Using GPT-3 as a Text Data Augmentator for a Complex Text Detector

In this work, we explore the problem of complex text detection. This problem is a frequent challenge when implementing text simplification pipelines. Identifying complex text segments can trigger text simplification models, making a better resource usage as state of the art Large Language Models are expensive to use. We focus in Spanish, as it is an under-represented language, given the lack of simple/complex paired datasets. We use a novel paired dataset in Spanish of financial educational texts to train and test our methods. To improve the performance of the classifier, we propose the usage of text simplifications generated with GPT-3 (data augmenter) to alleviate the need to label a large number of text segments as simple or complex. We use the BERT pre-trained model on Spanish data known as Spanish BERT (BETO) and explore the effect of augmenting target data in the model performance.