Name in original language
semplficazione dei testi con intelligenza artificiale generativa
Initiative overview
Clear and accessible communication is essential for ensuring that citizens can effectively navigate public services. This work examines the use of generative AI models to simplify administrative texts on the INPS, Italy’s National Institute of Social Security, web portal, which provides nearly 500 digital services related to pensions and welfare. As Europe’s largest welfare institution, INPS serves approximately 45 million citizens, making the clarity of its service descriptions crucial for accessibility and user experience.
INPS has been running a structured simplification programme aimed at making its content understandable to users with a middle-school education level. To date, over 700 texts have been manually simplified, using the Gulpease index to measure readability. The average score improved from 57 to 64, reflecting a substantial increase in accessibility.
Currently, INPS relies on a manual process to simplify the language of its Service Pages, a time-consuming and resource-intensive approach that requires specialised expertise. The need for clearer bureaucratic language has been recognised in Italy for decades. A 2002 directive, still in force, states that “all texts produced by administrations must be designed and written to be understood by those who receive them.” Despite this, institutional communication often remains complex, limiting citizens' ability to fully understand and access essential services.
While simplification enhances transparency and public trust, a major challenge is maintaining the legal precision required for official documentation. This work explores how Large Language Models (LLMs) can support INPS in transforming its manual simplification process into a semi-automated system that is both efficient and scalable. The goal is to develop a solution that can process large volumes of text while ensuring readability, accuracy, and compliance with legal standards.
To achieve this, the research establishes an evaluation framework based on text metrics and conducts user surveys to assess the accessibility and quality of AI-generated text compared with original and manually simplified versions.
Looking ahead, the potential for refining and expanding this approach is substantial. Possibilities include generating pre-simplified texts directly from legal norms, introducing a compliance-evaluation framework, and developing collaborative workflows in which AI and human experts iteratively refine simplifications, combining automation with expert oversight.
The experimentation on text simplification has been developed as a prototype and has successfully passed the evaluation stage. It is now transitioning into an operational business process within INPS and its technology partners, integrating a graphical interface to support a human-in-the-loop collaborative workflow.
