Data Release: Delving into Qualitative Implications of Synthetic Data for Hate Speech Detection

This repository contains the data release for the paper Delving into Qualitative Implications of Synthetic Data for Hate Speech Detection at EMNLP 2024, by Camilla Casula, Sebastiano Vecellio Salto, Alan Ramponi, and Sara Tonelli.

In this repository, we release the set of data that was manually annotated (3,500 examples in total). This subset does not include the original texts from the Measuring Hate Speech Corpus (Kennedy et al., 2020; Sachdeva et al., 2022), but only their comment IDs from the original dataset and the aggregated hate speech label we calculated for them, so the source MHS corpus should be first retrieved if one wishes to pair the source texts with their synthetic version. Here is a link to the Measuring Hate Speech Corpus.

Warning: These files contain hateful and upsetting language.

Description of the files

The annotation was divided into two sets:

• 500 examples annotated as human or llm-written. These are contained in the file human-vs-llm.tsv, which contains 5 columns:
  • comment_id: the original comment id that can be used to retrieve the original text from the MHS corpus.
  • label_x: the hate speech label we calculated after the aggregation process for the text corresponding to comment_id, based on the annotations of the MHS corpus (0 means no hate speech, 1means hate speech).
  • author: the source of the text. It is a string with the name of the LLM if the text was LLM-written (llama,mistral, or mixtral) or gold if it is an original real-world text.
  • text: the paraphrased or original text annotated by our annotators.
  • LLM?: the annotation by our annotators. It is TRUE if the annotator marked the text as LLM-written, FALSE if they thought it was a human.
• 3,000 examples (1,000 per model) annotated according to the other aspects we analyzed in the paper. These annotations are found in the files annotations-llama2-chat-7b.tsv, annotations-mistral-7b.tsv, and annotations-mixtral-8x7b.tsv. These files each contain 1,000 lines with 14 columns:
  • comment_id: the original comment id that can be used to retrieve the original text from the MHS corpus.
  • label_x: the hate speech label we calculated after the aggregation process for the text corresponding to comment_id, based on the annotations of the MHS corpus (0 means no hate speech, 1means hate speech).
  • synth_text: a string containing the LLM-generated paraphrase of the original text.
  • prompt_failure: whether the annotators deemed that the model was not able to correctly fulfill the instructions, and if so, the type of failure. Can be FALSE , Prompt failure, or Description of original gold.
  • hate_speech: whether our annotators found the synthetic text to contain hate speech or not. Can be No (no hate), Yes(hateful), or Unclear.
  • grammar_ok: if the synthetic text was deemed grammatically correct/realistic or not. Can be Yes or No.
  • world_knowledge_correct: whether the world knowledge present in the synthetic text is ok/realistic. Can be Yes or No.
  • target information: for each target category t in [origin, race, religion, gender, sexuality, age, and disability], there is a column target_[t], which is FALSE if that target is not present in the synthetic text, or a string detailing which target type it is if there is one under that category. We use the same targets as the original MHS corpus where possible for all categories but origin, since it can get extremely sparse, for which we only use the TRUEvalue if relevant.

Please note: for easier parsing and visualization of the files, we have changed all double inverted commas into single inverted commas in the texts.

Cite

The paper is set to appear in the proceedings of the EMNLP 2024 conference. It can be cited as:

Camilla Casula, Sebastiano Vecellio Salto, Alan Ramponi, and Sara Tonelli. 2024. Delving into Qualitative Implications of Synthetic Data for Hate Speech Detection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. To appear. Association for Computational Linguistics.

@inproceedings{casula-etal-2024-delving,
    title = "Delving into Qualitative Implications of Synthetic Data for Hate Speech Detection",
    author = "Casula, Camilla  and
      Vecellio Salto, Sebastiano and
      Ramponi, Alan and
      Tonelli, Sara",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    year = "2024",
    publisher = "Association for Computational Linguistics"
}