A Strategy Labelled Dataset of Counterspeech

Aashima Poudhar, Ioannis Konstas, Gavin AbercrombieNAACL 2024

TL;DR

This paper introduces a novel dataset of 1000 hate speech/counterspeech pairs, annotated with strategies derived from social sciences. They demonstrate that large language models (LLMs) can effectively classify these strategies, paving the way for more nuanced and targeted automated counterspeech generation.

What Problem Does This Paper Address?

Online platforms are rife with hateful and toxic language, impacting a significant portion of the global population. While traditional content moderation (like flagging and removal) has limitations, counterspeech offers a promising alternative to mitigate harm while upholding free speech. However, existing Natural Language Processing (NLP) research has largely overlooked the systematic study of how counterspeech works, particularly the diverse strategies employed. This paper aims to bridge this gap by developing a dataset that categorizes counterspeech based on established social science strategies and then evaluating the ability of LLMs to recognize these strategies.

Methods

The core of this research involves two main components: dataset creation and automated classification.

  1. Dataset Creation: The authors started with the "Multitarget-CONAN" dataset, which contains hate speech/counterspeech pairs. They then manually annotated a subset of 1000 examples with seven distinct counterspeech strategies. These strategies were derived from existing literature in social sciences and policy, refined through a pilot study with annotators, and included categories like Fact-Checking, Denouncing, Pointing Out Hypocrisy, and Questioning. To ensure reliability, they employed multiple annotators and measured inter-annotator agreement using Cohen's kappa and Fleiss' kappa.
  2. Automated Classification: To assess the potential of LLMs in this domain, they used a one-shot prompting approach with a GPT-3.5 model. This involved providing the model with a prompt that included the hate speech, the counterspeech, and a list of possible strategies, asking it to classify the counterspeech. The model's performance was evaluated using precision, recall, and F1 scores.

Datasets

  • Multitarget-CONAN dataset: This is the source dataset for hate speech/counterspeech pairs. The authors sampled 1000 examples from this dataset for their annotation.
  • Newly introduced benchmark: The paper introduces a new dataset of 1000 hate speech/counterspeech pairs annotated with seven specific counterspeech strategies. This dataset is made available upon acceptance of the paper.

Key Insights and Ideas

  • Counterspeech is strategic: The research confirms that counterspeech isn't monolithic; it employs a variety of deliberate strategies, much like other forms of communication.
  • LLMs can learn counterspeech strategies: The study demonstrates that even with a relatively small dataset and a simple one-shot prompt, LLMs show a promising ability to distinguish between these nuanced strategies.
  • Fact-checking is prevalent but not always effective: The dataset revealed that "Fact-Checking" was the most common strategy, despite research suggesting its effectiveness can be limited by cognitive biases. This highlights a potential disconnect between common practice and proven efficacy.
  • Annotation is challenging: Achieving high inter-annotator agreement was difficult, indicating the subjective nature of classifying counterspeech strategies. This underscores the need for clear guidelines and robust training.

Key Results

  • Inter-annotator agreement: After refinement and training, the two main annotators achieved a Cohen's kappa of 0.56 (moderate agreement) on the full dataset, indicating that consensus could be reached on strategy classification.
  • Automated classification performance: The GPT-3.5 model achieved a macro-averaged F1 score of 0.62 in classifying counterspeech strategies. This is a significant improvement over a majority class baseline (F1 of 0.57).
  • Strategy-specific performance: The model performed better on some strategies than others, suggesting areas for future model improvement.

Limitations and Future Work

  • Dataset size: The 1000-example dataset, while valuable, is relatively small and may not fully represent all counterspeech strategies, particularly less common ones.
  • LLM limitations: The current LLM performance, while promising, still has room for improvement, especially in recall. The authors acknowledge the limitations of capturing the full intricacy of human language.
  • Closed-source models: The use of commercial LLMs raises reproducibility concerns.
  • Future Work: The authors plan to explore more advanced prompting techniques, expand the dataset, and fine-tune LLMs specifically for strategy-driven counterspeech generation. They also aim to investigate the relative efficacy of different strategies in various contexts, moving beyond just classification.

Summary updated

BibTeX
@inproceedings{poudhar2024strategy,
  author = {Aashima Poudhar and Ioannis Konstas and Gavin Abercrombie},
  title = {{A Strategy Labelled Dataset of Counterspeech}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure