NeSy-MMCAD: A Neuro-Symbolic Multimodal Framework for Child-Abusive Meme Detection and Explanation with Emotion Consistency

ICLR 2026 Conference Submission16294 Authors

Keywords: Neuro-Symbolic AI, Multimodal Representation Learning, Harmful Content Detection, Emotion-Aware Classification, Knowledge-Guided Regularization

Abstract: Child-abusive memes pose a serious online safety threat by combining imagery, overlaid text, and humor to mask coercive or exploitative cues. Standard multimodal classifiers, while effective on surface features, often fail in subtle or low-resource cases. We present NeSy-MMCAD, a neuro-symbolic multimodal framework for child-abusive meme detection and explanation with emotion consistency. Our architecture integrates neural perception with symbolic reasoning: neural modules extract probabilistic predicates from images and text, capturing child/adult presence, nudity, violence, toxic language, coercion, and affective signals, while domain-informed rules encode commonsense constraints. A differentiable rule loss is jointly optimized with the classification loss, enforcing symbolic consistency while retaining flexibility to learn from data. Emotion-aware rules capture affective incongruities, and mitigation rules reduce false positives in benign contexts. To support this work, we curate DACAM (Dataset for Analysis of Child-Abusive Memes), a benchmark resource for evaluating harmful content detection. Experiments on DACAM demonstrate improvements in classification accuracy and interpretability over baseline multimodal models. Importantly, rule activations provide transparent explanations that link predictions to explicit constraints. These results demonstrate the effectiveness of combining neuro-symbolic reasoning, multimodal representation learning, and emotion consistency to enhance the reliability and accountability of AI systems for socially critical tasks such as child-abuse detection.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 16294