Archaeologists apply computer algorithms to analyze ancient trade data inscribed on clay tablets, using technology in search of lost cities.
Literary sources like Homer’s Iliad, for example, guided Heinrich Schliemann to discover the ancient Greek cities of Troy and Mycenae. Folklore and interviews with local residents helped Hiram Bingham III rediscover the ancient Inca city of Machu Picchu in Peru.
Today, archaeologists increasingly reach for digital tools to uncover the past.
Using supercomputers, researchers reported in late 2017 that they’d found 11 lost cities in modern day Iraq by analyzing quantitative data etched on ancient clay tablets.
Harvard assyriologist Gojko Barjamovic is part of the Old Assyrian Text Project, which focuses on the earliest documented long-distance trade network in history: a web of commerce that spanned what is now Iraq, Syria and Turkey during the Middle Bronze Age (roughly 1945-1730 BC).
Over more than a century of digging, assyriologists have translated more than 23,500 clay tablets linked to Old Assyrian culture. The tablets are inscribed in a cuneiform alphabet called Akkadian, and some record shipments of gold, silver, wool and other materials between urban centers.
“It’s sheer luck that these clay tablets are even still here,” said Barjamovic. It’s also fortunate that Akkadian can be translated almost word-for-word into English.
Barjamovic digitized the content of about 12,000 tablets to make it easier and quicker for archaeologists to analyze the ancient data using computers.
That allowed him to focus on tablets depicting cargo shipments between two cities, which would help him to search for other cities along ancient trade routes.
He identified 253 tablets describing more than 300 shipments of goods between 29 cities, including the trade hub of Kanesh in the middle of modern Turkey, and 12 others whose locations had been lost over time.
Applying AI to Find Ancient Secrets
With access to supercomputers in Chicago and France, Barjamovic and his colleagues applied a software algorithm based on geography and modern statistical economic models. It started with the assumption that cities traded more often with nearby cities.
The algorithm accessed trade volume and other data, such as commodity prices and population size, to estimate the distance between each pair of cities. Barjamovic said knowing how far apart two points are doesn’t tell how they are oriented to each other, even if the location of one city is known. But by combining all the estimates at once, Barjamovic and a team of economists could triangulate the rough locations of the unknown cities.
However, Barjamovic underestimated how complex the math would be.
“In the beginning, I thought it could be done with a pocket calculator.”
These researchers solved a classification problem, said Shashi Jain, innovation manager at Intel’s Software and Services Group. Jain’s team of AI engineers and data scientists work with the NASA Frontier Development Lab (FDL) to build complex maps of the lunar poles.
“If you know coordinates then you can predict the presence of other features based on the data set you do have,” said Jain. “They can figure out an average radius and then calculate out those intersections using that data.”
He said technology gives archaeologists a different way of thinking.
“Rather than looking at things through a cultural lens, researchers are applying statistical methods to data and coming up with something they couldn’t see before,” said Jain.
Data Crunching Points the Way
Because the model was highly non-linear — every trade link depended on every other one simultaneously — the calculations required serious computing power, according to economist Thomas Chaney of Sciences Po in Paris, one of Barjamovic’s collaborators.
The team crunched some data on the University of Chicago’s Acropolis server, with 500-worker clusters and as much as 100 Gb of memory. Chaney said each of the roughly 50 estimates took at least a day to run.
“From time to time, Gojko would find a new treasure trove of ancient texts to read, and that would increase the size of our dataset, allowing us to re-run everything with better data,” said Chaney.
Accuracy of the Algorithm
To check the model’s accuracy, the team ran the algorithm on three ancient cities whose locations were known. It got two out of three right. It identified the cities closer to the center of the trade network, which suggests the estimates become less accurate as data gets sparse toward the edge of the network.
“If you take out the known cities and let the computer do the computation, it finds them back,” Barjamovic said. “That’s a pretty good sign we’re on to something.”
City location wasn’t the only thing the project examined, said Barjamovic. A debate has long raged over what drove ancient cities to thrive and grow. Access to resources like minerals, fertile fields and rivers feed an economy, but the Old Assyrian Project’s results suggest that, at least in the region, the critical factor is where a city is located in a transportation network.
Barjamovic said cities at the intersections of natural trade routes do better than cities located far away from trade routes. He pointed out that cities in Turkey have existed 4,000 years to the modern day while others disappeared.
The new approach won’t make finding lost cities as easy as playing Pokémon Go, but it could complement more traditional methods. Barjamovic said it could be used anywhere with a large body of existing historical data, including Italian merchants in the Middle Ages and Jewish traders in Cairo.
More Quality Data, Better Results
The next stage of the Old Assyrian Project is to digitize more tablets.
“We need massive data — 12,000 texts is barely enough,” Barjamovic said.
As to whether Google Translate will someday have an “Old Assyrian” option, that probably won’t happen, he said. Translating Akkadian requires the judgement of an expert linguist, since many inscriptions are incomplete and some words can be written in a variety of different ways.
But it might be possible to look at the social side of the trade network by figuring out who wrote specific tablets. Roughly a third of the texts are letters and include the names of senders and receivers. Different merchants had distinct writing “styles,” even when stamping the same word into clay with a stylus.
But assembling a handwriting database large enough for statistical analysis means tagging millions of symbols individually — which is why Barjamovic has been working with historian Edward Stratford of Brigham Young University who is currently developing a crowdsourced database project called Tablet Ninjas.
“Humans are very good at pattern matching, and I have found that some humans like to help out with arcane languages,” Stratford said. He and a graduate student already developed a basic working model, and they are currently building an Android app.
If the team could extract the social information contained in the text, Barjamovic said, it could help them better understand small groups of intimately linked people along ancient trade routes.
“Why do rich people become richer? You could feasibly ask that of this 4,000-year-old material.”