READ revolutionizes access to handwritten documents
From the Middle Ages to today, from old Greek to modern English, from running text to tables or forms
Our latest milestone has put a big smile on our faces – there are now over 20,000 registered users of our Transkribus platform for Handwritten Text Recognition! People are working with Transkribus across the globe, using it to train hundreds of models to recognise texts of diverse dates, languages and styles.
Across the course of the READ project, we have welcomed over 13,000 new users of the platform and created a formidable network of people interested in opening up access to historical documents. We look forward to continued growth as we move into our next phase with READ COOP.
The Royal College of Physicians has been devoted to advancing medicine for the past 500 years and has amassed outstanding historical collections of rare books, medical instruments and medicinal plant specimens.
The RCP has recently digitised the 6000 sheets from the (mostly) nineteenth-century Herbarium of the Pharmaceutical Society of Great Britain. This collection comprises thousands of preserved plant specimens and their associated labels.
Dr Michael de Swiet, Dr Henry Oakley and Professor Anthony Dayan of the RCP then decided to work with the Transkribus team to try to recognise the text from the Herbarium collection.
The documents present various challenges for Handwritten Text Recognition (HTR) technology. They contain a mix of printed and handwritten text (in ink and pencil), various languages, abbreviations and specialist vocabulary. They are also written in several (similar) hands.
A first HTR model was trained on 29,083 transcribed words from the collection, using the pre-existing ‘English Writing M1’ model as part of the training process. The ‘English Writing M1’ model is trained to recognise the writing of the English philosopher Jeremy Bentham (1748 – 1832) and his secretaries – it is freely available to all Transkribus users for their experiments.
In the best cases, the resulting model can automatically transcribe pages from the collection with a Character Error Rate (CER) of around 10%.
The team at the RCP are pleased with these results and would be happy if they could be shared and improved upon by other people working with Herbarium material. If you would like to find out more about their work or have access to their HTR model, please contact the Transkribus team (firstname.lastname@example.org).