When we read a classical poem or a medieval treatise, we often assume we are looking at the author’s original words. In reality, we are looking at the end of a centuries-long "telephone game." Before the printing press, every copy of a text was handwritten by a scribe who might be tired, poorly lit, or confused by the script in front of them.
The research direction shared by philologists like Thomas Gärtner and machine learning researchers like Charlie Cowen-Breen is the systematic "debugging" of these historical errors. By modeling the physical and cognitive mechanics of how scribes made mistakes, researchers can reverse-engineer the original text from its corrupted remains.
The Mechanics of Corruption
Textual corruption is rarely random. It follows specific patterns dictated by the physics of the page and the limitations of human vision. One of the most common mechanisms is palaeographic confusion, where a scribe misinterprets visually similar characters.
In Ein Textvorschlag zu Ov. met. 1.591, Thomas Gärtner demonstrates how a single vertical stroke following a vowel could be misread, transforming a sharp, logical word into a generic adjective that fits the meter but weakens the narrative. Similarly, in Vier konjekturen zu griechischen romanen, Gärtner shows how these visual slips can replace high-impact verbs with mundane ones, effectively "muting" the author's original rhetorical intent.
Another mechanical error is eyeskip (or haplography). This occurs when a scribe’s eye jumps from one word to a similar-looking word a few lines down, causing them to accidentally omit the text in between. By identifying these "visual anchors" in a manuscript, researchers can hypothesize exactly what was lost, as seen in Gärtner’s work on Theognis and Lysias.
The Problem of Plausible Substitutions
The most difficult errors to detect are not the ones that result in gibberish, but the ones that make sense. Medieval copyists were not just machines; they were readers. When they encountered a passage they didn't understand, they often "corrected" it into something contextually plausible but fundamentally wrong.
As Karthik Narasimhan and Barbara Graziosi argue in An Annotated Dataset of Errors in Premodern Greek and Baselines for Detecting Them, these human-driven modifications create a layer of "noise" that is much harder to distinguish from the original text than modern digital typos. This is why standard Large Language Models often struggle with philology—they are trained to prefer what is "likely," whereas the original poetic intent might be rare or idiosyncratic.
Conjectural Emendation as Probabilistic Modeling
To fix these errors, philologists use conjectural emendation. This isn't guesswork; it is a form of probabilistic modeling. A successful conjecture must satisfy two conditions:
- Internal Logic: The new word must fit the meter, grammar, and cultural motifs of the time (e.g., the heart as a writing tablet in Textkritische Bemerkungen zu erotischen Epigrammen der Anthologia Palatina).
- Transcriptional Probability: The researcher must explain how the error happened. A fix is only convincing if you can show the visual or phonetic path from the original word to the corrupted one.
This same logic is now being digitized. The Logion framework, developed by Charlie Cowen-Breen and colleagues, uses BERT-based architectures to identify these gaps and suggest restorations. By training on massive datasets of ancient Greek, the model can spot where a word "feels" wrong based on the surrounding context, acting as a high-speed assistant to the human philologist.
Identifying Intentional Layers
Not all changes were accidental. Some were interpolations—marginal notes or ideological "corrections" that were later integrated into the main text.
Philip Payne highlights a fascinating physical example of this in Vaticanus Distigme-obelos Symbols Marking Added Text. He identifies specific marginal symbols in the Codex Vaticanus used by ancient scribes to flag passages they suspected were later additions. This suggests that even 1,600 years ago, copyists were acting as "gatekeepers" of textual integrity.
In modern contexts, this same problem appears as "simplified" editions of historical texts that subtly insert modern biases. Slobodan Ilić demonstrates in Tursun Bey Lost in Translation how these editorial interpolations can transform historical records into national myths, emphasizing the need for a systematic way to audit the integrity of non-scholarly editions against original manuscripts.
The Future: Human-in-the-Loop Restoration
The research suggests that the most effective way to restore lost literature is a "human-in-the-loop" approach. While machines like the ELECTRA-based discriminator or the Type-R pipeline can flag discrepancies at scale, the final judgment requires the deep contextual knowledge of a philologist. By combining the computational power to model every possible scribal slip with the human ability to recognize poetic genius, we can begin to see through the "noise" of history to the original intent of the authors.