Subclassifying Sadness: Enhancing Sentiment Analysis in Social Media with Large Language Model
Naseef Muhamed, Anju S. Pillai · 2025
This study addresses the underexplored area of secondary classification in sentiment analysis, focusing specifically on subclassifying negative sentiments, with an emphasis on "sad" sentiments. The primary objective of this work is to enhance the sentiment analysis by using the Large Language Model (LLM) to classify English language social media posts into three main emotional categories: happy, neutral, and sad. Additionally, the posts classed as "sad" are further classified into two categories: risky and non-risky. A BERT (bert-base-uncased) LLM is utilized in this work. Detecting risky posts can be highly beneficial, as they may indicate possible signs of suicidal thoughts, anxiety, or depression. Social media posts are used as the dataset for this work, which is analyzed to determine how well the LLM classifies sentiments and their subtypes. The proposed system achieved an accuracy of 92.5% in secondary classification using just six epochs and a reduced dataset of 1,200 posts, highlighting the effectiveness of using a pre-trained LLM models and fine-tuning it. The findings from this study emphasize the potential for timely interventions in mental health scenarios and demonstrate the significant impact of distinguishing risky from non-risky messages in real-time applications. Overall, by providing an effective method for subclassifying the "sad" emotion, this study enhances the sentiment analysis approaches, improving the understanding and response to negative emotional expressions in social media conversations.