On July 10, 2026, researchers Anguelos Nicolaou and colleagues published a paper titled Letter Lemmatization: One-to-one and Banded RNNs for Reversing Character-Set Simplification and Abbreviation in Medieval Text. The study addresses the challenges faced by medieval document transcribers due to inconsistent character-set practices and digitization policies.
Revolutionizing Character Mapping in Medieval Texts
The paper presents a novel approach to character mapping, focusing on one-to-one character mappings. The researchers trained character-level one-to-one Recurrent Neural Networks (RNNs) to reverse character-set simplifications. Remarkably, they achieved a recovery of half the Character Error Rate (CER) with just 20 lines of text.
This advancement is particularly significant for handwritten text recognition (HTR) post-correction, as the one-to-one networks demonstrated considerable improvements without accounting for insertions and deletions. This flexibility is crucial for the varied practices seen in medieval transcription.
Expanding Abbreviations with Banded RNNs
In addition to one-to-one mappings, the authors introduced a method using Banded RNNs, which operates with character-level alignment ground truth from parallel corpora. This technique successfully expands abbreviations found in medieval charter transcriptions, showcasing the versatility of their approach.
