How To Process Handwritten Latin with AI OCR (Odds And Ends)
So you’ve been trying to do some historical research, but you’re vexed by the old-time habit of English people not writing in English? You’ve come to the right place. Highbrow scholars used to conduct their correspondence in the languages of the ancients, using so many extra words that it’d put Charles Dickens to shame, except Dickens very sensibly wrote in his native tongue, which gives him a real advantage here. 1600s London literati — like the usual suspect, Baldwin Hamey — didn’t do us that favor.
Fortunately, with the advent of AI and its processing capacities, there’s a halfway decent solution. Obviously, AI is not truly intelligent, and it should never be treated as a substitute researcher. If it “hallucinates” something false and fictitious, and you don’t catch it, that’s completely your fault. However, when using AI's OCR function as a simple document-processing tool, this method can massively improve information availability in unprecedented ways.
There are a few other caveats I’ll offer right away:
The process is subject to error. It can and does misread letters sometimes, just as any human might. It also lacks the capability to tell you when it thinks it’s wrong, and is much likelier to make something up to fill in the gap. Do not take its accuracy at exact face value. (Note: The less readable the handwriting, the more important this is. AI can, and will, just try to guess. Use common sense.)
This method should not be used for formally published papers without extensive human review and editing. I shouldn’t have to say this, but I know I must. Academics have spotted published papers in so-called credible sources (later retracted) that contain stock GPT response phrases. In this case, the telltale giveaway was “Certainly, here is,” which GPT regurgitates with its answers for the sake of politeness. Not a good look for anyone.
This method should be used to determine the relevance of material, NOT exact translation. In my case, I’ve been looking for any instances where Baldwin Hamey Junior mentioned painting or artists. Feeding his pages of Latin writing into GPT helps me rule out irrelevant information, such as his many pages of paeans to his fellow doctors taken from this earth too soon. GPT works wonderfully well for this purpose, and it’s reliable — it’s unlikely to tell me an outright false claim. However, if I were to quote verbatim a GPT English translation of the Latin text, and claim that it meant Hamey held some specific opinion based on his word choices, that would be a horrible idea. Don’t do that.
Results can be inconsistent in quality, and free processing is rate-limited. For this example, I used ChatGPT with a subscription (GPT 4o is the model that works best), then ran it again with free ChatGPT, just to be sure of my results. On the free account, it processed three pages of Latin before telling me “Upgrade to ChatGPT Plus or try again tomorrow after 11:46 AM.” (I started at 1:15 PM. Figures.) it also didn’t catch the inclusion of a Greek word on page 3, but when I pointed out the error, it corrected it capably. Like most humans, it generally does a better job the more you actually explain to it. The value of a subscription (as of July 2024, $20 a month) will probably depend on how much bulk processing you actually need to do. It was worth it for me — I’ve got hundreds of pages to contend with — but if you’ve only got a few pages, don’t worry about it.
With all that aside, here’s the step-by-step instructions, as promised.
1: Acquire an image of your Latin text page.
For this I’ve used the “Bustorum aliquot Reliquiae,” the compilation of Hamey’s obituaries to his fellow doctors. I will note that the copy of “Bustorum” I processed is a later re-transcription of Hamey’s words in a beautifully consistent semi-cursive font, not Hamey’s own challengingly indecipherable handwriting. (That one is an uphill battle, not for the faint of heart.)
The index to “Bustorum” shows that Theodore Mayerne’s biography is on page 313.
Skipping ahead to 313, we find that Mayerne’s biography starts at the bottom of the page, and goes on for 16 more (!). Let’s just use the first page, because I’m trying to keep this brief.
This is a pretty clearly legible handwriting font, unlike certain doctors named Baldwin Hamey that I could speak of. But even with this comparatively easy text, it takes a lot of time to read it and transcribe it, let alone to translate it and determine whether there’s any useful information. And if there’s nothing relevant, then you’ve just wasted a massive amount of time for nothing at all.
Instead...
2: Send your image to the AI tool for processing.
Before slogging through it manually, let’s first visit the fairly well-known ChatGPT, located at https://chatgpt.com/. There are probably many other services that could do this, but I’m going to keep it basic. Note that this does not constitute an endorsement of ChatGPT specifically. I’m just using it here because it worked.
The box looks like this. (I’m always amused by the absurd prompt suggestions. I’ve never tried them.)
Drag and drop your first image into the “Message ChatGPT” part of the chat window.
Add the instruction “Please transcribe and translate this Latin” (or something similar).
