In a world filled with concerns about AI’s impact on job markets and apocalyptic scenarios, it’s essential to recognize the positive potential of artificial intelligence. One such remarkable application comes from a master’s thesis at Tel Aviv University, where Gai Gutherz and Professor Shai Gordin collaborated to accelerate the translation of ancient texts into modern languages. They focused on Ancient Akkadian cuneiform, one of the world’s oldest written languages, found on tens of thousands of clay tablets.
Piggybacking on Google’s Neural Machine Translation
Gutherz and Gordin’s project harnessed neural machine translation, the technology underlying Google Translate, for their ambitious task. The neural networks used in this translation approach were revamped in late 2016, leading to Google’s Neural Machine Translation (GNMT). GNMT boasts impressive speed, improved accuracy (60% fewer errors), and a better ability to handle rare words.
Training AI for translation
Incorporating GNMT into his work at Tel Aviv University, Gutherz embarked on teaching AI how to decipher Akkadian cuneiform. Like other language-oriented AI systems, the researchers trained GNMT using vast datasets. As the AI translated more text, its accuracy improved, akin to mastering a musical instrument through practice. For instance, GPT-3, a large language model, was trained on 300 billion human words.
Translating the voice of the past
The Akkadian cuneiform language poses a unique challenge. While there are only a few hundred experts worldwide who can decipher it, hundreds of thousands of untranslated clay tablets exist, making it a perfect candidate for Gutherz and Gordin’s project. However, the already-translated Akkadian tablets available for training were significantly smaller than vast language model datasets like ChatGPT’s.
The complexities of Akkadian-to-English translation
The choice of translating Akkadian to English presented additional complexities. Akkadian sentences follow either a subject-object-verb (SOV) or an object-subject-verb (OSV) order, while English follows a subject-verb-object (SVO) structure. Transliterating the Akkadian words word-for-word resulted in a high 97% accuracy. However, when adjusting the word order to English, the AI faced challenges with complex and nuanced sentences, requiring human intervention to finalize the translation.
The two-step Akkadian-to-English translation process
Gutherz and Gordin’s 2023 study, published in PNAS Nexus and available in Oxford University Press, outlines their two-step translation process. First, the researchers transliterate the Akkadian text. Next, they recombine the transliteration into a properly word-ordered English translation, refining the AI’s output to ensure accuracy and clarity.
Usefulness and criticisms
The potential of Gutherz and Gordin’s AI-driven project is immense. It could simplify the translation of obscure ancient texts into modern languages, proving especially valuable for Akkadian, a vital bridge language in the ancient Middle East for over three millennia. Akkadian was the dominant language, borrowing from Sumerian cuneiform and evolving through contact with various cultures.
However, the project has faced criticism regarding AI’s performance in poetic or abstract works. Philologists and translators like Nathan Wasserman, professor of Assyriology at the Institute of Archaeology at the Hebrew University of Jerusalem, believe AI is “still very far from being useful” for ancient text translation. Nonetheless, the project represents a productive step in the right direction, offering potential applications for tracking statistics, word incidences, and text comparisons.
AI is unlocking the secrets of ancient Akkadian cuneiform, allowing researchers to gain valuable insights into the past. Gutherz and Gordin’s project showcases AI’s tempered, reasoned, and high-minded applications, proving that it can be a valuable tool in preserving and understanding our rich historical heritage. As AI advances, there is potential for even greater discoveries in ancient texts, bridging the gap between the past and the present.
