Scholars who specialize in the study of ancient Mesopotamia, known as Assyriology, face the challenging task of decoding Akkadian texts written in cuneiform.
Cuneiform, named after its wedge-shaped marks made on clay tablets using a reed stylus, is one of the oldest forms of writing.
A breakthrough in this field has come from the collaboration between Israel’s Tel Aviv University and Ariel University. Their team of scientists has developed an artificial intelligence model that can translate Akkadian text in cuneiform into English, eliminating the need for years of manual translation efforts by experts.
Akkadian is an extinct East Semitic language that was spoken in ancient Mesopotamia (Akkad, Assyria, Isin, Larsa, and Babylonia) from the third millennium BC until the 8th century BC, when it was gradually replaced by Akkadian-influenced Old Aramaic among Mesopotamians. By the 10th century BC, two variant forms of the language, known as Assyrian and Babylonian, were in use in Assyria and Babylonia, respectively.
Translating Akkadian to English with Ai
Archaeologists have discovered hundreds of thousands of clay tablets from ancient Mesopotamia, which date back as far as 3,400 BCE.
These tablets are written in a script called cuneiform, but only a few experts can read and translate them due to their complexity.
A group of scientists, including Dr. Shai Gordin from Ariel University and Dr. Gai Gutherz, Dr. Jonathan Berant, and Dr. Omer Levy from TAU, have just released their study on a new machine-learning model for translating ancient Akkadian texts into English. However, the findings were published in the journal PNAS Nexus.
The team developed two different versions of the model, one that translated Akkadian from Latin script and another that translated from Unicode representations of the cuneiform signs.
However, the Latin transliteration version performed better, achieving a score of 37.47 in the Best Bilingual Evaluation Understudy 4 (BLEU4), which measures the level of similarity between machine and human translations of the same text.
The machine translation program works best when translating sentences that are 118 characters or less. However, the model sometimes produces “hallucinations,” which are syntactically correct English phrases but not necessarily accurate translations of the original text.
Moreover, Dr. Gordin noted that, in most cases, the translations are still usable as a rough draft of the text. The scientists propose that the model can be used in collaboration with human scholars, who can refine and correct the output to produce more accurate translations.