AI Helps Archaeology: MIT Google Decoding Lost Ancient Texts with Neural Networks

category:Internet
 AI Helps Archaeology: MIT Google Decoding Lost Ancient Texts with Neural Networks


Peng Mei News Internship Reporter Zhang Wei

Artificial intelligence technology is being used to decipher ancient characters that have been lost for a long time.

Recently, researchers at MIT and Google Artificial Intelligence Laboratory have proposed a neural network algorithm for automatically deciphering lost text. For the first time, this algorithm realized the automatic translation of Linear B in Mycenaean civilization of ancient Greece, and accurately translated 67.3% of the Cognates of Linear B into Greek. Next, AI and machine learning technologies may be used to decipher ancient Chinese characters that have not yet been deciphered.

Linear B appeared about 1400 B.C. and was deciphered by linguist Michael Ventris in 1953. The MIT and Google Artificial Intelligence Laboratory experiment is the first attempt to automatically translate Linear B.

According to the Massachusetts Science and Technology Review, in 1886, British archaeologist Arthur Evans stumbled upon a large number of inscribed stones and stone tablets on the Mediterranean island of Crete. In a follow-up study, he and other researchers determined that the stones and stone tablets were written in two different languages. The older one, later called Linear Character A, dates back to 1800 BC to 1400 BC, and the other, called Linear Character B, appeared about 1400 BC.

Early decipherment of both languages failed. Until 1953, an amateur linguist, Michael Ventris, succeeded in deciphering linear B. His success stems from two important inferences. First of all, Ventris speculated that many repetitive words in the linear B vocabulary were the place names of Crete Island, which was later proved to be correct. Secondly, he assumed that Linear B recorded the early forms of ancient Greece and then helped him decipher the rest of the text.

This time, Dr. Luo Jiaming of MITs Laboratory of Artificial Intelligence, Professor Regina Barzilay of MIT and Cao Yuan of Google Brain, based on the language change patterns recorded in historical linguistics, captured the character-level correspondence between homologous words by using the sequence-pair expression model, and established an automatic decipherment loss. Neural Network Algorithms for Ancient Texts. In this paper, researchers input lost text and non-parallel corpus in known related languages into the model, and then evaluate the alignment accuracy between lost text and corresponding words in known languages. This method not only accurately translates 67.3% of the linear B cognates into Greek, but also is used to translate Ugarite in the 15th century B.C., and the translation result is 5.5% higher than the traditional method. Researchers say their approach also shows an improvement in Roman language translation. Source: Wang Fengzhi _NT2541

This time, Dr. Luo Jiaming of MITs Laboratory of Artificial Intelligence, Professor Regina Barzilay of MIT and Cao Yuan of Google Brain, based on the language change patterns recorded in historical linguistics, captured the character-level correspondence between homologous words by using the sequence-pair expression model, and established an automatic decipherment loss. Neural Network Algorithms for Ancient Texts.

In this paper, researchers input lost text and non-parallel corpus in known related languages into the model, and then evaluate the alignment accuracy between lost text and corresponding words in known languages.

This method not only accurately translates 67.3% of the linear B cognates into Greek, but also is used to translate Ugarite in the 15th century B.C., and the translation result is 5.5% higher than the traditional method.

Researchers say their approach also shows an improvement in Roman language translation.