The St. Gall Abbey Library in Switzerland houses approximately 160,000 manuscripts of literary and historical significance dating back to the 8th century.
All of these manuscripts are handwritten on parchment and in languages rarely used in modern times.
To preserve these “historical treasures” of humanity, millions of such texts have been stored in libraries and monasteries around the world. Most of the collection is made accessible to the public through digital images. However, experts believe there is a significant amount of “extraordinary” material that has never been read, written in ancient languages.
Automated transcription system for manuscript pages.
Now, researchers at the University of Notre Dame are developing an artificial neural network to read complex ancient handwriting based on human perception. Walter Scheirer, an Associate Professor in the Department of Computer Science and Engineering at Notre Dame, shared:
“We are processing historical documents from centuries ago and in languages such as Latin, which are rarely used today. What we are aiming for is the automation of page transcription, mimicking the perception through the eyes of a professional reader. At the same time, it provides the ability to read text quickly and in a searchable manner.”
In a newly published study, Scheirer outlined how his team combines traditional machine learning methods with psychophysical approaches. This method measures the relationship between physical stimuli and mental phenomena.
For instance, the time it takes for a professional reader to recognize a specific character, assess the quality of handwriting, or identify the use of certain abbreviations.
Scheirer’s team studied digitized Latin manuscripts written by scribes at the St. Gall Abbey in the 9th century. Readers entered their manual transcriptions into a specially designed software interface.
The research team then measured reaction times during the transcription process to determine which words, characters, and passages were easy or difficult to transcribe. Scheirer explained that this method has created a network that aligns better with human behavior. Consequently, it reduces errors and provides more accurate and practical text reading.
“This is a strategy that is often not utilized in machine learning. We label the data through these psychophysical measurements. They stem directly from psychological studies on perception by conducting behavioral measurements.
Then we inform the network of common difficulties that can be corrected based on those measurements,” Scheirer explained.
However, according to Associate Professor Scheirer, this method still faces many challenges. His team is working to improve the accuracy of transcriptions, especially in cases where documents are damaged or incomplete. At the same time, they are calculating other aspects when a manuscript page may confuse the system.
The encouraging signal is that the team has successfully adapted the program to transcribe Ethiopian texts. They then adjusted it to a language with a completely different character set. This is considered a first step toward developing a program capable of transcribing and translating information for users.
