Fragmentary Latin inscriptions can be completed with AI
A new model is finding connections spanning the Roman world
History depends on the written word. But how can a historian interpret a text if its authorship or age are uncertain, and indeed some of those words are missing? The problem is not a new one. But where human experts have struggled, historians are turning to artificial-intelligence (AI) models for suggestions, with impressive results.
Over the past five years or so, the predictive abilities of artificial neural networks have increasingly been applied to reconstructing the past. In that time they have assisted with everything from piecing together smashed Babylonian tablets to deciphering the characters inscribed on ancient Chinese turtle shells. The most prominent example has involved AI enthusiasts using high-quality scans to digitally unfurl unopenable papyrus scrolls that were carbonised during the eruption of Mount Vesuvius in 79AD.
Now scientists will be able to use AI models to suggest likely dates and geographical origins for samples of Latin inscriptions, and even predict missing bits of text. Yannis Assael, a researcher at Google DeepMind in London, and Thea Sommerschield, a historian at the University of Nottingham, described their model, Aeneas (named for the mythological ancestor of the Romans), in a paper published this week in Nature. Aeneas can process images of Latin text as well as transcribed inscriptions, and is an iteration of an earlier model that focused on Greek inscriptions.
Aeneas represents the latest step towards the researchers’ goal of using AI models to do more than read individual texts. Drs Assael and Sommerschield hope to use large models, trained on tens of thousands of written sources, to glean unseen connections about ancient lives. Aeneas was trained on over 175,000 inscriptions, dating from the 7th century BC to the 8th century AD, and spanning Roman provinces from Britain to Mesopotamia.
The model can be extremely accurate: in tests, it dated unseen texts to within 13 years of the accepted figure. Crucially, it can also suggest other sources that may be connected. With an estimated 1,500 new Latin inscriptions discovered every year, from slaves’ epitaphs to emperors’ decrees, identifying relevant parallels is one of historians’ most important—and challenging—tasks. Human experts, even those with particularly impressive memories, have intimate knowledge of only specialist areas, while automated searches across the wider corpus are generally limited to strings of characters. Aeneas, by contrast, can search for thematic links across millennia, and the entire Roman world. Aeneas “helps us do things faster, and better", says Dr Sommerschield, but also goes “beyond what we could do already".
The researchers also tried out their model on contested inscriptions such as Res Gestae Divi Augusti, an account of the life of Rome’s first emperor, Augustus, carved into a temple wall in Ankara, Turkey; and a third-century AD altar text from Mainz in modern-day Germany. Aeneas far outstripped existing computer searches, they concluded, identifying “subtle and meaningful historical connections beyond literal matches, in ways that mirror expert-level reasoning". Aeneas found other texts composed decades apart that bore similarities to the altar text, for example, stretching from the German city of Bonn to Bulgaria, following the movements of the Roman army. Drs Assael and Sommerschield say it is essentially modelling how the Roman Empire was connected, through the movement of people, beliefs and ideas.
It is Aeneas’s ability to suggest missing text for gaps of uncertain length, however, that has some historians most excited. Many surviving Latin inscriptions are badly damaged, which means such a tool could generate new insights from existing material. For now, Aeneas’s gap-filling chops are less impressive than its dating ones. When presented with deliberately obscured text, the correct segments (up to 20 characters in length) featured among Aeneas’s top 20 predictions 46.5% of the time. When the length of the segment is unknown, this drops to 32.7%. All the same, says Charlotte Tupman, who lectures in classics and digital humanities at the University of Exeter, this represents a major leap forward. Most useful may be Aeneas’s ability to explain its reasoning, providing “saliency maps" that highlight which parts of the source influenced its predictions.
Working together
To test how their AI model augments human abilities, the team asked 23 historians to analyse and restore a list of texts that had identifying data removed. They found that, overall, historians working together with AI gave more accurate results than either on their own. That provides a “compelling case" for Aeneas to be incorporated into historians’ workflows, says Dr Tupman. The volunteers reported that the context provided by Aeneas was useful 90% of the time, and improved their confidence in key tasks by 44%. One said the similar texts Aeneas retrieved “completely changed my perception" of an inscription; another that they achieved in 15 minutes what would normally take a couple of days.
Drs Assael and Sommerschield suggest specialised tools like Aeneas could soon be integrated into chatbots, enabling historians to interrogate data as part of a more natural conversation. And they hope similar techniques will be applied to other languages and other types of texts, from tablets to papyri, perhaps even connecting archives of different civilisations. “The more data we have," says Dr Assael, “the more interesting patterns we can extract."
Curious about the world? To enjoy our mind-expanding science coverage, sign up to Simply Science, our weekly subscriber-only newsletter.
topics
