Artificial intelligence instantly translates 5,000-year-old cuneiform writing on LinkedIn
Translation is not simply about replacing a word with its equivalent word in another language. High-quality translation requires translators to understand how two languages connect ideas, and then use that knowledge to create a translation that maintains the nuances of the source language and makes it effortlessly understandable to native speakers.
Although the process is difficult, it is nothing compared to the challenge of translating an ancient language into a modern one. Translators must not only resurrect extinct languages from written sources, but also gain an in-depth understanding of how the cultures that produced them evolved over the centuries. Because of this, there are very few people who can translate ancient languages, and their best efforts are often surpassed by the large number of texts unearthed by archaeologists.
Akkadian is the native language of the Akkadian Empire, which arose around 2300 BC through the conquests of its founder, Sargon the Great. As a spoken language, Akkadian eventually split into Assyrian and Babylonian dialects, before being completely replaced by Aramaic in the early first millennium BC. Today, it is a truly extinct language, and there are no sublanguages to even carry on its legacy.
To help change that, a multidisciplinary team of archaeologists and computer scientists has developed an artificial intelligence that can translate Akkadian almost instantly and unlock historical records preserved on these 5,000-year-old stone slabs.
The team trained an AI model on cuneiform samples in the openly enriched annotated cuneiform corpus and taught it to translate in two different ways. First, the AI model learned to translate Akkadian from the transliteration of the original text. It also learned how to translate cuneiform directly. More specifically, it translates Unicode symbols for cuneiform script generated by another time-saving tool that automatically generates Unicode from images from raw tablets.
The AI model must then figure out how to handle the nuances of various genres in the sample—for example, the differences between literary works and administrative letters—and how to deal with the changes found in cuneiform writing over thousands of years. The AI model was then tested using BLEU4, an algorithm used to evaluate machine-translated text.
In the English transliteration test, the team's AI model scored 37.47 points. In the cuneiform test of English, it scored 36.52 points. Both scores are above their target baseline, within the range of high-quality translations. The results were surprising: the model was able to reproduce the nuances of each test sentence genre. While this is not one of the researchers' goals, they note in the study that this could open up possibilities for applications beyond translation.
"In almost all cases, the type is recognizable, regardless of whether the translation is correct or not," the team wrote. "A promising future scenario is for [models] to show users their translation-based list of sources, which is also particularly useful for academic purposes." #人工智能行业应用#
