I began a work placement with the Two Centuries of Indian Print project from the British Library working with my supervisor (Digital Curator) Tom Derrick, to automatically transcribe the Library’s Bengali books digitised and catalogued as part of the project. The OCR application we use for transcription is Transkribus, a leading text recognition application for historical documents. We also use a Google Sheet to instantly update each book’s basic information and job status.
In the first two days, I accepted training in how to use the Transkribus application by a face-to-face (virtual) demonstration from my supervisor since I didn't know how to use OCR. He also provided a manual for me to refer to in my practice. There are three main steps to complete a book transcription: uploading books, running layout analysis, and running text detection. We upload books from the British Library’s IIIF image viewer to Transkribus. I needed to first confirm the name and digital system number of a book from our team’s shared Google Sheet so that I could find the digital content of this book within the BL online catalogue. I would record the number of pages the book has into the Google Sheet at the same time. Then I copied the URL of the IIIF manifest and import this book into the collection of our project in Transkribus. After that, I would run layout analysis in Transkribus. It usually takes several minutes to run, and the more pages there are the more time it will take. Perfect layout analysis is where there is one baseline for each line of text on a page.
Although Transkribus is trained on 100+ pages, it still makes mistakes due to multiple causes. Title or chapter headers whose font size differs significantly from other text sometimes would be missed; patterned dividers and borders in the title page will easily been incorrectly identified as text; sometimes the color of paper is too dark, making it difficult to recognize the text. In these cases, the user needs to manually revise the recognition result. After checking the quality of the text analysis, I could then run text recognition. The final step is to check the results of the text recognition and update the Google Sheet.
Above: A view of a book in the Transkribus application, showing the page images and transcription underneath
During the three weeks of the placement, I handled a total of twelve books. In addition to the regular progression patterns described earlier, I was fortunate to come across several books that required special handling and used them to learn how to handle various situations. For example, the image above shows the result of text recognition for a page of the first book I dealt with in Transkribus, Dhārāpāta: prathama bhāg. Pāṭhaśālastha śiśu digera śikshārtha/ Cintāmani Pāl. Every word in this book is very short and widely spaced, making it very difficult for Transkribus to identify the layout. Because the book is only 28 pages long, I manually labeled all the layouts.
In addition to my work, I have had the pleasure of interacting with many British Library curators and investigators who are engaged in digitization. I attended a regular meeting of our project and learnt the division of labor of the digital project members. Besides, my supervisor Tom contacted some colleagues who work related to the digitization of Chinese collections and provided me with the opportunity to meet them, which has benefited me a lot.
The Principal Investigator for our 2CIP project, Adi, who also has been involved with research and development of Chinese OCR/HTR at the British Library, shared with me the challenges of Chinese OCR/HTR and the progress of current research at the British Library.
Curator for the International Dunhuang Project, Melodie, and a project manager, Tan, presented the research content and outcomes of the project. This project has many partner institutions in different countries that have collections related to the Silk Road. It is a very meaningful digitization project and I admire the development of this project.
The lead Curator for the British Library’s Chinese collections, Sara, introduced different types of Chinese collections and some representative collections in the British Library to me. She also shared with me the objective problems they would encounter when digitizing collections.
Three weeks passed quickly and I gained a lot from my experience at the British Library. In addition to the specifics of how to use Transkribus for text recognition, I have learned about the achievements and problems faced in digitizing Chinese collections from a variety of perspectives.
This is a guest post by UCL Digital Humanities MSc student Xinran Gu.