Identifying offensive and non-Offensive content in multimodal se

: Proposed the task of identifying offensive content and the different levels of offensiveness (in the form of intensity)

and predicting emotion and sentiment (and their intensities) as well for multimodal setting.

◦ : Introduced a large-scale multimodal dataset, that has been manually annotated with offensive and

non-offensive, emotion, sentiment and their respective intensities.

◦ : Prepared a baseline multitask transformer model with early modal fusion for performing classification.

◦ : To enhance the performance of the model further exploration was done using YOLOv5 pretraining with a

custom dataset through continual learning for feature extraction from image

◦ : BERT with GCN along with contrastive learning for textual embedding

◦ : Designed a multimodal multi-task framework using BERT-GCN that concurrently identifies the offensive

content as the primary task while the emotion, sentiment as the secondary task, also designed a novel fusion

mechanism for capturing multimodal features