Computer Vision (CV) is one of the hottest areas in machine learning research, where it has been widely applied to help machines learn to “see.” Practical applications include recognizing people’s faces in ID systems, environmental perception for autonomous vehicles, or diagnosing disease from medical scans.
An interesting new trend is to apply CV to help interpret ancient languages and scripts, such as the Japanese kuzushiji writing style. Now, researchers from the University of Chicago Oriental Institute (OI) and the Department of Computer Science have introduced an artificial intelligence tool called DeepScribe designed to read cuneiform tablets from 25 centuries ago.
Discovered by OI archaeologists in 1933, these particular tablets are inscribed with a complex, semi-alphabetic cuneiform script used by the Achaemenid Empire in Persia. Researchers have been studying and translating these ancient documents for decades, and have created a dictionary of the Elamite language, with students hand-labelling more than 100,000 cuneiform symbols.
Automating this process can significantly accelerate this time-consuming process and help researchers in the field conduct more in-depth analysis of the historical information locked in the cuneiform inscriptions.
The researchers trained a machine learning model called DeepScribe on some 6,000 annotated images from the Persepolis Fortification Archive. The model’s translation performance on the test cuneiform set has achieved about 80 percent accuracy.
The research team is continuing to improve accuracy, and hopes to eventually develop a generalized tool that archaeologists could use in studies of other ancient scripts.
A similar ancient language research project recently conducted by a group of Chinese researchers applied a multi-regional convolutional neural network to classify oracle bone rubbings, achieving accuracy close to that of experts in the field.
More information on the cuneiform research project is available on the University of Chicago website.
Author: Yuqing Li | Editor: Michael Sarazen