A Transformer-based Approach of Normalization of Historical French Text

Historical French texts pose unique challenges for modern readers and NLP systems due to highly inconsistent spelling and scarce standardized resources. This paper presents a system for the normalization of early modern French, focusing on enhanced corpus construction and customized sub-word tokenization. Existing aligned corpora were combined with a newly curated dataset based on La Gazette, a rich and stylistically coherent periodical, to provide greater coverage and diversity. To address the orthographic variation typical of the period, a custom sub-word tokenizer was trained to better capture morphological patterns, supporting a Transformer-based sequence-to-sequence model. The approach demonstrates how tailored data preprocessing and tokenization improve the accuracy and robustness of automatic normalization. This work contributes valuable resources and methods for processing historical French and lays the groundwork for broader applications in digital humanities and historical linguistics.