The growing demand for sentiment analysis solutions that are accurate and computationally efficient highlights the need for continued innovation. Applications such as real-time customer service automation and social media analytics require models capable of scaling to large datasets while effectively handling complex and noisy emotional text, all under strict resource constraints. One promising direction for improvement is the integration of emotion-aware word embeddings. Although these embeddings have been shown to enhance the ability of BERT to capture nuanced emotional distinctions, their application to DistilBERT remains largely unexplored. This gap is significant given the increasing demand for models that balance accuracy and efficiency in practical applications.This thesis addresses this research gap by systematically evaluating the fine-tuning and integration of emotionaware word embeddings into DistilBERT. The aim is to improve its ability to detect subtle emotional distinctions while maintaining its lightweight, resource-efficient architecture. Using the GoEmotions dataset and integrating emotion-aware word embeddings into DistilBERT, this thesis evaluates two model configurations to improve emotional nuance recognition. Employing robust training and evaluation techniques, the thesis aims to improve performance and efficiency for real-world applications.