Findings of the WOAH 5 Shared Task on Fine Grained Hateful Memes Detection

Lambert Mathias, Shaoliang Nie, Aida Mostafazadeh Davani, Douwe Kiela, Vinodkumar Prabhakaran, Bertie Vidgen, Zeerak WaseemACL 2021

TL;DR

This paper presents the findings of the WOAH 5 shared task, which focused on fine-grained detection of hateful memes. The task introduced two sub-challenges: identifying the protected category attacked by the meme and the type of attack used. This work highlights the complexity of multimodal hate speech detection and the need for models that can understand the interplay between text and images.

What Problem Does This Paper Address?

The proliferation of online hate speech is a significant societal concern, and social media platforms are increasingly tasked with moderating such content. However, accurately detecting hate speech, especially in the form of memes, is a complex challenge. Memes combine text and images, and their hateful nature often arises from the interaction between these modalities, making it difficult for systems that analyze them separately. Existing approaches often rely on binary classifications of textual content, which are insufficient for capturing the nuances of hate speech. This paper addresses the need for more sophisticated methods to detect hateful memes by focusing on fine-grained classification of what is being attacked (protected categories) and how it is being attacked (attack types).

Methods

The WOAH 5 shared task involved two subtasks:

  • Task A: Protected Category Detection: Identifying the protected group(s) targeted by the meme (e.g., race, disability, religion, nationality, sex).
  • Task B: Attack Type Detection: Identifying the specific type of attack used (e.g., contempt, mocking, inferiority, slurs, exclusion, dehumanizing, inciting violence).

Both tasks were framed as multi-label classification problems, as a single meme could target multiple categories or employ multiple attack types. The evaluation metric used was the standard ROC AUC for multi-label classification. Participants were encouraged to develop multimodal systems, though some systems focused solely on text.

Datasets

The shared task utilized a reannotated version of the hateful memes dataset from Kiela et al. (2020). This dataset was expanded with fine-grained labels for protected categories and attack types, annotated by professional moderators. For non-hateful memes, labels were set to "none" for both categories. The dataset was split into training, development (seen and unseen), and testing (seen) sets.

Key Insights and Ideas

  • Multimodality is Crucial: The paper emphasizes that hateful memes are inherently multimodal, and models need to learn to understand the interaction between text and images holistically, rather than processing them independently.
  • Fine-Grained Detection is Necessary: Moving beyond simple "hateful" vs. "not hateful" classifications is essential for a deeper understanding and more effective mitigation of online hate. Identifying specific attack types and targets provides richer insights.
  • Text-Only Models Can Be Surprisingly Strong: Despite the multimodal nature of memes, some text-only systems performed remarkably well, even outperforming multimodal baselines. This suggests that the textual component can carry significant information, or that there might be biases in the dataset that text-based models can exploit.
  • Pre-trained Representations are Key: All participating systems leveraged pre-trained multimodal or unimodal representations, highlighting their importance in achieving strong performance on complex tasks like hate speech detection.

Key Results

The paper reports the results of the shared task submissions. While specific winning systems are detailed, the overall performance across participants is presented in Table 4. For Task A (protected category), the best performing system achieved an ROC AUC of 0.914. For Task B (attack type), the best performing system achieved an ROC AUC of 0.913. The majority baseline performed poorly, achieving 0.70 for Task A and 0.72 for Task B. The paper notes that systems exploiting image-level metadata and state-of-the-art multimodal representations like VLBERT showed strong performance on Task A.

Limitations and Future Work

The authors acknowledge that detecting hate speech remains technically difficult, with many unsolved challenges. They suggest that future work should focus on scaling fine-grained annotations to other hate speech datasets to develop classifiers that can better detect the nuances of hate speech. The annotated datasets are made publicly available to encourage further research in this area.

Summary updated

BibTeX
@inproceedings{mathias2021findings,
  author = {Lambert Mathias and Shaoliang Nie and Aida Mostafazadeh Davani and Douwe Kiela and Vinodkumar Prabhakaran and Bertie Vidgen and Zeerak Waseem},
  title = {{Findings of the WOAH 5 Shared Task on Fine Grained Hateful Memes Detection}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure