May 2023 – Aug 2023 · 3 mos · Remote

• Objective
• Create a novel multimodal model to analyze sentiment and emotion roles in tweeting rumors vs. non-rumors
• Approach
• Created a multimodal dataset of images and tweets, cleaned and preprocessed it using regex and other libraries
• Used BERT for text and ResNet50 for image features, followed by manual annotation for emotion and sentiment
• Designed a model with BiLSTMs, GRL and novel Cross Fusion Module for integrating multimodal outputs
• Introduced new novel Inter-segment and Contextual Intermodal Attention modules within the base model
• Conducted analysis of related research papers and efficient prompting techniques for enhancing LLM results
• Actively developing LLM-integrated model for Fake News, harnessing its power and introducing common sense
• Result
• Developed an innovative model for effective sentiment and emotion extraction from rumor or non-rumor, integrat-
• ing common sense, and actively developing a novel multimodal framework where LLMs are an integral component