Linguagem simples com LLMS: avaliação de simplificação textual no contexto administrativo da UFC Inova

Abstract

Plain Language consists of a set of practices to write easier to read public documents, aiming to expose relevation information quicker, and also promote accessiblity and transparency. In 2025, the Brazilian government sanctioned the law n° 15.263, which establishes the National Policy of Plain Language. The UFC Inova institution already publishes plain language versions of some of Its documents since 2021, but they are written manually, requiring extra time and work of the colaborators. Using that backgroudn, this work discuss the use of Generative Artificial Inteligence (GenAI) to perform text simplification, comparing the performance of proprietary and open Large Language Models (LLMs). The evaluation used consolidated readability indexes, morphosyntactic metrics related to text complexity and semantic similarity metrics on the generated texts. Two iterations of text simplification were performed, in the first one, the best proprietary and open models were Gemini 2.5 Flash and Qwen2.5:14b, respectively, but Gemma3:4b and Gemini 2.5 Pro had the best semantic similarity, with BERTSCores of 0.826 and 0.760. After the audience evaluation in an online survey, the process was adapted to perform summarization and simplification jointly. The results of this second iteration show the efficiency of summarization combined with simplification, having 7 out of 9 models with significant improvements on the average sentence size and Qwen3:14b having a performance on par with Gemini 2.5 Pro’s. Generated texts were compared with the existing simplified version and the models with the best results were, again, Gemini 2.5 Pro and Gemma3:4b, both with a 0.665 BERTScore, the similarity scores were lower in general, sitting around 0.622 and 0.684.