Touché23-Human-Value-Detection

Task 4: https://touche.webis.de/semeval23/touche23-web/index.html

sharedtask.ipynb - code for main BERT model, higher performing model

logisticregression.ipynb - code for baseline logistic regression model

DOI: https://doi.org/10.5281/zenodo.6814563 Version: 2022-07-11

Dataset for Touché / SemEval 2023 Task 4; ValueEval: Identification of Human Values behind Arguments. Based on the original Webis-ArgValues-22 dataset accompanying the paper Identifying the Human Values behind Arguments, published at ACL'22.

The dataset currently contains 5220 arguments. We are, however, looking for more argument datasets (conclusion + stance + premise) to annotate and incorporate, especially datasets from different cultures and genres. Please send suggestions to our task or organizers mailing lists.

Argument Corpus

The annotated corpus in tab-separated value format. Future versions of this dataset will contain more arguments and be split into "-training", "-validation", and "-testing" files to represent the corresponding sets for the evaluation.

• arguments-training.tsv: Each row corresponds to one argument
  • Argument ID: The unique identifier for the argument
  • Conclusion: Conclusion text of the argument
  • Stance: Stance of the Premise towards the Conclusion; one of "in favor of", "against"
  • Premise: Premise text of the argument
• labels-training.tsv: Each row corresponds to one argument
  • Argument ID: The unique identifier for the argument
  • Other: Each other column corresponds to one value category, with a 1 meaning that the argument resorts to the value category and a 0 that not
• level1-labels-training.tsv: The same as labels-training.tsv but for the 54 level 1 values of the taxonomy (used in human annotation). Though not used for the 2023 task (except for the annotation), participants can still use them in their approaches.

Value Taxonomy

The value-categories.json describes the 20 value categories of this task through examples. Format:

{
  "<value category>": {
    "<level 1 value>": [
      "<exemplary effect a corresponding argument might target>",
      ...
    ], ...
  }, ...
}

The level 1 values are not used for the 2023 task (except for the annotation), but are still listed here for some might find them useful for understanding the value categories. See our paper on Identifying the Human Values behind Arguments for the complete taxonomy.

Authors

• Johannes Kiesel, Bauhaus-Universität Weimar, [email protected]
• Milad Alshomary, Paderborn University, [email protected]
• Nicolas Handke, Universität Leipzig, [email protected]
• Xiaoni Cai, Technische Universität München, [email protected]
• Henning Wachsmuth, Paderborn University, [email protected]
• Benno Stein, Bauhaus-Universität Weimar, [email protected]

Version History

• 2022-07-11

  • Exchanged the values.json from original dataset with task-specific value-categories.json

• 2022-07-09

  • Initial

License

This dataset is distributed under CC BY-SA 4.0.