Generating Multi-Level Accessible Dataset Descriptions for German Open Data Portals Using Local Language Models
Abstract
Open data portals face a persistent challenge: datasets and their metadata often remain incomprehensible to non-specialist audiences, limiting the democratic potential of open government data. While recent advances in large language models offer promising approaches to automated description generation, their application to multi-level accessible descriptions of structured datasets, particularly in non-English contexts, remains underexplored. This paper presents the design, implementation, and preliminary evaluation of a modular pipeline that generates dataset descriptions at three accessibility levels using locally deployed language models. The system integrates deterministic dataset profiling (Frictionless Data and ydata-profiling) with generative AI, employing Qwen3.5-9B for description generation and an initial two-model evaluation setup for automated quality assessment (Prometheus 2 and Mistral 7B), which was later critically examined against human expert judgement. We address practical constraints of public sector deployment, including reproducibility, digital sovereignty through local model execution, and alignment with operationalized criteria derived from German accessibility standards. Our implementation processes datasets from GovData.de, generating descriptions conforming to DIN 8581-1 (Plain Language German), DIN SPEC 33429 (Easy Language German), and Standard German for general public audiences. This work contributes both a reusable technical architecture and methodological insights for accessible metadata generation in open data ecosystems, with particular attention to cross-lingual prompting strategies and the limitations of the DCAT-AP.de metadata standard.
