Deciphering Implicit Hate: Evaluating Automated Detection Algorithms for Multimodal Hate

Austin Botelho, Scott A. Hale, Bertie Vidgen

TL;DR

This paper tackles the challenge of detecting "implicit hate" – hate speech that's subtly embedded in content, especially in multimodal formats like memes. By developing a new annotated dataset and evaluating various detection models, they show that incorporating both visual and textual context significantly improves detection accuracy, particularly for nuanced forms of hate.

What Problem Does This Paper Address?

Online hate speech is a pervasive problem, but much of the research on automated detection has focused on explicit forms (like slurs). This leaves a significant gap in addressing "implicit hate," which uses more nuanced strategies like "dog whistling" – intentionally ambiguous language that conveys hateful messages only to specific audiences. This problem is amplified in multimodal content, such as memes, where text and images combine to create meaning that's harder for algorithms to decipher than text alone. Existing hate speech datasets often overlook non-textual features, hindering the development of effective multimodal detection systems.

Methods

The researchers developed a new taxonomy for hate speech, categorizing it by Primary attribute (Hate, Counterspeech, Reclaimed, Neutral), Modality (text, image, or both), and Strategy (Explicit vs. Implicit, with further sub-categories for implicit hate like "Pseudo-factual," "Normative statements," "Coded," and "Creative expressions"). They then trained and evaluated four classes of models: unimodal-image, context-invariant unimodal-text (LSTM), context-aware unimodal-text (Transformer), and multimodal models. The multimodal models combined image features (extracted using Xception) with text features (processed by Transformer models like RoBERTa).

Datasets

The study utilized a re-annotated sample of 5,000 multimodal tweets from the MMHS150K dataset. This new annotation enriched the existing dataset with labels for primary category, modality, and strategy, providing a more granular understanding of hate speech. The authors make this corpus, along with their annotation guidelines, code, and best-performing models, publicly available.

Key Insights and Ideas

  • Multimodal context is crucial for implicit hate: While text-only models can be good at detecting explicit hate, incorporating visual information significantly boosts performance, especially for subtle forms of hate.
  • Context-aware models outperform static ones: Transformer-based models that consider word context are substantially better than simpler models.
  • Ambiguity is a major hurdle: All models perform significantly worse on content where annotators disagree, highlighting the inherent difficulty in defining and detecting hate speech, especially when it's subtle.
  • Implicit hate is more prevalent and harder to detect: The dataset revealed that a majority of hateful tweets were implicitly hateful, and models struggled more with these than with explicit hate.

Key Results

The multimodal model achieved the highest overall F1 score of 0.771, outperforming other models. The context-aware unimodal-text model was the most accurate on the subtask of implicit hate detection. While the multimodal model had a slightly lower accuracy than the context-aware unimodal-text model on explicit hate, it had a lower propensity for false positives. Notably, all models performed considerably worse on ambiguous content, with the multimodal model showing the best performance (F1 score of 0.692) on such challenging cases.

Limitations and Future Work

The paper acknowledges that annotator disagreement significantly impacts model performance, suggesting that further improvements would require more robust annotation processes or models that can better handle ambiguity. The authors also note that while their multimodal model shows improvement, the gains from incorporating image data are currently minor compared to advancements in text-based models. Future work could focus on developing more sophisticated methods for fusing multimodal information and exploring how to better disambiguate subtle or reclaimed hate speech.

Summary updated

Abstract

Accurate detection and classification of online hate is a difficult task. Implicit hate is particularly challenging as such content tends to have unusual syntax, polysemic words, and fewer markers of prejudice (e.g., slurs). This problem is heightened with multimodal content, such as memes (combinations of text and images), as they are often harder to decipher than unimodal content (e.g., text alone). This paper evaluates the role of semantic and multimodal context for detecting implicit and explicit hate. We show that both text- and visual- enrichment improves model performance, with the multimodal model (0.771) outperforming other models' F1 scores (0.544, 0.737, and 0.754). While the unimodal-text context-aware (transformer) model was the most accurate on the subtask of implicit hate detection, the multimodal model outperformed it overall because of a lower propensity towards false positives. We find that all models perform better on content with full annotator agreement and that multimodal models are best at classifying the content where annotators disagree. To conduct these investigations, we undertook high-quality annotation of a sample of 5,000 multimodal entries. Tweets were annotated for primary category, modality, and strategy. We make this corpus, along with the codebook, code, and final model, freely available.
BibTeX
@inproceedings{botelho2021deciphering,
  author = {Austin Botelho and Scott A. Hale and Bertie Vidgen},
  title = {{Deciphering Implicit Hate: Evaluating Automated Detection Algorithms for Multimodal Hate}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure