AI for Akkadian

« previous post | next post »

Article by Melanie Lidman in The Times of Israel (6/17/23):

Groundbreaking AI project translates 5,000-year-old cuneiform at push of a button

‘Google Translate’-like program for Akkadian cuneiform will enable tens of thousands of digitized but unread tablets to be translated to English. Accuracy is debatable.

Opening and key paragraphs:

Cuneiform is the oldest known form of writing, but it is so difficult to read that only a few hundred experts around the world can decode the clay tablets filled with wedge-shaped symbols. Now, a team of archaeologists and computer scientists from Israel has created an AI-powered translation program for ancient Akkadian cuneiform, allowing tens of thousands of already digitized tablets to be translated into English instantaneously.

Globally, libraries, museums, and universities have more than half a million clay tablets inscribed with cuneiform. But the sheer number of texts, and the tiny number of Akkadian readers — a language no one has spoken or written for 2,000 years — means just a small fraction of these tablets have been translated.

“What’s so amazing about it is that I don’t need to understand Akkadian at all to translate [a tablet] and get what’s behind the cuneiform,” said Gai Gutherz, a computer scientist who was part of the team that developed the program. “I can just use the algorithm to understand and discover what the past has to say.”

The project began as a thesis project for Gutherz’s masters degree at Tel Aviv University. In May, the team published a research paper in the peer-reviewed PNAS Nexus, from the Oxford University Press, describing its neural machine translation from Akkadian to English.

Neural machine translation, also used by Google Translate, Baidu translate, and other translation engines, works by converting words into a string of numbers, and uses a complex mathematical formula, called a neural network, to output a sentence in another language in a more accurate and natural sentence construction than translating word-for-word.

Translation is an art form, so it can be difficult to measure numerically what constitutes a “good” translation, Gutherz said. In order to rate the translations, the researchers used the Best Bilingual Evaluation Understudy 4 (BLEU4), an evaluation tool developed in the early 2000s to automatically measure the accuracy of machine-created translations.

According to the study, the neural machine translation provided a BLEU4 score of 36.52 for cuneiform to English, and a score of 37.47 for transliterated cuneiform to English. BLEU4 scores are from 0 to 100, with 0 being the lowest and 100 being a perfect translation, which even a human translator could not achieve. Around 37 is considered fairly good for an early-stage translation model, explained Gutherz.

Gutherz said that Google Translate, a privately-funded commercial tool that has been in existence for over a decade, would get a BLEU4 score of about 60 translating from Spanish to English.

In 2020, Gutherz, archaeologist Prof. Shai Gordin of Ariel University, and others published a paper about using AI to translate Akkadian cuneiform to a transliterated Latin script. The transliterated script reads as a nonsensical collection of letters and numbers to the untrained eye, but is a common “language” that allows archaeologists and researchers to study and discuss cuneiform across the world.

In the 2020 paper, the team was able to use AI to achieve 97 percent accuracy from Akkadian cuneiform to transliterated Latin script. This is a much simpler process since it works by translating the cuneiform symbols to a single word, and keeping the words in the same order that they were found.

Translating Akkadian to English or transliterated script to English is a much more complicated process because it requires the computer to string together full phrases or sentences that make sense in English, which is written in a different syntactical order.

Predictably, the AI had a higher level of accuracy for formulaic texts, such as royal decrees or divinations, which follow a certain pattern. More literary and poetic texts, such as letters from priests or treaties, had a higher incidence of “hallucinations,” an AI term meaning that the machine produced a result that is completely unrelated to the text provided.

One of the things that most surprised the researchers is that the translations captured the style or rhythm of a certain genre so that they could determine — simply based on the style of the translation — if the text was a formulaic legal document, astrological report, or scholarly letter.

The biggest challenge for training the AI model was the limited amount of material — images of tablets and translated tablets — that the team had available to train the AI model. Even the largest online databases of Akkadian tablets have only tens of thousands of entries.

“The amount of data you train on is correlative to how well you can perform, and the more data you have, the better your models will be,” said Gutherz. “ChatGPT works so well because they managed to train it on basically the entire internet. For us, the main task at the beginning was to gather all the possible translations we could get, to generate as many examples as possible.”

The team drew their samples from ORACC, the Open Richly Annotated Cuneiform Corpus, an online database from the University of Pennsylvania. For the data they were able to scan, the researchers used 90% of the material for training (50,544 sentences), 5% for validation (2,808 sentences), and 5% for testing (2,808 sentences).

The article continues on for many paragraphs about the prospects and pitfalls of AI tools such as ChatGPT.  One thing is certain:  they are here to stay and will continue to improve and take on more useful tasks that are repetitive, mechanical, and boring.  Another thing, at least to my mind, is that they will not replace creative, controlling human intelligence.


Selected readings

[Thanks to Hiroshi Kumamoto]

Comments are closed.