TL;DR
This paper presents the first large-scale quantitative analysis of hate speech on Parler, a platform known for its lax moderation. By developing and applying advanced hate speech detection models, the researchers identify distinct groups of users, including "hate mongers," and reveal their unique characteristics and behaviors, offering crucial insights into the dynamics of hate on alternative social media platforms.
What Problem Does This Paper Address?
Mainstream social media platforms have increasingly implemented stricter content moderation policies to combat hate speech and toxicity. This has led to a migration of users, including those with extremist views, to alternative platforms like Parler and Gab, which brand themselves as "free-speech" networks with minimal moderation. While the existence of hate speech on these platforms is acknowledged, there's a significant lack of quantitative, large-scale analysis, particularly on Parler, which has experienced rapid growth. This paper aims to fill this gap by providing a comprehensive analysis of hate speech proliferation on Parler, understanding its characteristics, and identifying the users who actively spread it, thereby offering a deeper understanding of the challenges posed by these less-moderated online spaces.
Methods
The researchers employed a multi-faceted approach, combining advanced natural language processing (NLP) techniques with network analysis.
- Post-Level Hate Speech Detection: They fine-tuned a BERT-based transformer model (specifically, DistilBERT) on a newly annotated dataset of 10,000 Parler posts. This fine-tuning was crucial, as they found that general-purpose hate speech detection models (like Jigsaw Perspective) and models trained on other platforms (like Twitter) performed poorly on Parler data, highlighting the unique nature of hate speech on this platform. They also compared their fine-tuned model against other established methods, including a lexical approach (HateBase).
- User-Level Hate Speech Classification: To identify "hate mongers" and other user groups, they adapted a diffusion-based model. This model leverages the social network structure (repost relationships) and the post-level hate speech predictions. They refined the seeding process for this diffusion model by using their fine-tuned BERT classifier, which they argue significantly improves the accuracy of user classification compared to standard diffusion models.
Datasets
- Parler Dataset: The authors created a new, annotated dataset of 10,000 Parler posts. This dataset was sampled and then manually annotated by 112 students to identify hate speech. They report a high agreement level among annotators (72% agreement, Cohen's Kappa of 0.44). This is the first publicly available annotated corpus of Parler.
- Gab Datasets: For comparison and evaluation, they utilized three existing annotated Gab datasets, totaling 42,100 posts, with 7.7K (18.4%) tagged as hateful.
- Gab User-Level Annotated Dataset: They used a user-level annotated dataset of 798 Gab users from Das et al. (2021) to validate their diffusion-based user classification model.
Key Insights and Ideas
- Platform-Specific Hate Speech: The study strongly suggests that hate speech is not monolithic across platforms. Models trained on general or other specific platforms perform poorly on Parler, necessitating platform-specific fine-tuning. This highlights the evolving nature of hateful language and the need for adaptive detection methods.
- "Hate Mongers" are Highly Active and Influential: A significant finding is that a relatively small percentage of users (around 16% in Parler) are responsible for a disproportionately large amount of hate speech. These "hate mongers" are more active, more central in the network, and have a greater propensity to spread content.
- Distinct User Groups: Beyond "hate mongers," the analysis identified other user groups, including "hate flirts" and "standard users," with distinct characteristics in terms of activity, sentiment, and network role. This nuanced categorization moves beyond a simple binary of hateful vs. non-hateful.
- Social Network Structure Matters: The integration of social network signals (repost relationships) with content analysis proved crucial for accurately identifying and characterizing hate-promoting users.
Key Results
- Post-Level Classification: Their fine-tuned BERT model significantly outperformed other models on both Parler and Gab datasets, demonstrating the effectiveness of platform-specific fine-tuning.
- User-Level Classification: They identified that approximately 16.1% of active Parler users can be classified as "hate mongers" (HM). These HM users, despite being a minority, generated 41.23% of the content on Parler and 71.38% on Gab.
- Engagement and Popularity: Hate mongers exhibit higher engagement levels, posting more, replying more, and having significantly more followers compared to other user groups.
- Emotional and Sentiment Analysis: HM users tend to express more negative sentiment and convey emotions like anger and sadness more frequently than other user groups.
- Platform Differences: Gab users, on average, are older in terms of account age compared to Parler users, potentially reflecting the different trajectories and user bases of these platforms.
Limitations and Future Work
- Time Span Disparity: The study acknowledges that the Parler and Gab datasets were collected over different time periods, which could influence the comparison. However, they argue that the datasets span similar durations and that the temporal disparity might even offer an advantage in examining model generalization.
- Ethical Considerations: The authors emphasize the ethical implications of automatically classifying users as "hate mongers," noting that such labels could lead to false positives and potentially harmful consequences. They advocate for a conservative approach and suggest involving human oversight ("man-in-the-loop") for user labeling.
- Future Work: The researchers propose two main directions for future research:
- Comparing their findings with more traditional social platforms like Twitter to understand how hate speech dynamics differ.
- Developing early detection methods for hate mongers, aiming to identify them based on their very first steps on a social platform.
Summary updated