Researchers use AI to translate texts from 5,000 years ago into forgotten language; and the machine understands better decrees than poems

 The oldest known writing system is the cuneiform, invented in Mesopotamia by the Sumerians in the 4th millennium B.C. (a millennium before the Egyptian hieroglyphs) to write on clay tablets with punch or chisel (no paper and papyrus).

Over the time it was in use (until the 1st century AD) it evolved from an ideographic system to a phonetic (syllabic) and underwent variations to adapt to the different languages that adopted it as their own, most of them unrelated to each other.

This meant that the advances achieved in the translation of certain tablets could not always be applied to those found on other sites. Add to that the low availability of experts in such ancient and minority languages and you will have hundreds of thousands of boards in your hands waiting to be slowly translated.

Most of these documents remain untranslated and inaccessible due to their large number and limited number of experts able to read them.

That’s until artificial intelligence is used to lend a hand. And now, in a scientific paper published by Oxford University Press, a team of researchers has managed to apply a deep learning model to translate academic texts with a high level of accuracy.

AI helps to translate cuneiform tablets from thousands of years (Image: Mbzt - CC BY 3.0)

Akkadian is an ancient Semitic language (such as Arabic or Hebrew), which for centuries has become the most used in Iraq and Syria, replacing the ancient Sumerian. Its cuneiform version includes characters that can have several simultaneous functions (equivalent to sounds, ideas...) which makes translation difficult.

It turned out that an AI model (already available in a GitHub online repository) shows surprising accuracy in translating formal Acadian texts, whether bureaucratic (actual decrees) or religious (omens)because all are in compliance with certain fixed standards.

However, texts of a more literary type cause AI to "hallucinate" much more frequently (i.e., "fragmented" results unrelated to input data, such as a random ChatGPT). Besides:

"The model achieves the best results in short and medium sentences of approximately 118 characters or less."

It is worth mentioning that the characters are not equivalent to our letters: sentences usually have much more information than a tweet, for example.

In addition, the model implements two types of translation, the so-called 'C2E' (direct translation from cuneiform to English) and 'T2E' (which first transliterates to the Latin alphabet of cuneiform and then translates). The results were tangentially better for this second modality (T2E).

However, the goal is not so much to leave translation in the hands of machines, but to create a methodology based on "human-machine collaboration" that helps accelerate the work of academics and Acadian students.

On the other hand, according to the researchers, measures are being taken to implement the functionalities of the new model in an existing online application called Babylonian Engine, specially designed to analyze cuneiform texts with computational methods.

Click on the button and see the AI that with only $1 works for You. Click and learn more.

Click here