The clay cuneiform tablets of the Sumerians, Hittites and Babylonians are among the oldest written documents in the world. Researchers have now developed a new AI method to decipher the texts, which are up to 5,000 years old. They first trained an artificial intelligence to separate the individual characters on a board from each other. The second step was to interpret these cuneiform characters. The process could help to decipher cuneiform tablets more quickly in the future and also compare them with each other.
The first versions of cuneiform were developed by the Sumerians in Mesopotamia around 3300 BC. Many of these symbols consisted of pictograms that initially resembled pictorial writing, but over time became increasingly abstracted and became stylized phonetic and syllabic symbols. Over the millennia, cuneiform spread across much of the Near and Middle East and formed the basis for the writings of the Akkadians, Assyrians, Hittites and Babylonians.
It’s the details that matter
Archaeologists have already discovered tens of thousands of cuneiform tablets during excavations in the area where these cultures were spread. Decipherments of the writing show that the contents of these tablets cover a wide range of topics: “Everything can be found on them: from shopping lists to court rulings. The tablets provide a look into man’s past several millennia ago,” explains Hubert Mara from the Martin Luther University Halle-Wittenberg (MLU). “However, many cuneiform tablets are heavily weathered and difficult to decipher even for trained eyes.”
Particularly difficult, but crucial to the meaning of the characters, is the question of whether an imprint represents an incised line and a triangular wedge. In order to interpret the signs correctly, not only optimal lighting conditions but also a lot of background knowledge are necessary. To make matters worse, the writing system at the time was very complex and was used for several languages. “So far, accessing the contents of the cuneiform tablets has been difficult – you have to know exactly what you are looking for and where,” says Mara. Although there are initial attempts with adaptive programs, these only produced moderate results, especially with poorly preserved cuneiform tablets.
Decipherment in two stages
Mara and his team therefore investigated the extent to which artificial intelligence can help decipher the cuneiform texts. To do this, the researchers developed a two-stage system that was trained on the basis of around 200 3D scans of cuneiform tablets and additional information about the writing systems. In the first step, the AI system learned to recognize and separate the individual characters and their related components on a written tablet. The next step was a more detailed analysis of the individual characters and the distinction between lines and wedge-shaped notches.
Together, this process makes it possible to decipher and digitize the cuneiform texts in a manner similar to the OCR (Optical Character Recognition) method already used for photographs and digital scans of modern documents. In initial tests, the prototype of the new AI system was already able to decipher cuneiform tablets from different time periods, as Mara and his team report. Their digital decipherment aid was particularly accurate with smaller cuneiform tablets, but showed minor weaknesses in the edge areas of larger text blocks.
More precise than previous methods
Overall, the new two-stage AI system performed better than previous methods, which often used adaptive algorithms. Because these AI models were usually only trained on the basis of photographs of the boards, there was still a problem with the precision of the decipherment, as the team explains. In contrast, their newly developed AI system learned from 3D scans, but was then also able to decipher photos of cuneiform tablets. “Surprisingly, our system even works very well with photographs that actually represent poorer source material,” reports Ernst Stötzner from MLU.
Once converted into digital computer text, the cuneiform characters become easier to read. At the same time, digital decipherment makes it possible to search through the contents of many panels and compare them with each other. This opens up completely new perspectives for research, as the researchers explain. So far, their prototype has only mastered two of the twelve known cuneiform languages. But the AI system is expandable and can also be trained to use other languages and fonts.
Source: Martin Luther University Halle-Wittenberg; Specialist article: GCH 2023 – Eurographics Workshop on Graphics and Cultural Heritage, doi: 10.2312/gch.20231157