Uncertainty Estimation for Complex Text Detection in Spanish

Text simplifcation refers to the transformation of a source text aiming to increase its readiblity and understandability for a specific target population. This task is an important step towards improving inclusivity of such target populations (i.e., low scholarity or visually/hearing impaired groups). The recent advancements in the field brought by Large Language Models improve the performance of machine based text simplification approaches. However, using Language Models to simplify large text segments can be resource demanding. A more simple model to classify whether the text segment is worth to simplify or not can improve resource efficiency, in order to avoid unnecessary text prompts to the Large Language Models. Furthermore, text simplicity categorization can also be used for other purposes, such as text complexity measurement. The discrimination of text segments into simple and complex categories might lead to a number of false positives or negatives for a not well-tuned model. A way to control the acceptance threshold, is the implementation of an uncertainty score for each prediction. In this work we explore two simple uncertainty estimation approaches for complex text identification: a Monte Carlo Dropout and an Deep Ensemble Based approach. We use an in-house dataset in the financial education domain for our tests. We calibrated the two implemented methods to find out which performs better, using a Jensen-Shannon based distance between the correct and incorrect outputs of the discriminator. Our tests showed an important advantage of the Monte Carlo Dropout over the Deep Ensemble Based method.