I attended the ACL talk where he presented this. While the method was reasonably well-justified, it seems clear that the data was flawed, leading to a flawed model, leading to a false result. The author failed to ask himself the question "what if this didn't work?".
Here's what I would identify as the main flaw: To identify the language of a centuries-old document, he trained his model on a modern multilingual text -- the Universal Declaration of Human Rights.
(A questioner in the audience asked "If you needed a document in hundreds of languages, why didn't you use the Bible?", to which he had no real answer.)
So he ran this somewhat anachronistic model and it told him that the text was a good statistical match for being in... Esperanto. He threw this prediction out on the basis that it makes absolutely no sense. You'd think this would be the first sign that the model did not fit the data.
The #2 prediction was Modern Hebrew, which of course did not exist in the 15th century, so he said "that's close enough to biblical Hebrew" and claimed that as the result.
He asked a Hebrew speaker he knew to decipher a couple of sentences based on the model. The Hebrew speaker could not actually decipher them. The author pressed on and the Hebrew speaker obligingly produced some unsatisfying guesses at sentences, which sound remarkably like sentences from unsuccessful attempts to decipher Linear B texts.
And I don't understand why anyone would so confidently report a modification of the #2-ranked prediction of the model. We might as well go ahead and say "AI confirms: the Voynich Manuscript is in ancient Esperanto!"
The much more realistic interpretation is that the Voynich text matches the statistics of none of the language samples that the model was trained to recognize, and the model is overconfident in the probabilities it outputs.
I think the fact that the method was cool must have been part of the paper's acceptance at ACL. But it does not give us a real answer about the Voynich manuscript.
I studied that manuscript a bit at university, and the questioner in the audience was definitely on the right track wrt the Bible being valuable as a source of training data. We trained on the pre-15th century translations of the Bible in a number of languages. Off the top of my head, there was Old Church Slavonic, Vulgate Latin, Bibilical Hebrew plus a number of other languages people had proposed over the years, but the statistics from Voynich Herbal A and Herbal B were never very similar to any of the languages. The closest fit we found was for Manchu, trained using a Manchu translation of the 'Dao de Jing' which made us hopeful for Basanik's Manchu Hypothesis or a modified Proto-Manchu hypothesis (http://www.ic.unicamp.br/~stolfi/voynich/04-05-20-manchu-the...) but as we dug deeper, the n-gram frequencies ended up being wildly off.
Over the years, I have been more and more doubtful that the Voynich Manuscript is natural language at all. Hopefully someone will come along and prove me wrong, but I have stopped hoping that any of these headlines is the answer for quite a while.
He only claims to have potentially identified the document's language as a form of Hebrew using a language detection model trained on UN documents. That let him incidentally decipher a few lines by literally typing them into Google Translate, but he doesn't claim to be an expert on Hebrew or have the ability to translate the whole document. He's looking for Hebrew experts to test his theory.
But by the time the story makes it into it's third re-write on a chain of bad news websites, it's "Man claims his AI deciphers unbreakable code that stumped Enigma codebreakers" or whatever.
The language identification method they use (does it count as AI?) assumes that the Voynich Manuscript text is in a language they check for, the words of which may have been anagrammed. There is no consensus that this is the case. The Voynich Manuscript could be entirely meaningless, or encrypted using a different mechanism, or be written in a language they don't check for. They then find the closest match. The use of anagrams adds an extra degree of freedom, which makes finding spurious matches more likely.
So they identified Hebrew, which is at least plausible. They then translate one line of the manuscript, using Google Translate, into English.
There was no control. They should have applied their method to text known to be meaningless, and questioned their approach if meaning was found. Gordon Rugg (whom they cite) and myself (here: http://web.onetel.com/~hibou/voynich/generated-voynich-manus... ) have generated meaningless Voynichese text.
I haven't applied their entire approach, but Google Translate translates the first line of my generated manuscript, when the language is identified as Bengali, as "Do not worry about the fact that the person is very friendly".
Brb, writing an article about how Gordon Rugg and Donald Fisk's previously thought-to-be-meaningless text was deciphered by an AI. Did they channel ancient Bengali spirits when writing it? Read more to find out.
Even if he's wrong, it's super cool to watch his explanation. It's making so much sense.
I doubt any AI can do better then him here, so much vague clues you have to get to the bottom of and loads of slight matches because it's an old language that only has relatives, but no equivalents today.
Yes! I came here to post this. He builds on work by Stephen Bax widely recognized in the "Voynich community" as being the most insightful. As a trained linguist, this guy does the grunt work to build a model, which he then uses (in his "Voynich update" video) to predict an outcome successfully. That is a huge indicator that he's onto something. What a fascinating series of videos!
exactly. Stephen Bax was the most promising before Volder Z took it this step further :)
I'm so glad someone finally shares my enthusiasm, I have very high hopes to have a full translation in my lifetime. [SPOILERS:] It sounded like someone just needs to find some old knowledgeable Sinti/Roma.
Unrelated to the content of the story, but interesting nonetheless...
While reading the article (which was low quality with a flawed attention-grabbing headline, as others have pointed out), I noticed 3 consecutive articles linked on the right side were for porn-related stories. I don't know too much about the International Business Times, but the name would have made me think it was a relatively upstanding news source. Plus, the .co.uk always adds a bit of extra class for us Americans.
So, I looked the place up on Wikipedia[1]. Here's what I found...
"In late 2011, Google allegedly moved the outlet's articles down in search results in response to excessive search engine optimization activity."
"Reporting in 2014, Mother Jones claimed that IBT journalists are subject to constant demand to produce clickbait; one former employee reportedly complained that management issued 'impossible' demands, including a minimum of 10,000 hits per article, and fired those who couldn't deliver."
And it goes on...
So, yep, pretty much a click-bait factory. No wonder the story is so mangled.
Here's what I would identify as the main flaw: To identify the language of a centuries-old document, he trained his model on a modern multilingual text -- the Universal Declaration of Human Rights.
(A questioner in the audience asked "If you needed a document in hundreds of languages, why didn't you use the Bible?", to which he had no real answer.)
So he ran this somewhat anachronistic model and it told him that the text was a good statistical match for being in... Esperanto. He threw this prediction out on the basis that it makes absolutely no sense. You'd think this would be the first sign that the model did not fit the data.
The #2 prediction was Modern Hebrew, which of course did not exist in the 15th century, so he said "that's close enough to biblical Hebrew" and claimed that as the result.
He asked a Hebrew speaker he knew to decipher a couple of sentences based on the model. The Hebrew speaker could not actually decipher them. The author pressed on and the Hebrew speaker obligingly produced some unsatisfying guesses at sentences, which sound remarkably like sentences from unsuccessful attempts to decipher Linear B texts.
And I don't understand why anyone would so confidently report a modification of the #2-ranked prediction of the model. We might as well go ahead and say "AI confirms: the Voynich Manuscript is in ancient Esperanto!"
The much more realistic interpretation is that the Voynich text matches the statistics of none of the language samples that the model was trained to recognize, and the model is overconfident in the probabilities it outputs.
I think the fact that the method was cool must have been part of the paper's acceptance at ACL. But it does not give us a real answer about the Voynich manuscript.