Detection of Persuasion Techniques in Memes - Multimodal Deep Learning System

Developed a Multimodal Deep Learning system to identify 22 distinct persuasion techniques in memes using the SemEval-2021 Task 6 dataset. Employed transformer-based models for text feature extraction and CNNs for image analysis. Integrated attention mechanisms for feature fusion and optimised the model using Focal Loss to handle class imbalance. Achieved comprehensive evaluation using accuracy, micro and macro F1 scores, focusing on enhancing performance for minority classes.

Tools: Python, TensorFlow, Keras, Transformers, Pre-trained ML Models