Google's DeepMind AI fills in the blanks on broken ancient texts
From confining plasma for nuclear fusion to solving 50-year protein folding problems, DeepMind's AI is starting to prove itself on the frontiers of modern science. To aid research into the history of human writing, the company has now turned its technology to the task of restoring incomplete ancient texts, where it has performed with an impressive degree of accuracy.
In the last five or so years we've seen DeepMind's AI technology take some impressive leaps that have demonstrated its practical potential and ability to tackle some of the most complex problems in science, like those mentioned above. Other examples include improving the accuracy of rain forecasting, beating the world's best Go players and cutting costs at parent company Google's data centers.
In turning its hand to ancient history, DeepMind now aims to bring new clarity to ancient inscriptions that have become damaged, moved from their original location or are unable to be accurately dated. To do so, DeepMind researchers teamed up with historians and scientists in Italy, England and Greece, and trained the AI on the largest digital dataset of Greek inscriptions, using both individual characters and complete words as inputs.
The resulting AI tool, which the team has called Ithaca after the Greek island described in Homer's Odyssey, was able to restore damaged texts with an accuracy of 62 percent, identify their original location with 71 percent accuracy and correctly date texts to within 30 years of their true date of creation.
Because Ithaca is designed as a research tool, the finished version generates several text prediction hypotheses, which historians can then choose from. When identifying the original location of a text, instead of a single solution it offers a map with a probability distribution for 84 different ancient regions, and does the same thing for possible years when dating a text.
In experiments, expert historians achieved a 25-percent accuracy when restoring ancient texts, but when using Ithaca, this accuracy jumped up to 72 percent. The hope is that the technology can unlock the potential of AI systems working in cooperation with human experts, bringing new clarity to our understanding of past civilizations.
DeepMind says it is already working on versions trained on other ancient languages, such as Hebrew, Demotic and Mayan. It has made the code for Ithaca open source, and launched a free interactive version of it online.
The research was published in the journal Nature.