Internet memes are a dominant form of online communication, but their unique structure makes them a potent vehicle for spreading hate speech, misogyny, and political propaganda. Unlike traditional text-based abuse, a meme's harmful intent rarely resides in the text or the image alone. Instead, the harm emerges from the juxtaposition of both modalities.
To automate the detection of harmful memes, researchers have shifted away from unimodal analysis toward multimodal fusion. The goal is to build architectures that do not merely process text and images in parallel, but actively model the complex, non-linear interplay between them.
The Challenge of Bimodal Juxtaposition
Why do standard content moderation tools fail when encountering a meme? The answer lies in what researchers call the implicit and nuanced nature of multimodal communication.
Consider a meme analyzed by Smriti Singh et al. in “Female Astronaut: Because sandwiches won’t make themselves up there”: Towards Multimodal misogyny detection in memes. The text contains a standard trope about making sandwiches, which a text classifier might flag as mildly disrespectful or entirely benign depending on context. The image is a standard photograph of a female astronaut—completely innocuous on its own. A unimodal system inspecting either asset individually would find nothing actionable. However, when layered together, the image subverts the text (and vice versa) to deliver a deeply misogynistic message wrapped in sarcasm.
As highlighted by Ziyang Luo et al. in their evaluation benchmark GOAT-Bench : Safety Insights to Large Multimodal Models through Meme-Based Social Abuse, this reliance on implicit metaphors creates a massive blind spot for AI safety. Because the harmful meaning is emergent rather than explicit, models must be capable of cross-modal reasoning to detect the underlying abuse.
Architectures for Multimodal Fusion
To capture this emergent meaning, researchers have developed several distinct strategies to fuse visual and textual features.
Late Fusion and Ensembling
The most straightforward approach is to extract independent embeddings from state-of-the-art unimodal encoders (such as BERT or RoBERTa for text, and ViT or ResNet for images) and combine them at the final classification layers. For instance, Da Li et al. in AMS_ADRN at SemEval-2022 Task 5: A Suitable Image-text Multimodal Joint Modeling Method for Multi-task Misogyny Identification utilized multi-stage training frameworks to optimize separate text and image models before ensembling their predictions. While computationally efficient, simple late fusion often struggles when the relationship between text and image is highly sarcastic or indirect.
Attention-Driven and Bilinear Fusion
To force the model to learn how specific words interact with specific visual regions, researchers employ intermediate fusion mechanisms. Mayukh Sharma et al. demonstrated in R2D2 at SemEval-2022 Task 5: Attention is only as good as its Values! A multimodal system for identifying misogynist memes that cross-attention mechanisms can dynamically weight the interactions between modalities, forcing the text embeddings to attend to relevant visual cues.
Going a step further, Pranaydeep Singh et al. introduced the use of Multi-modal Compact Bilinear (MCB) Pooling in LT3 at SemEval-2021 Task 6: Using Multi-Modal Compact Bilinear Pooling to Combine Visual and Textual Understanding in Memes. MCB pooling allows the network to compute a higher-order outer product of the visual and textual feature vectors without the exponential computational cost, capturing intricate, multiplicative interactions between the two data streams.
[Text Tokens] ---> [Text Encoder] ---\
---> [MCB Pooling / Cross-Attention] ---> [Emergent Meaning Representation]
[Image Pixels] ---> [Vision Encoder] ---/
Joint Vision-Language Models
Instead of forcing two separate models to talk to each other, joint architectures map text and vision into a single shared semantic space from the beginning. Gustavo Lorentz et al. in INF-UFRGS at SemEval-2022 Task 5: analyzing the performance of multimodal models benchmarked several of these unified architectures (such as ViLBERT and VisualBERT), proving that early, deep cross-modal alignment is highly effective for detecting subtle misogyny. Similarly, Ayme Arango et al. leveraged contrastive language-image pre-training in HateU at SemEval-2022 Task 5: Multimedia Automatic Misogyny Identification, showing that models pre-trained on massive image-text pairs (like CLIP) possess a strong baseline intuition for how visual concepts align with textual descriptions.
Disentangling Visual and Semantic Roles
A key breakthrough in making these fusion models work is guiding them to look at what matters inside the image. Raw pixel arrays contain a lot of noise; models perform better when they can isolate semantic entities.
Jason Ravagli et al. addressed this in JRLV at SemEval-2022 Task 5: The Importance of Visual Elements for Misogyny Identification in Memes by incorporating an explicit object detection step (using Mask R-CNN) prior to multimodal fusion. By identifying the presence of people or specific objects, the downstream VisualBERT model could ground its textual analysis in concrete visual entities rather than abstract global image features.
Once entities are identified, the model must understand their narrative function. Propaganda and hate speech often rely on framing individuals or groups as targets. To capture this, Shaik Fharook et al. introduced a semantic role labeling approach in Are you a hero or a villain? A semantic role labelling approach for detecting harmful memes.. By training models to explicitly categorize meme subjects into roles like "hero," "villain," or "victim," the system maps out the underlying persuasive or harmful narrative structure of the meme.
Navigating Cultural and Linguistic Nuances
The challenge of meme analysis is further compounded by language and culture. Memes are highly localized, frequently employing regional dialects, code-mixing, and culturally specific metaphors.
In multilingual and code-mixed environments, standard English-centric models fail completely. Gitanjali Kumari et al. tackled this in CM-Off-Meme: Code-Mixed Hindi-English Offensive Meme Detection with Multi-Task Learning by Leveraging Contextual Knowledge, demonstrating that detecting offense in Hindi-English (Hinglish) memes requires specialized multi-task learning frameworks that explicitly extract fine-grained linguistic features alongside external contextual knowledge.
For low-resource regional languages, standard deep learning layers can struggle to optimize with limited data. To bridge this gap, Manasha Arunachalam et al. introduced Kolmogorov-Arnold Networks (KANs) to the problem space in Misogynistic Meme Detection in Dravidian Languages Using Kolmogorov Arnold-based Networks. By replacing fixed linear weights with learnable univariate functions on edges, these networks more effectively integrated multimodal embeddings for low-resource Tamil and Malayalam memes.
Finally, the field is moving away from simple binary classification ("hateful" vs. "not hateful") toward explainability. In BANMIME : Misogyny Detection with Metaphor Explanation on Bangla Memes, Md Ayon Mia et al. utilized Chain-of-Thought (CoT) fine-tuning on Large Multimodal Models. This approach forces the model to generate a step-by-step textual explanation of the meme's underlying visual metaphors before outputting a final classification label. This not only improves accuracy in culturally conservative contexts but also provides human moderators with a clear rationale for why a specific multimodal juxtaposition is harmful.
Key Foundational Work
Racist or Sexist Meme? Classifying Memes beyond Hateful (2021) — contributed 1 concept elements to this direction.
Deciphering Implicit Hate: Evaluating Automated Detection Algorithms for Multimodal Hate (2021) — contributed 1 concept elements to this direction.
Alpha at SemEval-2021 Task 6: Transformer Based Propaganda Classification (2021) — contributed 1 concept elements to this direction.