Monitoring Online Misinformation Using LLM-Based Models​

Sarah Ainani

Co-Presenters: Individual Presentation

College: Hennings College of Science Mathematics and Technology

Major: BS.COMPUTER/SCI

Faculty Research Mentor: Martin Kollapally, Navya  

Abstract:

Social media platforms have become an essential arena in everyday life, users share a varying array of emotional content, personal struggles, and mental health reflections on social issues. With the ongoing increase in emotional discourse online, there has become a need and a market to develop tools that can accurately depict the intensity of emotional expressions. This research looks into how natural language processing (NLP) can be used to detect and analyze emotional content in user generated text platforms like reddit and twitter. Many users also express emotions indirectly, sarcasm, humor, and cultural references, or memes which can further confuse or cause a struggle for the model to understand. Additionally users may mask true emotions with languages that appear neutral or come off as contradictory. The study will investigate how natural language processing can be leveraged to detect and analyze emotional content to developed an improvement to technical agents when working with human language.This study theorizes that a content-aware trained LLM-based model, given multi-label emotion datasets, will significantly outperform traditional models with no prior knowledge or training. The intention is to explore primary and secondary emotions as well as ambiguity to better understand and mimic human nature. A Dataset with over 1000 social media posts was compiled from Reddit and Twitter. Each post underwent manual annotations that include three emotional scales: primary emotion, secondary emotion, and emotional obscurity. Posts were also flagged for their relevance to mental health, and common societal concepts like couch hopping and burnout. The research team utilized these annotated notes to train a BERT-based model. The model's performance was compared to different Hugging Face models based on classifiers.