Send it.
ChatGPT will output a response accordingly, as follows:
Copy the whole thing (you can use the convenient copy button for this).
Paste it into another document. Which brings us to step 3…
3: Read the response to confirm whether the page is relevant to your research.
Copied into text:
**Transcription:** ``` Hic Caroli cum principis Fu Regis. Ille Electorum primus post Praesidem; Hic primus ab illis, in nostra Pharmacopoeiae prima editione. Cetera me latent, praeter ultima Juni oris, qui rure senectam diem obijt mense Januarii, 1654. Dor de Mayem Theodorus de Mayene Vir centius equestris inter ``` **Translation:** ``` Here lies Charles, with the Prince of the King. He was the first of the Electors after the President; He was the first among them, in our first edition of the Pharmacopoeia. The rest is hidden from me, except for the last day of June, he died in the countryside in his old age in the month of January, 1654. Lord of Mayem Theodore of Mayene A centurion among knights ``` The text appears to be an epitaph or a commemorative note detailing the life and contributions of Charles, mentioning his role among the electors and his connection to the first edition of the Pharmacopoeia. It concludes with his death in January 1654, with Theodore of Mayene (or the writer) adding a personal note.
Since we’re not looking to translate that upper portion of the text (“Here lies Charles” etc), I’ll skip that. But it's a solid start.
Inconveniently, Theodore’s obituary cuts off at the beginning of the first sentence. With the goal in mind of an accurate translation, it’ll be best for me to to send the second page, copy the extracted Latin, compile pages 1 and 2 of the Latin together to complete the sentence, and so forth.
Here’s the result of processing the second page:
Copied:
**Transcription:** ``` Anglos, ac inter Allobroges, septentrionalem ad oram la- cús Lemani, Albonae Baro, in Aula Britannica, plus decies quatuor annis. Archi- atorum Comes excessit 22 Martij 1654. Erant Meher- cúli plurima, quae Mayer- nium, ad hoc fatifigij exe- erunt. Eruditio. Sagacitas, Religio, Comitas, et mate- ``` **Translation:** ``` The English, and among the Allobroges, to the northern shore of Lake Geneva, Baron of Albona, in the British Court, for more than fourteen years. The Count of the chief physicians died on the 22nd of March 1654. By Hercules, there were many things that exhausted Mayernium with this labor. Learning. Sagacity. Religion, Courtesy, and ma- ``` This text seems to be another commemorative note or a part of a biographical entry, detailing the achievements and roles of a noble figure associated with the British Court, specifically mentioning his time at the court, his titles, and his death. The text ends mid-sentence, suggesting it is part of a longer entry.
Here you can start seeing the shortcomings of the automatic processing, especially on the line that’s the third from the bottom. What on earth is “fatifigij?” Maybe a trendy brand name, but surely not Latin.
And now for step 4:
4: Manually correct the errors in the transcript.
If the material is irrelevant to your research, don't worry about it— after all, now you know what it’s actually about, and whether or not it matters. However: If this is an important document, the Latin must be hand-reviewed and corrected. You cannot skip this.
It helps a lot if you know Latin for this part. If you don’t know the language at all, that’s a bit of a disadvantage, and you should probably get some help to check it over, if you can find anybody. Fortunately, once you (or some other obliging soul) has fixed it, GPT can re-process the piece of text with correct Latin, as many times as you want or need.
Here’s what we get for putting pages 1 and 2 of Mayerne together, without correction:
Dor de Mayem Theodorus de Mayene Vir centius equestris inter Anglos, ac inter Allobroges, septentrionalem ad oram la- cús Lemani, Albonae Baro, in Aula Britannica, plus decies quatuor annis. Archi- atorum Comes excessit 22 Martij 1654. Erant Meher- cúli plurima, quae Mayer- nium, ad hoc fatifigij exe- erunt. Eruditio. Sagacitas, Religio, Comitas, et mate-
Now we’ll return to the original Latin page and see what the text actually says.
After some work, here is the fixed version, with all corrections underlined and in bold. The end of the next sentence has been omitted (I can pick that up after processing page 3, and so forth).
Dor de Mayern Theodorus de Mayerne Vir census equestris inter Anglos, ac inter Allobroges, septentrionalem ad oram la- cús Lemani, Albonae Baro, inque Aula Britannica, plus decies quatuor annis. Archi- atrorum Comes excessit 22 Martij 1654/5. Erant Meher- cúle plurima, quae Mayer- nium, ad hoc fastigii euexe- runt.
