Emotion Recognition in Social Networks Using Large Language Models

Abstract

With the rapid development of the internet, people have become increasingly reliant on social media platforms, generating vast amounts of social network data. Among these, sentiment data stands out as a particularly valuable category. By analyzing sentiment data, it is possible to predict the evolution of social network opinions and promptly monitor users emotional changes for early intervention. This paper discusses various large language models from the perspective of reducing sentiment classification categories, evaluating their performance through core metrics including accuracy, precision, and recall. Through seven different approaches—such as adding output linear layers, enhancing text recognition capabilities, implementing three-level joint prompt structures, and integrating COMET knowledge reasoning—this paper fine-tuned large language models to achieve more precise identification of emotional data in social networks. The optimized models demonstrated improved overall performance. This comprehensive review of large language models in social network sentiment analysis provides scholars with actionable insights for better fine-tuning strategies in this field.