Hate Speech Dataset from a White Supremacy Forum

Ona de Gibert, Naiara Perez, Aitor García-Pablos, Montse CuadrosEMNLP 2018

TL;DR

This paper introduces a novel, manually annotated dataset of hate speech from the white supremacist forum Stormfront, focusing on sentence-level analysis. This resource aims to improve the detection of hate speech by providing a more nuanced and context-aware dataset than previously available, enabling better research into this critical online safety issue.

What Problem Does This Paper Address?

The proliferation of online content has led to a significant increase in hate speech, posing a challenge for social media platforms and policymakers. While hate speech is generally understood as communication disparaging a group based on characteristics like race or religion, a universally agreed-upon definition remains elusive. This ambiguity makes automated detection difficult. Existing datasets often rely on broader definitions of "toxic" or "offensive" language, or are not publicly available. Furthermore, many studies analyze comments at the document level, potentially missing hate speech embedded within otherwise clean text or over-labeling entire comments based on a single hateful sentence. This paper addresses the need for a more precise, publicly available, sentence-level dataset specifically from a hate-group forum to better understand and combat hate speech.

Methods

The authors developed a custom annotation tool to manually label thousands of sentences extracted from the Stormfront forum. The annotation process involved creating detailed guidelines, refined through multiple rounds of labeling and discussion among annotators. Key to their approach is the consideration of context, allowing annotators to view the entire post and even previous posts in a thread to make informed decisions. They also introduced a "RELATION" label for sentences that, while not hateful in isolation, contribute to hate speech when considered together. After data collection, they performed baseline experiments using standard machine learning models (SVM, CNN, LSTM) to evaluate the dataset's utility and identify challenges in hate speech detection.

Datasets

The primary dataset introduced in this paper is a collection of 10,568 sentences extracted from the Stormfront forum, manually annotated for hate speech. This dataset is publicly available on GitHub. The authors also reference existing datasets and resources such as Hatebase, Kaggle's Toxic Comment Classification Challenge, and various tweet-based datasets, but their focus is on the novel Stormfront dataset.

Key Insights and Ideas

  • Context is Crucial: The study highlights that annotators frequently needed to consult the context of a post (and sometimes previous posts) to accurately classify sentences, especially for identifying hate speech. This underscores the difficulty of sentence-level hate speech detection without contextual information.
  • Subjectivity and Nuance: The annotation process revealed the inherent subjectivity in defining hate speech, leading to the development of detailed guidelines and the "RELATION" label to capture subtle forms of hate speech that emerge from the combination of sentences.
  • Stormfront as a Source: Extracting data from a dedicated white supremacist forum like Stormfront provides a rich, albeit challenging, source of hate speech, allowing for a focused study on the language and targets of such discourse.
  • Vocabulary Predicts Hatefulness: Analysis of word scores (using PMI) revealed that derogatory terms and group-specific slurs strongly correlate with hate speech, while neutral or forum-specific political terms are less indicative.

Key Results

The dataset consists of 10,568 sentences, with the vast majority (86.09%) labeled as NOHATE and a smaller portion (11.29%) as HATE. The "RELATION" and "SKIP" (non-English or irrelevant) categories accounted for the remaining sentences. Baseline experiments showed that while NOHATE sentences were more accurately classified (around 70-80% accuracy), HATE classification accuracy was lower (around 50-70%). LSTM-based classifiers generally performed best. Notably, including additional context during annotation led to slightly lower classification accuracy for automatic models, suggesting that the nuances requiring human context are difficult for machines to grasp.

Limitations and Future Work

The authors acknowledge that the annotation process was demanding and that the definition of hate speech, particularly what constitutes an "attack," remains somewhat ambiguous. They also note that the dataset is unbalanced, with significantly more non-hate speech examples. Future work suggested includes incorporating world knowledge and richer conversational context into automatic classifiers, exploring the impact of the "RELATION" label on classification, and further characterizing the dataset's content. They also propose that active learning paradigms could benefit future labeling efforts due to the dataset's unbalanced nature.

Summary updated

BibTeX
@inproceedings{gibert2018hate,
  author = {Ona de Gibert and Naiara Perez and Aitor García-Pablos and Montse Cuadros},
  title = {{Hate Speech Dataset from a White Supremacy Forum}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure