In Machine Learning, class imbalance is one of the most common challenges in classification tasks. This issue becomes even more pronounced in Natural Language Processing domains such as TASS, a Spanish-language emotion detection corpus characterized by a marked disproportion among categories. Traditional oversampling methods like SMOTE, based on k-nearest neighbors, lose effectiveness when applied to the high-dimensional spaces generated by modern language models. This work presents a probabilistic balancing framework that models the distribution of RoBERTa CLS embeddings (768 dimensions) using the covariance matrix estimated through the Ledoit–Wolf method, Lasso regression, and Elastic Net. From these distributions, realistic synthetic instances are generated for minority classes, drastically reducing the imbalance ratio without introducing semantic noise. The balanced embeddings are then classified using a lightweight multilayer perceptron (MLP), which eliminates the need for costly transformer fine-tuning. When evaluated on the TASS 2020 dataset, the best proposed algorithm achieved a Macro F1 score of 82.45%, outperforming the previously reported best result (55.3%) by 27.15% points [1]. These results demonstrate that synthetic generation guided by probabilistic models is an effective and computationally efficient alternative for emotion detection in high-dimensional, highly imbalanced scenarios.