D-HUMOR: Dark Humor Understanding via Multimodal Open-ended Reasoning -- A Benchmark Dataset and Method

TL;DR

The paper tackles the challenge of detecting dark humor in multimodal memes by introducing the D-HUMOR dataset, a large Reddit-based collection annotated for dark humor presence, target, and intensity. It proposes a reasoning-augmented architecture that generates and iteratively refines structured explanations using a Vision-Language Model and a Role-Reversal Self-Loop, followed by fusion of OCR text, image features, and reasoning signals through a Tri-stream Cross-Reasoning Network. The approach yields state-of-the-art results on dark humor detection, target identification, and intensity prediction, with ablations showing the crucial importance of the reasoning component. By releasing the dataset and code, the work provides a resource and methodology to advance multimodal humor understanding and content moderation in real-world social media settings.

Abstract

Dark humor in online memes poses unique challenges due to its reliance on implicit, sensitive, and culturally contextual cues. To address the lack of resources and methods for detecting dark humor in multimodal content, we introduce a novel dataset of 4,379 Reddit memes annotated for dark humor, target category (gender, mental health, violence, race, disability, and other), and a three-level intensity rating (mild, moderate, severe). Building on this resource, we propose a reasoning-augmented framework that first generates structured explanations for each meme using a Large Vision-Language Model (VLM). Through a Role-Reversal Self-Loop, VLM adopts the author's perspective to iteratively refine its explanations, ensuring completeness and alignment. We then extract textual features from both the OCR transcript and the self-refined reasoning via a text encoder, while visual features are obtained using a vision transformer. A Tri-stream Cross-Reasoning Network (TCRNet) fuses these three streams, text, image, and reasoning, via pairwise attention mechanisms, producing a unified representation for classification. Experimental results demonstrate that our approach outperforms strong baselines across three tasks: dark humor detection, target identification, and intensity prediction. The dataset, annotations, and code are released to facilitate further research in multimodal humor understanding and content moderation. Code and Dataset are available at: https://github.com/Sai-Kartheek-Reddy/D-Humor-Dark-Humor-Understanding-via-Multimodal-Open-ended-Reasoning

Figures (5)

• Figure 1: A comparison of clean and dark humor (Source: Reddit).
• Figure 2: A subset of memes from D-Humor dataset, classified by their dark humor presence, target, and intensity. (Source: Reddit)
• Figure 3: Overview of the proposed reasoning‐augmented multimodal architecture. Each meme (image + OCR text) is first processed by VLM to generate and refine a structured explanation via the Role-Reversal Self-Loop mechanism. Text encoders encode the OCR transcript and the self-refined explanation into textual embeddings, while the image encoder extracts visual embeddings from the meme image. A Tri‐stream Cross‐Reasoning Network (TCRNet) then applies pairwise scaled dot-product attention across the three modalities, text, image, and reasoning, to produce attended representations. These representations are concatenated into a unified feature vector, which is passed through a classification head to predict dark humor presence, target category, and intensity.
• Figure 4: Self-loop mechanism using Role Reversal (RR) Prompting. The input meme is passed to a language model which generates an initial explanation. This explanation is then refined iteratively through Role Reversal prompting, where the model assumes the role of a reviewer to critique and improve its own output. After $N$ such refinement cycles, a final explanation is obtained. This mechanism enables the model to simulate self-awareness and improve explanatory coherence without external supervision.
• Figure 5: The charts represent an ablation study of TCRNet on Dark Humor Meme Analysis. Each ablation removes one component (OCR Text, Image, or Explanation) to study its impact. The full TCRNet results are included in \ref{['tab:results_comparison']}.