A deep learning artificial intelligence (AI) model can predict missing words, fragments, and sentences from cuneiform tablets that are up to 4500 years old.
Clay tablets engraved with cuneiform text in the Akkadian language are key tools for understanding the cultures that existed in Mesopotamia (more or less the area of present-day Iraq) between 2500 BC and 100 AD Many of these tablets date to the age, are damaged and lack key sections of the text.
The computer scientist Gabriel Stanovsky Hebrew University of Jerusalem and colleagues from different departments have collaborated to use artificial intelligence and unravel the secrets of these plates, completing the missing cuneiform text.
Encode tables in cuneiform writing
In the past, research has already "read" old documents (letters of the renaissancethe rolls of herculaneum), but never with this type of approach.
The team used a deep learning AI model already trained in 104 different languages. These include some Semitic languages such as Hebrew, which shares similarities with Akkadian. They then trained the algorithm by transcribing 10.000 tablets into cuneiform writing. The AI model was able to suggest contextually accurate words and phrases to fill in the gaps. Take it as a kind of T9, but with the Mesopotamian.
How do we know that the suggestions are relevant? The researchers also tested AI on already known parts of the tablets, and completion was excellent there as well. The artificial intelligence has reconstructed the sentences in cuneiform writing with an amazing 89% accuracy, in some cases even expanding the possible interpretations of the texts.
The importance of knowing languages
"The main finding of this study," says Stanovsky, "is that the use of other languages really helped encode Akkadian." Indeed, without pre-training the model on those 104 different languages, accuracy was nearly 30 percentage points lower.
It is a tool that in the next few years, I am sure, will unleash enormous potential for the decoding of important historical documents.
References: arxiv.org/abs/2109.04513
