Towards Automatic Generation of Messages Countering Online Hate Speech and Microaggressions

Mana Ashida, Mamoru KomachiNAACL 2022

TL;DR

This paper explores using large language models (LLMs) like GPT-3 to automatically generate "counternarratives" – messages that push back against online hate speech and microaggressions. They found that GPT-3, when prompted with a few examples, produced the highest quality counternarratives, and they've released a new dataset (CHASM) to help further research in this area.

What Problem Does This Paper Address?

Online hate speech and microaggressions are growing problems, causing harm and creating toxic environments. While detecting these harmful messages is one approach, simply flagging them can infringe on free speech. A more constructive approach is "counterspeech" or "counternarratives" – messages that actively counter the harmful content. However, manually creating these responses is labor-intensive. This paper aims to automate this process by using LLMs to generate effective counternarratives, extending beyond just hate speech to include the more subtle, often unintentional, "microaggressions."

Methods

The researchers explored using pre-trained language models (GPT-2, GPT-Neo, and GPT-3) for counternarrative generation. Instead of fine-tuning the models, which requires extensive data, they focused on "prompting." This involves providing the LLM with a few examples of the task (hate speech + counternarrative) to guide its generation. They designed a "few-shot" prompt strategy, using examples from existing datasets and their own curated prompts.

To evaluate the generated counternarratives, they developed a human evaluation framework with three key dimensions:

  • Offensiveness: Whether the generated message itself is offensive.
  • Stance: Whether the message effectively opposes the original offensive input.
  • Informativeness: Whether the message is specific and provides useful information, rather than being generic.

Datasets

The study utilized existing datasets for hate speech inputs:

  • CONAN: A dataset of hate speech and counternarrative pairs.
  • Multitarget-CONAN: Similar to CONAN but with multiple targets.
  • Knowledge-grounded hate countering: Another dataset for hate speech.

For microaggression inputs, they used:

  • Social Bias Inference Corpus (SBIC): Specifically, the "microaggression" category from its development set, which is based on the SELFMA dataset.

The researchers also introduced a new dataset:

  • CHASM (Countering Hate Speech and Microaggressions): This corpus contains machine-generated counternarratives annotated with offensiveness scores of the original abusive posts.

Key Insights and Ideas

  • Prompting is a viable alternative to fine-tuning: For counternarrative generation, prompting LLMs with a few examples is a computationally efficient way to achieve good results, especially when task-specific data is scarce.
  • GPT-3 excels at generating high-quality counternarratives: Among the models tested, GPT-3 consistently produced the most effective, safe, and informative counternarratives.
  • Microaggressions are a unique challenge: Countering microaggressions is more difficult than hate speech due to their subtle nature and the lack of immediate recognition of their offensiveness.
  • Human evaluation is crucial but complex: Defining and consistently measuring the quality of counternarratives through human judgment is challenging, highlighting the need for well-defined annotation schemes.
  • Potential for misinformation: While LLMs can generate informative counternarratives, they can also inadvertently produce misinformation or "hallucinate" facts, requiring careful post-editing or fact-checking mechanisms.

Key Results

  • GPT-3 outperformed GPT-2 and GPT-Neo in generating counternarratives that were rated as safe, disagreeing with the input, and informative.
  • The study found that GPT-3's generated text length was more similar to human text compared to GPT-2 and GPT-Neo.
  • CHASM dataset creation: The paper introduced a new corpus of annotated machine-generated counternarratives, which is valuable for future research.
  • Inter-annotator agreement for offensiveness, stance, and informativeness was comparable to previous studies, with GPT-3 showing higher agreement for offensiveness.

Limitations and Future Work

  • Misinformation and Hallucination: The authors acknowledge that generated "informative" counternarratives sometimes contained misinformation. Future work could involve integrating fact-checking mechanisms.
  • Focus on English: The CHASM corpus is currently only in English. Expanding to other languages would be beneficial.
  • Outdated Hate Speech Content: The hate speech examples in the dataset may not fully reflect the most current forms of online abuse.
  • Types of Counternarratives: The paper suggests exploring different types of counternarratives (e.g., humor, warnings) and how to evaluate their effectiveness.
  • Qualitative Analysis of Microaggressions: Further qualitative analysis is needed to understand the nuances of generating effective microinterventions.
  • Automated Evaluation Metrics: Developing more robust automated metrics for evaluating counternarrative quality is an ongoing challenge.

Summary updated

BibTeX
@inproceedings{ashida2022towards,
  author = {Mana Ashida and Mamoru Komachi},
  title = {{Towards Automatic Generation of Messages Countering Online Hate Speech and Microaggressions}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure