Grasping emotions in text, an important act in applications like sentiment analysis, as models need to recognize the deeper human feelings and the textual meaning. We experimented on large and small transformer-based language model, BERT-base-cased and DistilBERT-base-cased architectures for detecting emotion in sentiment analysis tasks require intense understanding of human emotions as well as textual interpretations. The dataset we selected contains 20,000 samples, 6 different emotion labels joy, fear, sadness, surprise, anger and love, has an imbalance across the labels. Fine-tuning has been done on both models, and the focus of evaluation is classification of emotions, where both models have been analyzed on their capacity to classify sentiments correctly. Our aim is to check whether larger models with higher counts of parameters can capture precise emotional content within text better than their smaller counterparts, and results suggest the BERT model was bit more accurate in comparison to DistilBERT. This class-imbalanced label problem is handled by placing more weight on the minority side and then fine-tuning is done. Though there is improvement in performance of models, still the results indicate that a large transformer-model BERT has an upper hand in the domain of emotion sentiment sensing than a small transformer-model DistilBERT, but smaller model DistilBERT performance is increased to almost close reach to the performance of larger model BERT; in fact the DistilBERT model outperformed the BERT model in classifying emotions with fewer samples, but BERT remained the overall outperformer.