Quasar@EEUCA 2026: Multimodal Deep Learning for Vaccine Stance Detection in Memes

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

The Quasar@EEUCA 2026 system addresses the challenging task of vaccine stance detection in multimodal memes, requiring interpretation of both textual and visual cues. Developed for the VaxMeme 2026 Shared Task at EEUCA 2026, the system employs a soft-voting ensemble approach. It combines DeBERTa-v3-large and RoBERTa-large for robust textual representation with CLIP (ViT-B/32) for integrated vision-language understanding. To overcome dataset limitations and class imbalance, the system incorporates domain-specific preprocessing, including random token deletion, image enhancement, and balanced class oversampling. Ablation studies identified balanced class oversampling as the most impactful component for performance improvement. The final system achieved a macro F1-score of 0.8306, securing 8th position among 25 participating teams, demonstrating the efficacy of ensemble-based multimodal learning for this specific detection task.

Key takeaway

For Machine Learning Engineers developing systems for social media content analysis, particularly for nuanced opinion detection in multimodal memes, prioritize ensemble-based approaches. Your models should combine strong text and vision-language representations, like DeBERTa-v3-large and CLIP. Crucially, implement balanced class oversampling to mitigate dataset imbalance, as this significantly boosts performance. This strategy can improve the accuracy of your stance detection systems, making them more robust against complex, imbalanced real-world data.

Key insights

Ensemble-based multimodal deep learning effectively detects vaccine stance in memes by combining text and vision models with data balancing.

Principles

• Capturing subtle semantic interactions is crucial for multimodal stance detection.
• Class imbalance significantly impacts performance in stance detection tasks.
• Ensemble models combining diverse representations improve multimodal task performance.

Method

A soft-voting ensemble combines DeBERTa-v3-large, RoBERTa-large, and CLIP (ViT-B/32). It uses domain-specific preprocessing, random token deletion, image enhancement, and balanced class oversampling.

In practice

• Apply balanced class oversampling to address dataset imbalance.
• Combine large language models with vision-language models for multimodal tasks.
• Use soft-voting ensembles for robust multimodal classification.

Topics

• Multimodal Deep Learning
• Vaccine Stance Detection
• Memes Analysis
• Ensemble Models
• DeBERTa-v3-large
• CLIP (ViT-B/32)

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• MLLMs Get It Right, Then Get It Wrong: Tracing and Correcting Late-Layer Textual Bias — MLLMs often correctly process visual information in early layers, but a "late-layer textual override" bias shifts final predictions towards text.
• I Know What You Meme, Even If it Emerged Today: Understanding Evolving Memes through Open-World Knowledge Acquisition — Meme understanding requires dynamic, open-world knowledge acquisition beyond static model parameters.
• Paying More Attention to Visual Tokens in Self-Evolving Large Multimodal Models — Self-evolving LMMs often under-condition visually; VISE uses invariance rewards to enforce spatial and semantic consistency.
• CORE: Conflict-Oriented Reasoning for General Multimodal Manipulation Detection — Multimodal misinformation's essence lies in intrinsic conflicts, which CORE detects by enhancing MLLMs for generalizable manipulation detection.
• Memes-as-Replies: Can Models Select Humorous Manga Panel Responses? — LLMs show promise in understanding social cues for humor, but struggle with visual information and subtle wit in meme reply selection.
• Multi-Modal Guided Multi-Source Domain Adaptation for Object Detection — Multi-modal inputs like depth and text can enhance multi-source domain adaptation for robust object detection.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.