Google Translate Decodes World’s Oldest Language with AI

An ancient language with roughly untranslated documents now has a translator that works in minutes AI

The world's oldest language, Akkadian, was translated using artificial intelligence (AI) by a mixed team of computer scientists and historians. The team, led by a Google software engineer and an Assyriologist from Ariel University, used the same technology that powers Google Translate to construct an AI model capable of instantaneously reading the ancient characters found on cuneiform tablets. 

Akkadian, the language of the Akkadian Empire that flourished in modern-day Iraq from the 24th to 22nd century BCE, presents unique translation issues. Understanding its meaning is like going without a North Star since there are no descendant languages and a lack of cultural background. The Akkadian writing system used cuneiform, distinguished by sharp, intersecting triangular symbols carved on clay tablets with the wedge-shaped end of a reed. Yet, because of their sheer volume and the restricted number of specialists who can read them, most of these texts remain untranslated and unavailable.

The vastness of existing cuneiform writings significantly outnumbers the small number of Akkadian linguists. As a result, a great bank of information about this critical early civilization, frequently referred to as the world's first empire, remains unexplored. Linguistic attempts to translate Akkadian writings need help to keep up with the growing quantity of tablets uncovered by archaeologists. On the other hand, the incorporation of AI into the cuneiform interpretation process has the potential to change this environment.

The AI model developed by the team specializes in two types of translation cuneiform to English translation and cuneiform transliteration. The model's translation quality, as judged by the Best Bilingual Evaluation Understudy 4 score, produced outstanding results. The model outperformed the team's expectations, scoring 36.52 and 37.47 for the two translation categories and providing high-quality translations. The BLEU4 score goes from 0 to 100, with 70 representing the best possible result for a highly experienced human translation. 

Computer-generated translations have always been fragile and untrustworthy, unable to grasp the full complexity of idioms and nonliteral language that defy conventional grammatical constraints. Recent advances in artificial intelligence, such as the cuneiform translator, have dug into the more delicate parts of the language. Despite its outstanding accomplishments, the cuneiform AI translator still creates mistakes and hallucinations, which are frequent in AI systems. The AI model performs well for translating shorter lines and formulaic documents like administrative records. It also reproduces genre-specific characteristics during translation, which piqued the researchers' interest. In the future, AI will be taught on greater samples of translations. The AI translator assists researchers by creating preliminary translations that human specialists may modify and check.