Artificial intelligence translates ancient texts instantly – 07/25/2023 – Marcelo Viana

Around 2300 BC, the conquests of King Sargon the Great made Akkad the first empire in history. The Akkadians adapted to their language the cuneiform script invented by their neighbors to the south, the Sumerians, whom they admired: centuries after the Sumerian people had ceased to exist, the Akkadian monarchs still styled themselves “kings of Sumer and Akkad”.

Akkadian split into Assyrian and Babylonian dialects, which, in the first millennium BC, were replaced by Aramaic, and fell into oblivion. But these peoples bequeathed us their texts, written on countless clay tablets, from which it was possible, in the 19th century, to decipher their dead languages ​​for millennia.

But reading these texts is difficult and time-consuming. Many clay tablets are damaged or broken. And cuneiform writing is complex: the same symbol can have different meanings depending on the context. If translation between current languages ​​already requires knowledge of the respective cultures, imagine the difficulty with languages ​​of civilizations extinct long ago. There are few qualified specialists and, therefore, most texts from Mesopotamia have never been read.

In work published in March of this year in the journal PNAS Nexus and on the digital platform GitHub, researchers from Israel and Germany presented an artificial intelligence algorithm that uses convolutional neural networks, technology similar to that of the popular Google Translate, to switch from Acadian to English almost instantly.

First, the text needs to be digitized, converting the symbols in the clay into cuneiform-specific Unicode standard codes. This is done automatically from tablet photos. The ORACC scan bank was used to train the algorithm to translate using two different systems. In the first one (T2E), the translation is made from a previous transliteration of the Unicode codes into the Latin alphabet. In the second (C2E), Akkadian in Unicode is translated directly into English.

Subjected to a standardized test, the new algorithm performed better than expected, with an advantage for T2E. Not surprisingly, he deals better with formal texts, such as royal decrees, than literary texts, such as hymns or prophecies. The great surprise is that, even when the result requires revision, the algorithm is able to identify the type of text and maintain the respective style in the translation.

And the great advantage of “machine learning” technology is that the algorithm will improve by itself as it is used more and more!

PRESENT LINK: Did you like this text? Subscriber can release five free hits of any link per day. Just click the blue F below.