New RNN method decodes medieval abbreviations with 50% less errors

Anguelos Nicolaou, Maria Pia Tiseo, Tamas Kovacs, Nicolas Renet, and Georg Vogeler have published a paper titled 'Letter Lemmatization: One-to-one and Banded RNNs for Reversing Character-Set Simplification and Abbreviation in Medieval Text,' accepted for the ICDAR 2026 workshop VINALDO. The work tackles the challenge of fluid character sets in digitized medieval corpora, where transcribers used inconsistent abbreviations and simplified scripts. The authors propose two character-level RNN approaches: a one-to-one mapping network trained with self-supervision that can undo character simplifications using as few as 20 text lines, achieving a 50% reduction in character error rate without handling insertions or deletions. This is particularly effective for post-correction of handwritten text recognition (HTR) output.

The second contribution is a Banded RNN architecture that leverages character-level alignment ground truth from parallel corpora (same text in different character sets) to expand abbreviations in medieval charter transcriptions. The researchers also define a novel heuristic that computes semantic similarity between characters from any two arbitrary character sets, calling the overall process 'letter lemmatization.' They provide a rich Python library that efficiently implements all methods, enabling other scholars to apply these techniques to similar historical document processing tasks.