Ancien ou moderne ? Pistes computationnelles pour l’analyse graphématique des textes écrits au XVIIe siècle

Simon Gabay, Philippe Gambette, Rachel Bawden, Benoît SagotLinx 2022 · DOI

Abstract: L’abandon des systèmes graphiques au profit de l’orthographe contemporaine dans l’écrasante majorité des éditions de textes du xviie siècle a fait disparaître leur richesse graphématique. Cette dernière est logiquement restée en grande partie méconnue, et donc sous-exploitée, en dépit de son rendement interprétatif certain. En privilégiant une approche pratique, sur corpus, plutôt que théorique, et en s’appuyant sur une recatégorisation des différents systèmes concurrents à cette époque en scriptae françaises, nous nous proposons de poser les bases d’une étude scriptométrique de la langue classique, s’intéressant à l’analyse de documents singuliers, manuscrits comme imprimés.

Introduction: The Orthographic Gap

When we read the masterpieces of the 17th century today, we almost exclusively encounter them in modernized French spelling. While this accessibility is vital for general readers, it creates a significant "orthographic gap" for researchers. In the 1600s, French spelling was not standardized; it was a battlefield between the "Ancients," who favored etymological, Latin-based spellings to preserve tradition, and the "Moderns," who pushed for phonetic simplification and clarity.

Table 1: Comparison of spelling conventions between the 'Ancients' and 'Moderns'.

Manual analysis of these original spellings is labor-intensive and often ignored in modern scholarship. This paper introduces a "scriptometric" approach to bridge this gap, using computational tools to analyze the complex graphematic richness of original documents without losing the nuance of the era's ideological shifts.

Methods: Scriptometry as Data Science

The researchers propose a workflow that treats spelling variation as a data science problem, utilizing three primary techniques:

  1. Alignment-Based Approach (ABA): By aligning original "diplomatic" transcriptions with modernized versions at the character level, the system automatically extracts transformation rules. For example, comparing estoit (old) to était (modern) allows the system to identify rules regarding the removal of the 's' and the addition of accents.
  2. Normalization Models: Using the MOSES toolkit, the team trained statistical machine translation models to "translate" 17th-century French into its contemporary equivalent.
  3. Principal Component Analysis (PCA): To interpret thousands of variations, PCA is used to cluster texts. This mathematical technique maps complex spelling data into a 2D space, revealing which texts share similar "spelling DNA."

Case Study: Bossuet’s Ideological Orthography

A key insight of the study comes from the analysis of manuscripts by the orator Jacques-Bénigne Bossuet. By tracking specific words across documents spanning 1648 to 1660, the researchers identified a surprising evolution in his writing habits.

Table 2: Orthographic variations in Bossuet's manuscripts across different years, showing the shift from modern to ancient forms.

As a young student, Bossuet often used "Modern" phonetic spellings (e.g., meme). However, as he rose through the church hierarchy and sought to project traditionalist authority, he reverted to the "Ancient" style (e.g., mesme). This "orthographic U-turn" suggests that spelling was a deliberate tool for social and political positioning.

The Printer’s Shadow and Composite Books

The study also examined Boileau’s Œuvres diverses (1674). PCA visualizations demonstrated that the volume was not an orthographic monolith. Reprints of older satires retained their original, conservative spellings, while new works published for the first time in the same volume adopted more modern conventions. This proves that a single book can contain multiple chronological and technical layers of spelling, reflecting the habits of different printing houses or periods.

Key Results and Future Work

The automated normalization model achieved a remarkable 97.4% lexical accuracy, identifying approximately 70 distinct transformation rules that define the transition to modern standards. The system successfully clustered texts based on their publication history, proving that machines can handle the "chaos" of pre-standardized French with precision.

While the current system relies on high-quality training data like the FREEMnorm corpus, future work aims to move toward "unsupervised" detection. This would allow the machine to identify patterns without needing a modernized version for comparison, ultimately helping editors create digital editions that are both accessible and faithful to the original linguistic history.

Summary updated

Abstract

L’abandon des systèmes graphiques au profit de l’orthographe contemporaine dans l’écrasante majorité des éditions de textes du xviie siècle. a fait disparaître leur richesse graphématique. Cette dernière est logiquement restée en grande partie méconnue, et donc sous-exploitée, en dépit de son rendement interprétatif certain. En privilégiant une approche pratique, sur corpus, plutôt que théorique, et en s’appuyant sur une recatégorisation des différents systèmes concurrents à cette époque en scriptae françaises, nous nous proposons de poser les bases d’une étude scriptométrique de la langue classique, s’intéressant à l’analyse de documents singuliers, manuscrits comme imprimés.
BibTeX
@inproceedings{gabay2022ancien,
  author = {Simon Gabay and Philippe Gambette and Rachel Bawden and Benoît Sagot},
  title = {{Ancien ou moderne ? Pistes computationnelles pour l’analyse graphématique des textes écrits au XVIIe siècle}},
  booktitle = {Linx},
  year = {2022},
  doi = {10.4000/linx.9346},
}
Figure