Social media platforms now have a significant impact on public opinion because they provide a wealth of user-generated content that captures emotions, viewpoints, and fascinating topics. This study presents a novel framework for analyzing emotional trends on social media in two phases: Optimized transformer models, such as DistilBERT for topic classification, RoBERTa for sentiment analysis, BERTweet for emotion detection, and BERTopic for keyword extraction, are included in Phase 1. Phase 2 focuses on refining these models through instruction-based fine-tuning, boosting their understanding of language and context on social media. By employing these models, we intend to determine which specific themes evoke distinct emotions and explore the connection between topics and emotions. Our method offers more profound understandings of social media sentiment and has applications in crisis management, political debate, and brand monitoring. The shortcomings of traditional models are addressed by this approach, which provides a more advanced understanding of public opinion in the digital age.