Transforming Digital Access in Parliament: Topic Modeling Meets Retrieval Augmented Generation

The Swiss Parliament is leveraging Artificial Intelligence (AI) to enhance efficiency, accessibility, and transparency in legislative processes, while ensuring AI deployment aligns with democratic values and legal standards through robust oversight mechanisms. This paper presents two pilot projects initiated by the Swiss Parliamentary Library in collaboration with the Bern University of Applied Sciences, focusing on the automatic indexing and intelligent information retrieval of parliamentary businesses. Each parliamentary business is assigned a set of 232 topics to facilitate searchability. The first pilot project focuses on automatic indexing of parliamentary businesses and involves the implementation of a topic modeling pipeline. This is achieved by training machine learning models for three Swiss languages (German, French and Italian) on a dataset consisting of parliamentary businesses originally labeled with topics by the Swiss Parliamentary Library staff beginning from 1991 to 2024. We achieved Macro-F1 scores exceeding 0.5 across all models and, through a user interface provided to Parliament staff, obtained additional feedback that informed adjustments to the confidence score, ultimately prioritizing recall over precision. In the second pilot project, a Retrieval Augmented Generation (RAG) system was developed to make the parliamentary businesses searchable in the three Swiss languages. An interface was created with a feedback mechanism to allow users to not only interact with the RAG system but also give direct feedback where the system works and where it fails. These findings were utilized to inform further refinements to the RAG system, thereby enhancing its overall performance. In both use cases - automated indexing and RAG - users prioritized recall over precision.