Social media (YouTube) political sentiment multi-label analysis

In an era where presidential elections and political discourse are increasingly digital, understanding public sentiment through online comments becomes crucial. This project explores the relationship between YouTube comments and political preferences, challenging the null hypothesis that asserts no connection between contextless comments and candidate preference. Utilizing various natural language processing (NLP) tools, including Bing-Liu, Vader Sentiments, and Large Language Models (LLMs) like ChatGPT-4, we delve into multi-label sentiment analysis for political figures.
Our methodology encompassed a rigorous data collection process from YouTube, leveraging custom scrapers and the computational power of KSU's BEOCAT servers. We navigated through challenges like content limitations and the need for comment sanitization to comply with community guidelines. The study tested different models, including logistic regression and Graph Convolutional Networks (GCNs), against a baseline of Max Label/Zero R classification.
Results showed varying degrees of success, with individual label accuracies ranging from moderate to high. However, the overall accuracy of our final model using GCNs stood at 39%, indicating the complexity and difficulty of multi-label classification in political sentiment analysis yet also the success of graph convolutional networks at identifying complex contextless sentiment. This project not only sheds light on the potential of NLP in political analytics but also opens up avenues for future work in real-time sentiment analysis during political events, albeit with ethical considerations.
This research contributes to the growing field of NLP implementation in public opinion analysis, with implications extending beyond politics into other consumer-centric industries.