Познакомьтесь с Энеем: искусственным интеллектом, который может заполнить пробелы поврежденных латинских текстов

The best result comes when a person and a model work together.

The artificial intelligence (AI) model can predict where ancient Latin texts come from, estimate how old they are, and restore the missing parts. The model, called Aeneas and described today in Nature, was developed by some team members who created a previous artificial intelligence tool that could decipher ancient Greek inscriptions.

Learning ancient inscriptions, known as epigraphy, is a difficult task because some texts lack letters, words or sections, and languages change over time. Historians analyze texts by comparing them with other inscriptions containing similar words or phrases. But the search for these other inscriptions is incredibly time-consuming, says co-author Thea Sommerschild, an epigrapher from the University of Nottingham, Great Britain.

Another problem is that new inscriptions continue to open, so there is too much information for any person, says Ann Rogerson, who studies Latin texts at the University of Sydney, Australia.

To facilitate the restoration, translation and analysis of inscriptions, a team consisting of researchers from universities in the United Kingdom and Greece, as well as from Google's artificial intelligence company DeepMind in London, has developed a generative AI Model trained on inscriptions from the world's three largest Latin epigraphy databases. The combined data set contained text of 176,861 inscriptions - plus images of 5% of them - with dates from the seventh century BC to the eighth century AD. The model includes three neural networks, each of which is designed for different tasks: restoring missing text; predicting where the text comes from; and evaluating its old time. Along with the results, Aenea also provides a list of similar inscriptions from the data set in support of his answer, ranked according to the extent to which they relate to the original inscription.

"Eneas can instantly retrieve corresponding parallels from our entire data set," because each text has a unique identifier in the database, says co-author Yannis Assael, a researcher at Google DeepMind.

The team checked the accuracy and usefulness of the model by asking 23 epigraphs to restore the text that was removed from the inscriptions. Specialists were also asked to date and determine the origin of the inscriptions, both separately and using a model. The experts themselves dated the inscriptions to about 31 years from the moment of the correct answer. The dates predicted by Aeneas were correct for 13 years.

When it came to determining the geographical origin of the inscriptions and restoring parts of the text, specialists who had access to a list of similar inscriptions of the model and its forecasts were more accurate than specialists working alone or only on the model. Specialists also dated the inscriptions for about 14 years after the correct answer, when they had a list and forecasts of the model.

Assistant historian

Then the model was tested on a well-known text called Res gestae divi Augusti, which details the life of the Roman emperor Augustus. The model's predictions about the age of the inscriptions were similar to the predictions of historians, and the tool was not misled by the dates mentioned in the text. He also chose spelling options and identified other features that the historian would use to predict age and origin.

Aeneas also performed well when inspecting the altar with Latin inscriptions. He included in his list of similar inscriptions another altar from the same region, which, according to the team, was noted because the model was not told that the two altars were geographically related or belong to the same period of time.

Rogerson says that the model can be used to analyze huge amounts of data that exceed one person. It can also help historians find inscriptions similar to those they are working on, which can take weeks or even months manually, and can be useful for students who study epigraphy, she says.

The model's answers seem to be more justified than the responses of popular artificial intelligence tools, Rogerson adds. "He gives a hypothesis based on the evidence base on which he works, so this is a rational assumption, not a wild blow in the dark."

However, the team behind Aeneas said that the model was limited because its training database was smaller than that of other models such as ChatGPT and Microsoft Copilot, which could affect its performance on unusual inscriptions. Rogerson says that "Anea" may not be so useful for inscriptions that are unique or date back to the period when fewer artifacts are available.