Experts from Tel Aviv University, in collaboration with Ariel University, created a program to decipher an ancient language that is difficult to interpret. An accurate “automatic” translation of cuneiform characters into English. An example of collaboration between man and machine in a field in which experts are lacking. To date, hundreds of thousands of tablets remain to be deciphered.
Tel Aviv () – Exploiting the potential of artificial intelligence (AI) to decipher a language rooted in the past and often difficult to interpret, facilitating the task of historians, linguists and translators. This is what a group of Israeli researchers from Tel Aviv University (TAU) have done in collaboration with Ariel University, who have developed a model that allows the “automatic translation” of Akkadian texts in cuneiform characters into the most common and understandable modern English. .
Specializing in the archaeological, historical, cultural, and linguistic study of ancient Mesopotamia, Assyriologists have spent years trying to interpret texts in cuneiform, one of the oldest known forms of writing. The reference to the “wedge shape” recalls its use in the past, when the signs were printed on a clay tablet.
Akkadian, a language spoken in ancient Mesopotamia (which corresponds to present-day Iraq), was an East Semitic language used mainly by Assyrians and Babylonians. It is the oldest known Semitic language, and is based on a writing system first exploited by the Sumerians. It is said that the language owes its name to the city of Akkad, the city of King Sargon, founder and father of the empire, the largest inhabited center of the time, although until now there are no certain traces of it.
Over the decades, archaeologists have unearthed hundreds of thousands of cuneiform clay tablets dating back to 3,400 BC, far more than the few scholars who can understand and translate them. Shai Gordin from Ariel University, Jonathan Berant and Omer Levy from TAU, along with other colleagues, have recently shared the fruits of their studies in the specialized journal pnas in an article titled “Translating from Akkadian to English by Neural Machine Translation.”
During the study and design phase, the team developed two versions of the model, one of which translated Akkadian from Latin script representations of cuneiform signs and the other from unicode representations of cuneiform signs. The first, corresponding to Latin transliteration, is the one that has given the best results over time, with a score of 37.47 in the Best Bilingual Evaluation Understudy 4 (BLEU-4), a parameter used to evaluate the correspondence between human and artificial translation of the same text. The program, as they explained in a note, was especially effective when translating short sentences of less than 120 characters; when exceeding this limit, “hallucinations” (a syntactically correct text, but imprecise) appeared.
The program could be especially useful in a first phase of translation, in the category of “man-machine collaboration”, to then leave room for human intervention to refine the translated text in the final version. There are hundreds of thousands of cuneiform manuscripts related to the political, social, economic, and scientific life of ancient Mesopotamia. “And most of these texts,” the researchers explained to the Jerusalem Post– remain inaccessible to the limited number of experts able to understand them.
Translation is a “complex process” because it requires not only the presence of “people who know two different languages”, but also be able to understand “different cultural contexts”. In this perspective, they continued, “digital tools can help” thanks to elements such as “optical character recognition (OCR) and machine translation.” Ancient languages, the note concluded, pose even greater problems of complexity because “their reading and understanding requires the knowledge of a long-gone linguistic community” and the texts themselves “can be highly fragmentary.”
