Multimodal Fusion for Harmful Content Detection

18 concepts17 papers1 proposal

On social media, harmful content rarely presents itself in a single, easily identifiable format. Consider an internet meme: it might pair a completely benign image of a household appliance with a seemingly innocent text caption. Separately, a text classifier and an image classifier would flag neither. But when a human views them together, the juxtaposition reveals a deeply misogynistic or propagandistic message.

This phenomenon is known as emergent meaning—the harmful intent is not contained within the text or the image alone, but surges from the interplay between them. To detect such content, machine learning models must move beyond isolated unimodal analysis and instead perform multimodal fusion, integrating diverse streams of information to capture their combined context.

The Failure of Unimodal Analysis

Traditional content moderation relies heavily on text-based filtering or isolated image analysis. However, as the authors of MMVAE demonstrate, unimodal baselines struggle significantly with internet memes because they miss the semantic bridge between the visual and textual components.

If a model analyzes only the text extracted via Optical Character Recognition (OCR), it lacks the visual grounding necessary to understand irony, sarcasm, or targeted abuse. Conversely, if it only looks at the image, it misses the explicit framing provided by the text. To close this gap, researchers have shifted toward architectures that explicitly model the cross-modal interaction.

Mechanisms of Multimodal Fusion

How do models actually combine these distinct modalities? The research landscape reveals three primary strategies for fusion: feature concatenation, joint latent spaces, and structural graphs.

1. Feature Extraction and Late Fusion

The most straightforward approach is to pass the text and image through separate, state-of-the-art pre-trained models, extract their respective embedding vectors, and combine them before classification.

[Image] ---> [Vision Encoder (e.g., ResNet/Swin)] ---\
                                                      ---> [Fusion Layer] ---> [Classifier]
[Text]  ---> [Text Encoder (e.g., BERT)] ------------/

For instance, Hossan et al. and Ratul et al. utilize deep convolutional networks like ResNet alongside BERT variants to extract high-level representations, fusing them via dense layers. Similarly, Mahadevan et al. leverage contrastive models like CLIP to ensure that the extracted visual and textual features already share a aligned semantic space before being fed into downstream classifiers like Support Vector Machines (SVMs).

2. Joint Latent Space Modeling

Simple concatenation can sometimes fail to capture deeper, non-linear correlations between modalities. To address this, Gu et al. introduced a Multi-modal Multi-task Variational Autoencoder (MMVAE). Instead of just gluing features together, the architecture learns a shared, coordinated latent distribution that co-represents both visual and textual features, allowing the model to capture the nuanced interaction space where harmful meaning emerges.

3. Graph-Based and Non-Linear Interplay

Other approaches treat the components of a meme as interconnected entities. Srivastava proposed leveraging Graph Neural Networks (GNNs) to model the relationships between textual tokens and visual objects as a network, framing misogyny detection as a graph classification problem.

To capture highly complex, non-linear interactions during fusion, Arunachalam et al. explored Kolmogorov-Arnold Networks (KANs), replacing traditional multi-layer perceptrons with learnable activation functions on the edges of the network to better map the subtle cross-modal dependencies.

The challenge of multimodal fusion compounds when moving away from English to low-resource or culturally nuanced languages. In regions speaking Dravidian languages like Tamil and Malayalam, online speech is frequently characterized by code-mixing (e.g., writing Tamil words using the Roman script) and highly localized cultural metaphors.

To tackle this, recent frameworks pair advanced vision transformers with specialized language models:

  • Sayma et al. combined MalayalamBERT with a Swin Transformer to capture localized textual context alongside visual cues.
  • Chakraborty et al. paired IndicBERT with ResNet18, demonstrating that multimodal fusion is uniquely effective when the offensive message is implicit and relies entirely on cultural context.

As highlighted by Chakravarthi et al. in their shared task findings, these integrated architectures consistently outperform unimodal alternatives, proving that visual elements provide the critical anchor text-only models lack in under-resourced linguistic environments.

Beyond Text and Images: Content and Network Structure

The core intuition of multimodal fusion—that malicious intent hides in the intersection of different data types—extends beyond memes. Online information operations (such as coordinated propaganda campaigns) rarely rely on a single deceptive post. Instead, their impact relies on the interplay between what is being said and how it is being spread.

[User Content Embeddings] -------\
                                 ---> [Graph Foundation Model] ---> [Operation Detection]
[Network Structure Embeddings] --/

To counter this, Minici et al. developed IOHunter, a framework that fuses user content embeddings with network structure embeddings. By evaluating both the textual semantic signal and the social network topology simultaneously, the model exposes coordinated manipulation that would appear completely organic if analyzing the text or the network in isolation.

Whether combining pixels with text or words with network graphs, multimodal fusion works because it mirrors human comprehension: it evaluates the message not in a vacuum, but within its full, emergent context.

Key Foundational Work

MMVAE at SemEval-2022 Task 5: A Multi-modal Multi-task VAE on Misogynous Meme Detection (2022) — contributed 2 concept elements to this direction.
UMUTeam at SemEval-2022 Task 5: Combining image and textual embeddings for multi-modal automatic misogyny identification (2022) — contributed 1 concept elements to this direction.
Poirot at SemEval-2022 Task 5: Leveraging Graph Network for Misogynistic Meme Detection (2022) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_multimodal_fusion_for_content_analysis_4cd0ca25,
  title = {multimodal fusion for content analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multimodal-fusion-for-content-analysis/rdr_4cd0ca25102e439d87670f20f3cc5f7e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_multimodal_fusion_for_content_analysis_4cd0ca25_v1,
  title = {multimodal fusion for content analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multimodal-fusion-for-content-analysis/rdr_4cd0ca25102e439d87670f20f3cc5f7e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake