FigSIM: Suicide Meme Dataset—A New Challenge for Multimodal Content Moderation

FigSIM is the first dataset with fine-grained annotations for suicide-related memes, providing a new evaluation benchmark for multimodal content moderation models. The dataset contains 1049 memes, annotated from three dimensions: suicide severity, rhetorical phenomena, and suicide-related content, aiming to address the complex dilemmas in moderating suicide memes on social media.

Source Information:

• Original Author/Maintainer: Paper author team (arXiv)
• Source Platform: arXiv
• Original Title: FigSIM: A Dataset for Fine-grained Suicide Severity and Figurative Language in Suicide Memes
• Original Link: http://arxiv.org/abs/2606.02523v1
• Publication Date: June 1, 2026

Memes have become an important part of internet culture, but the emergence of suicide memes poses moderation challenges: they can be help signals, black humor, or even dangerous content. Existing multimodal models struggle to accurately understand their complex semantics. Both over-moderation and under-moderation cause problems—over-moderation stifles emotional expression, while under-moderation allows harmful content to spread. The core dilemma now is that memes rely on cultural background and context; when suicide is involved, understanding becomes much harder. The multi-layered and ambiguous nature of semantics is a shortcoming of existing systems.

The FigSIM dataset achieves a breakthrough in fine-grained annotation, covering three key dimensions:

1. Suicide Severity Grading: Non-binary classification with fine-grained levels, covering a continuous spectrum from mild hints to explicit expressions, providing a basis for differentiated handling;
2. Rhetorical Phenomenon Recognition: Annotates rhetorical devices like metaphor and irony, addressing the gap between literal understanding and deep semantics in models;
3. Suicide-related Content Detection: Identifies sensitive content such as specific suicide methods and tools, tackling the challenge of meaning differences of items in different contexts.

The research team evaluated 16 mainstream unimodal and multimodal models using FigSIM, covering traditional machine learning to the latest large models. The results show:

• Even the most advanced models have obvious limitations, especially a significant performance drop when handling memes with rhetorical devices;
• Models generally underestimate high-severity suicide content, which may miss dangerous signals;
• Reasons include uneven training data distribution (biased towards safe categories) and the complexity of rhetorical language exceeding the models' understanding capabilities (requiring cultural background, context, and common sense reasoning).

The difficulty in moderating suicide memes stems from three challenges:

1. Multimodal Characteristics: Images and text complement or contradict each other; literally negative text paired with specific images may convey self-mockery or help signals;
2. Context Dependency: The same template has different meanings in different communities or at different times; cultural gaps are a fundamental challenge for global moderation;
3. Ethical and Data Constraints: Annotation of sensitive topics and data acquisition need to balance academic value and ethical responsibility; the current dataset size (1049 memes) is still limited.

Practical significance of FigSIM:

• For platforms: Evaluate and optimize existing moderation systems, identify model blind spots;
• For researchers: Provide a standardized evaluation benchmark to drive technological progress.

Recommendations: Automated moderation should not be the only line of defense. It needs to be combined with human moderation, community norm guidance, and mental health support resource connection. Especially for sensitive topics, it requires technical judgment + manual review + professional support.

FigSIM represents an important progress in AI content moderation, but it also reveals technical limitations. For suicide-related topics, we need to balance intelligent algorithms (risk identification) and ethical boundaries (cautious decision-making). The research value lies not only in the dataset but also in revealing the complexity of the problem—suicide memes require multi-dimensional understanding and multi-level intervention. Future research needs to improve model performance while focusing on fairness, interpretability, and humanistic care, ensuring that technology serves human well-being.