Aeneas, an AI of Google Deepmind called as the Trojan hero who landed in Lazio, reconstructs the missing parts in the inscriptions and helps to date them.
An artificial intelligence co-creator from Google Deepmind, Aneneascould become the best ally of historians in reconstructing the ancient Latin inscriptions. The IA, trained on texts and images of tens of thousands of writings engraved in the stone in various places and eras of the history of Rome, can complete the missing letters and words in the fragments that have been lost And to hypothesize where and when the text was written, based on the regional and temporal variations of the language. His skills were illustrated in an article on Nature.
Ancient braces
Thea Sommerschield, historical and epigraphist of the University of Nottingham, in the United Kingdom, imagined and created with the company of IA of Google Deepmind, in London, and with other researchers expert in ancient languages a system capable of lighten the heavy reconstruction work of enrollments in the latin language damaged.
Comparing the pieces of epigraphs (the stone writings) with other inscriptions is a painstaking task, because the language changes in time and places, and because every day new inscriptions come to light – too material, for the memory of a single person.
At school from the classics
The group of scientists made a generative resistantThat is, capable of creating new content on the basis of what has been learned, trained on 176,861 inscriptions taken by the three largest epigraphs in Latin epigraphs, in addition to the photos of 5% of them. The writings had been produced from the seventh century AC to the eighth century AD.
The model, renamed AneneasAeneas, like the Trojan hero who landed in Lazio after a thousand vicissitudes and founder of the lineage from which Rome was born, is based on three neural networks (i.e. algorithms of machine learning inspired by the human brain) with three different tasks: reconstruct the missing texts; understand where they were written; understand how ancient they are.
Excellent performance
The IA managed to reconstruct the missing text, and to estimate the date of realization of the registrations With precision within 13 years When scientists tried it on epigraphs whose dates were known. Aeneas even identified the Roman province in which the inscriptions had been drawn up with an accuracy of 72%.
In the tests on deliberately corrupt Latin inscriptions to simulate the damage of centuries of history, the IA system has reached an accuracy of 73% in completing missing parts Up to 10 characters long. He even managed to reconstruct fragments in which the total length was unknown, with an accuracy of 58%.
I did this …
Particularly useful for historians is the fact that the model show the list of similar registrations Take up by the databases studied and used to provide its response, put in order of relevance with respect to the registration to be completed.
This support material reveals the logic of his suggestions and allows humans who work with it to verify the validity of the results. Finding Materal similar to that to be completed can take months of work, and this rapid comparison would also be very useful to those who are still studying to become epigraphist.
In fact, it is in the collaboration with the historians that Aenes has given its best: 23 epigraphists they achieved better resultsboth in completing and in contextualizing the texts, when they were able to work with the IA.
And when the model was put to the test with the Res gestae divi Augusti“The companies of the Divine Augustus”, a text on the life of the Roman emperor Augustus whose dating has long been debated, Aneneas has reached the same chronological conclusions as most historians without letting themselves be missed by the many dates contained in the registration.
Deepmind has announced that it will make Aenes publicly available online for anyone who wants to use it.
