Get the latest tech news
How AI is unlocking ancient texts
From deciphering burnt Roman scrolls to reading crumbling cuneiform tablets, neural networks could give researchers more data than they’ve had in centuries.
This time, the researchers took advantage of a breakthrough in machine learning called the transformer model, which captures more complex language patterns than an RNN is able to by analysing different characteristics of an input — such as characters or words — in parallel, weighting them according to the context. It isn’t possible to know when it has contributed to research unless the authors choose to acknowledge it, says Sommerschield, but examples reported so far include the re-dating of Athenian political decrees, and an investigation of tablets from the fourth century bc that contain questions put to the Oracle of Dodona in northwestern Greece. For example, Katerina Papavassileiou at the University of Patras, Greece, and her colleagues used an RNN to restore missing text from a series of 1,100 Mycenaean tablets from Knossos, Crete, containing accounts of sheep herds written in a script called Linear B in the second millennium bc 9.
Or read this on Hacker News