A Normalisation Layer for Old Turkic Text Editions in OCR-Based Workflows

For over a century, Old Turkic texts have been edited using divergent Latin-based transcription systems developed independently across German, Turkish, Russian, Japanese, French, and Chinese scholarly traditions. Because most of these editions survive only as printed volumes or PDFs, optical character recognition (OCR) is the primary route to digitisation, yet OCR does not resolve, and often amplifies, the transcriptional heterogeneity already embedded in the source editions. Even error-free character recognition would not yield interoperable text, since the incompatibility originates in editorial decisions rather than in the recognition process. This paper discusses a deterministic, representation-level normalisation framework that operates immediately after OCR extraction. The framework is additive, reversible, and never corrective. It preserves the raw OCR output unchanged in an orig layer and generates a canonical transcription in a reg layer using Text Encoding Initiative (TEI) P5 choice structures, with all philological interpretation frozen in static, edition-aware mapping tables rather than executed dynamically. Applied to 4,485 OCR-derived tokens from 13 editions published between 1919 and 2023, the approach allowed 94–97% of the tokens to be normalised using the defined mapping rules. By separating preservation from normalisation, it offers a reproducible basis for large-scale Old Turkic corpus construction and cross-edition search across different transcription traditions.