EXIST 2026 International Challenge
Multimodal systems for TikTok sexism analysis, with three first-place evaluation results and a CLEF 2026 paper.
- Date
- September 2026
Sexism in a video can depend on what is said, what appears on screen, and how those signals fit together. As part of team AIdonnow, I built multimodal pipelines for EXIST 2026 that turn speech, on-screen text, and visual descriptions into representations that classification models can use.
Approach
- Combined Whisper speech transcription, OCR, and Qwen vision-language descriptions to reconstruct each video's content as text.
- Compared classical classifiers and multilingual encoders with Qwen3-8B fine-tuned using QLoRA.
- Evaluated physiological features from eye tracking, heart rate, and EEG, including the settings where they helped or hurt performance.
Impact
- Ranked 1st in Task 3.2 hard, Task 3.2 soft, and Task 3.3 hard evaluation for TikTok sexism analysis.
- First author of AIdonnow at EXIST 2026: Multimodal Text Reconstruction and Sensor Fusion for TikTok Sexism Detection, co-authored with Javier Martínez Llinares and published in the CLEF 2026 Working Notes on 4 October 2026.