Robin  and
      Morger, Felix  and
      {\"O}hman, Joey  and
      Adesam, Yvonne  and
      Borin, Lars  and
      Dann{\'e}lls, Dana  and
      Forsberg, Markus  and
      Isbister, Tim  and
      Lindahl, Anna  and
      Malmsten, Martin  and
      Rekathati, Faton  and
      Sahlgren, Magnus  and
      Volodina, Elena  and
      B{\"o}rjeson, Love  and
      Hengchen, Simon  and
      Tahmasebi, Nina},
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.506",
    doi = "10.18653/v1/2023.emnlp-main.506",
    pages = "8137--8153",
    abstract = "We present Superlim, a multi-task NLP benchmark and analysis platform for evaluating Swedish language models, a counterpart to the English-language (Super)GLUE suite. We describe the dataset, the tasks, the leaderboard and report the baseline results yielded by a reference implementation. The tested models do not approach ceiling performance on any of the tasks, which suggests that Superlim is truly difficult, a desirable quality for a benchmark. We address methodological challenges, such as mitigating the Anglocentric bias when creating datasets for a less-resourced language; choosing the most appropriate measures; documenting the datasets and making the leaderboard convenient and transparent. We also highlight other potential usages of the dataset, such as, for instance, the evaluation of cross-lingual transfer learning.",
}

Thanks to Felix Morger for adding this dataset.