Google DeepMind’s Aeneas model can restore fragmented Latin text

At its best, AI is a tool, not an end result. It allows people to do their jobs better, rather than sending them or their colleagues to the breadline. In an example of “the good kind,” Google DeepMind has created an AI model that restores and contextualizes ancient inscriptions. Aeneas (no, it’s not pronounced like that) is named after the hero in Roman mythology. Best of all, the tool is open-source and free to use.

Ancient Romans left behind a plethora of inscriptions. But these texts are often fragmented, weathered or defaced. Rebuilding the missing pieces is a grueling task that requires contextual cues. An algorithm that can pore over a dataset of those cues can come in handy.

Aeneas speeds up one of historians’ most difficult tasks: identifying “parallels.” In this setting, that means finding similar texts arranged by wording, syntax or region. DeepMind says the model reasons across thousands of Latin inscriptions. It can fetch parallels in seconds before passing the baton back to historians.

DeepMind says it turns each text into a historical fingerprint of sorts. “Aeneas identifies deep connections that can help historians situate inscriptions within their broader historical context,” the Google subsidiary wrote.

One of Aeneas’ most impressive tricks is restoring textual gaps of unknown length. (Think of it as filling out a crossword puzzle where you don’t know how many letters are in each clue.) The tool is also multimodal, meaning it can analyze both textual and visual input. DeepMind says it’s the first model that can use that multi-pronged method to figure out where a text came from.

DeepMind says Aeneas is designed to be a collaborative ally within historians’ existing workflows. It’s best used to offer “interpretable suggestions” that serve as a starting point for researchers. “Aeneas’ parallels completely changed my perception of the inscription,” an unnamed historian who tested the model wrote. “It noticed details that made all the difference for restoring and chronologically attributing the text.”

Alongside the release of Aeneas for Latin text, DeepMind also upgraded Ithaca. (That’s its model for Ancient Greek text.) Ithaca is now powered by Aeneas, receiving its contextual and restorative superpowers.

Researchers can take Aeneas for a spin at DeepMind’s “Predicting the Past” website. It also open-sourced the model’s code and dataset.