As you can see, the accuracy rate is pretty good. Most of the errors are skipped or mistaken letters, and many of them are things that would trip up a human, too.
The error in “inque Aula Britannica” is because of the peculiarly abbreviated method that many of these writers consistently used for “que,” in which it becomes a small symbol at the end of the Q. GPT didn’t know what to do with this and just wrote the word as “in.”
The error in “Martij” has to do with the Latin J vs I; the letter J as we know it did not exist in the classic Latin alphabet, and was a medieval addition. By the 1600s, it was a firmly ingrained habit to write “ii” Latin words as “ij” instead. In order to get a proper result from the translator, I prefer to correct it to the original “ii.”
And the only fully butchered word, “fatifigij” which appears to be “fastigii,” is solvable by correctly identifying the letters as they appear in other words:
-The “ij” vs “ii” problem is addressed above.
-The “g,” which could be confusable for a “q,” appears higher up on the page in “Allobroges,” which is a very specific word referring to a group of Gallic people.
-And the “s,” which is so often written in a way that modern readers understandably confuse with “f,” appears on the previous page within “equestris,” another word that is unlikely to be anything else.
You might say that the meticulous typo correction seems like an awful lot of work. You’re right. But if you got lucky, and you found something important buried in endless pages upon pages of Latin, it has to be transcribed correctly. And this is where GPT helps, with the raw output of the processed text. As every writer knows, it’s easier to edit a bad draft than to start from scratch.
With the new version in hand, let’s return to GPT for the final step 5:
5: Instruct the tool to re-translate the fixed version of the Latin text.
(Disclaimer: I had a typo here. It says "euexe-erunt" instead of "euexe-runt." GPT caught and corrected for my own human error. Thanks!)
Copied:
``` Dor de Mayern Theodore de Mayerne A man of equestrian census among the English, and among the Allobroges, to the northern shore of Lake Geneva, Baron of Albona, and in the British Court, for more than fourteen years. The Count of the chief physicians died on the 22nd of March 1654/5. By Hercules, there were many things that raised Mayerne to this height. ```
This still doesn’t look perfect — I’m questioning “The Count of the chief physicians” — but it’s decent. So let’s ask what that’s all about:
Not quite satisfactory yet. Let’s give it another shot:
Dor de Mayern Theodore de Mayerne A man of equestrian rank among the English, and among the Allobroges, on the northern shore of Lake Geneva, Baron of Aubonne, and in the British Court, for more than fourteen years. The chief of physicians died on the 22nd of March 1654/5. By Hercules, there were many things that raised Mayerne to this height.
“Equestrian rank” is likely a colloquial use of “equestris” referring to Mayerne’s status, maybe his knighthood. But, right now, this is accurate enough for me. In other words: I can say, with certainty, that Pages 1 and 2 describe the life, times, and travels of the illustrious Theodore de Mayerne.
Which isn’t actually the Mayerne information I needed. But this passage worked very well for the how-to guide, so I really can’t complain.
At this point, you might be inclined to say that this tool could render human Latin specialists obsolete. But, all things considered, that’s very unlikely. Nothing can replace the human eye for translation, especially for dead languages. A scholar might know the exact meaning and significance of “vir census equestris” as Hamey wrote it. GPT clearly does not. (I don’t, either, so I’m not judging.)
But, with that said, AI-powered OCR serves an unfilled function — rapid and efficient bulk processing.
There is a very large amount of useful historical material buried in Latin texts, and a very small number of people who are capable of manually extracting it. If I hadn’t discovered I could use GPT for this, I wouldn’t have been able to use an alternative method — it simply couldn’t have been done at all. Hiring an expert for Latin translating on the necessary scale (hundreds of pages) would be prohibitively expensive, especially with my inconvenient lack of budget. And, especially for Hamey’s notes and other documents, there’s no way to know which pages would be relevant in the first place. Even with my Latin education (albeit, from a decade ago), the substance and nuance was impenetrable to me. I’m certain I’m not the only one who has run into this quandary.
I believe that this method will give the academic world a leg up when it comes to processing extensive raw data from old texts, especially in foreign languages. GPT can’t do everything — it struggles with unclear material (as we all do) — but it saves a lot of toil and drudgery, allowing researchers to devote their time to the tricky business of really doing the research. Once the raw information from the Latin is available to read and process in English, it’s far easier to use that knowledge to accomplish what scholars originally set out to do: making connections, drawing conclusions, and sharing their discoveries with everybody else.
Thanks for reading. Hope this helps.
