TL;DR:
- AI translation tool deciphers ancient Akkadian cuneiform texts in seconds.
- Interdisciplinary researchers combine computer science and history to create the AI model.
- Translation challenges arise due to cultural nuances and unknown language contexts.
- Vast amounts of knowledge remain untapped in untranslated cuneiform tablets.
- AI achieves high-quality translations with the potential for further improvement.
- AI-assisted translations save time and labor for human researchers.
- Future prospects include AI providing sources used for translations and aiding scholarly purposes.
Main AI News:
The challenging task of deciphering dead languages has long perplexed scholars. From Egyptian hieroglyphics to Mayan glyphs, unraveling the mysteries of these ancient scripts has taken years, even centuries. Linear B, an early form of Greek, remained a puzzle for over three millennia. This is where the potential of artificial intelligence (AI) shines, tackling complex problems like these with unprecedented speed. Even for languages already deciphered, challenges persist. Consider Akkadian cuneiform, one of the oldest written languages, with nearly a million untranslated texts. Now, thanks to AI, decoding these ancient writings takes only seconds.
In a recent journal article, an interdisciplinary group of computer science and history researchers unveiled their groundbreaking AI model for instant translation of ancient glyphs. Spearheaded by a Google software engineer and an Assyriologist from Ariel University, the team leveraged the technology behind Google Translate to train their AI model on existing cuneiform translations.
A Beacon of Hope for Translation Challenges
When it comes to deciphering dead languages, especially those without descendant languages, understanding meaning without extensive cultural context is akin to navigating without a North Star. Akkadian is precisely such a language. As the language of the Akkadian Empire, situated in present-day Iraq from the 24th to 22nd centuries BCE, Akkadian existed both as a spoken and written language. Its cuneiform script utilized a collection of sharp, intersecting triangular symbols. Akkadians would often write by marking clay tablets using the wedge-shaped end of a reed (cuneiform derives from the Latin term meaning “wedge-shaped”). Today, countless clay tablets have endured the passage of time, adorning the halls of various universities and museums, containing invaluable knowledge.
Translation is far more than a mere decryption of foreign words or phrases. Frequently, a statement in one language lacks an exact or straightforward equivalent in another, accounting for cultural nuances and structural disparities. Achieving high-quality translation demands a profound understanding of both languages’ structures, the surrounding cultures, and the historical underpinnings that shape them. Translating a text while preserving its original tone, rhythm, and even humor is a delicate art—one that becomes exceedingly challenging when the cultural context of the language remains largely enigmatic.
The staggering number of cuneiform texts far surpasses the limited pool of linguists capable of deciphering Akkadian. Consequently, vast amounts of knowledge about this early civilization, often regarded as the world’s first empire, remain untapped. Currently, the number of existing tablets and the rate at which archaeologists unearth new ones far outpaces the translation efforts of linguists. However, with the integration of AI into the cuneiform interpretation process, this situation may change.
“Documenting the political, social, economic, and scientific history of ancient Mesopotamia, hundreds of thousands of clay tablets inscribed in the cuneiform script remain untranslated and inaccessible due to their sheer number and the limited quantity of experts able to read them,” the team noted.
The AI system offers two types of translation—cuneiform to English translation and cuneiform transliteration (phonetic rewriting). Achieving impressive scores of 36.52 and 37.47, respectively, on the Best Bilingual Evaluation Understudy 4 (BLEU4), a metric for translation quality, the AI outperforms the team’s initial targets. These scores indicate high-quality translations, with 70 being the highest achievable score for a highly skilled human translator.
Moving Beyond Brittle Translations
Traditionally, computer-generated translations were prone to errors and considered unreliable. According to Tom McCoy, a computational linguist at Princeton University, translation programs relying on grammatical rules often miss the depth of meaning conveyed by idioms and nonliteral language, slipping through the cracks of formal grammar. However, recent advancements in AI, including the cuneiform translator, have begun to delve into the subtler aspects of language. This heralds an exciting era of AI-driven computational linguistics.
“In recent AI, the big breakthrough has been statistical processing, which entails a different type of mathematics compared to the rigid rules employed in previous methods,” explained McCoy. “Statistics helped overcome the limitations of earlier approaches. Now, we’re working with machine learning and deep learning. Machines are capable of learning all these idiosyncrasies, idioms, and exceptions to rules, elements that were absent in the previous generation of AI.”
Nevertheless, the AI-powered cuneiform translations still exhibit occasional errors and “hallucinations” common to AI models. For instance, a phrase originally translated as “Why should we (also) conduct the lawsuit before a man from Libbi-Ali?” was rendered as “They are in the Inner City in the Inner City.” Despite these occasional missteps, the tool significantly reduces the time and human labor required for the initial processing of texts.
“AI is currently remarkable but unreliable. While it can accomplish remarkable feats, one can never wholly trust the output it generates,” cautioned McCoy regarding using AI for translation. “The best case for utilizing AI is when a task is labor-intensive and challenging for humans but, once AI provides output, humans can easily verify it.“
The AI model demonstrates the highest accuracy when translating shorter sentences and formulaic texts such as administrative records. Surprisingly, it also succeeds in capturing genre-specific nuances during translation—a discovery that delighted the researchers. To further enhance accuracy, the AI will undergo additional training on more extensive translation samples, as outlined by the researchers.
For now, the AI can assist researchers by producing preliminary translations that humans can subsequently verify for accuracy and refine with nuanced understanding.
“An exciting future scenario would involve the [model] presenting the user with a list of sources used for its translations, a feature particularly valuable for scholarly purposes,” the researchers envisioned.
Conclusion:
The development of AI-powered translation for ancient languages represents a significant breakthrough in the field. This technology has the potential to unlock invaluable knowledge from untapped sources, offering new insights into ancient civilizations. While the current AI system is remarkable but not infallible, it saves considerable time and effort for human translators. The integration of AI into the translation process opens up opportunities for scholarly research and paves the way for further advancements in the field of computational linguistics.
As AI continues to improve, it will play a pivotal role in deciphering complex dead languages and reshaping the landscape of historical and linguistic studies. Businesses operating in the translation industry should closely monitor these advancements and explore potential applications for their services, as AI-driven translation tools could significantly enhance efficiency and accuracy in the future.
