Automatic Text Simplification (ATS) aims to make texts easier to read while still preserving their meaning. Large Language Models (LLMs) have become the predominant approach for this task, yet most work has focused on a few domains, such as Wikipedia, news and medical texts. The Public Administration (PA) domain, covering legal notices, council documents, and online government portals which citizens access daily, has been largely left unexplored, especially in languages other than English. This project presents a systematic exploration of ATS on PA texts across three corpora: SimPA (English), Admin-It-L2 (Italian) and ClearSim (Spanish), addressing both sentence and paragraph-level text.

A series of experiments covering a multitude of open-source LLMs of varying size and training data examined custom domain-specific prompts, in-context learning via few-shot prompting, and cross-lingual prompting. Simplifications were evaluated using token-based metrics (SARI, ROUGE, BLEU and BERTScore) and the suitability of other metrics such as LENS and readability scores was reviewed. Further depth of simplification quality was found through qualitative analysis of LENS and model simplifications.

Findings highlight clear strategies and selections for future work. The Qwen2.5-7B model excelled and outperformed other models across all three languages. The project reveals drawbacks of token-based metrics and how prompts and few-shot counts can drastically influence sentence-level simplification. On the other hand, LENS showed promising correlation with human judgement compared to standard metrics. The project also demonstrates the contrast of ATS for different textual types, where paragraph-level simplifications behave less consistently than sentence-level experiments. Together, these findings build a strong foundation for future ATS work in the PA domain.