Robust NLP for Orthographically Unstable Historical Corpora

This direction is merged into another direction: historical orthographic variation. Its last version stays readable and citable here.

Version 1 of 135 papers41 conceptsfrozen 2026-10-04

Modern Natural Language Processing (NLP) assumes a world of standardized spelling. When a model sees the word "said," it maps it to a specific vector. But in a 17th-century English letter, that same word might appear as sayed, sed, or saide. This "orthographic instability" is what Claudine Moulin, Natalia Filatkina, and Richard Eckart de Castilho call the "pain in the neck" of computational linguistics in Analyzing Formulaic Patterns in Historical Corpora.

To unlock historical archives, researchers are developing techniques that treat non-standard spelling not as "errors" to be fixed, but as a complex mapping problem between high-variance historical signals and low-variance modern standards.

The Problem: The Fragility of Modern Tools

Most NLP pipelines—taggers, parsers, and lemmatizers—are trained on modern, edited text (like Wikipedia or news). As Marcel Bollmann notes in A Large-Scale Comparison of Historical Text Normalization Systems, even slight deviations in spelling cause these tools to fail.

This isn't just a matter of "old" words; it's a lack of a unified standard. Benoît Sagot highlights in Développement d'un lexique morphologique et syntaxique de l'ancien français that a single word in Old French might have multiple spellings within the same manuscript. This creates an explosion in vocabulary size that obscures the underlying morphological patterns.

Normalization as Machine Translation

The most common strategy to bridge this gap is Historical Text Normalization: mapping archaic variants to their modern equivalents.

Researchers have found success by framing this as a translation problem. In Revisiting NMT for Normalization of Early English Letters, Tanja Säily and colleagues use Neural Machine Translation (NMT) to "translate" 15th-century English into modern spelling.

The intuition here is that spelling variations often follow phonetic or regional patterns. A character-based NMT model can learn that a "y" in a certain context in 1600 often maps to an "i" today. To ensure these models don't hallucinate, researchers often filter the output through historical dictionaries like the OED or use weighted voting ensembles of different models, as seen in Normalizing Early English Letters to Present-day English Spelling.

Robustness Without Normalization

Normalization has a downside: it can erase the very linguistic diversity researchers want to study. Simon Gabay and Rachel Bawden argue in Ancien ou moderne ? Pistes computationnelles pour l’analyse graphématique des textes écrits au XVIIe siècle that these "writing systems" (scriptae) contain ideological and authorial signatures that modernization destroys.

Instead of changing the text to fit the model, some researchers change the model to fit the text:

The "Messy" Reality of Primary Sources

The ultimate goal of this research is to move away from "clean" edited editions and toward the "messy" reality of primary sources. Whether it is Ari Vesalainen and Filip Ginter extracting migration data from phonetic Finnish church records in Creating a Historical Migration Dataset from Finnish Church Records, 1800–1920, or Sarah Schulz and Jonas Kuhn distinguishing between OCR errors and intentional archaic spellings in Multi-modular domain-tailored OCR post-correction, the direction is clear: building NLP that is as flexible and context-aware as a human philologist.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_robust_nlp_for_orthographically_unstable_12b30347,
  title = {Robust NLP for Orthographically Unstable Historical Corpora},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-nlp-for-orthographically-unstable-historical-corpora/rdr_12b30347e30947889337e7830dff5ab9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_nlp_for_orthographically_unstable_12b30347_v1,
  title = {Robust NLP for Orthographically Unstable Historical Corpora},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-nlp-for-orthographically-unstable-historical-corpora/rdr_12b30347e30947889337e7830dff5ab9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}