This study presents an innovative approach to reconstructing historical thunderstorm and hail activity using documentary evidence. It combines a systematic, source-critical analysis with Transformer-based language models. A particular strength lies in the development of a structured classification framework that accounts for differences in source types, linguistic variability, and uncertainty while yielding physically plausible long-term patterns in convective weather observations. Applying ThunderstormBERT and HailBERT further demonstrates the potential of automated, language-based methods to efficiently extract meteorological information from large, heterogeneous historical archives. The analysis reveals the robustness of the reconstructed hail and thunderstorm signals, as well as the challenges associated with intermediate-intensity and winter events. Overall, the manuscript offers a reproducible framework for transforming historical textual evidence into quantitative climate information, providing a promising basis for extending the observational record of long-term convective weather variability.