Feature Densities for Uncertainty Quantification for Complex Text Detection in Spanish

Text simplification is crucial for enhancing content accessibility, particularly for audiences with low literacy or sensory disabilities. Despite recent advances using large language models (LLMs), their computational expense and predominant control by private entities hinder practical deployment. Efficiently detecting complex text segments requiring simplification is thus vital for resource optimization. This work addresses uncertainty quantification for complex text detection in Spanish-an under-explored area -to improve model transparency and enable targeted retraining. We train a BETO-based classifier and compare three uncertainty-quantification techniques-MonteCarlo Dropout, Deep Ensembles, and a novel Feature Density Estimation that operates in latent space, leverages representations to model feature distributions, offering post-training computational efficiency. Experiments on a financial education dataset (5,314 text pairs) show that Feature Density Estimation matches Monte Carlo Dropout's performance (Jensen-Shannon distances: 0.42 vs. 0.43) at significantly lower computational cost, while Deep Ensembles underperformed (0.34). Statistical analysis confirms Feature Density Estimation as a lightweight, effective uncertainty quantification alternative for low-resource language applications.