Researching and AI
I am starting research on a new project about an American woman in England named Beatrice Warde. She was one of the first woman typographers and also played a crucial role in educating much of England, and the world, about typography. She was involved with the production and sale of Times New Roman. Plus there is a fun romance I am discovering. It should make for an excellent novel.
As part of the research, I found a collection on Beatrice Warde at the University of Birmingham’s library. I was astonished to see a 27-paged document listing all the items they have from her, including many letters, journals, pictures and poems. They even have locks of hair. I wish I could travel to England and go through the boxes myself, but my current situation with young kids (besides the financial commitment) will not allow for that. Instead I followed their website till I found a form to request digital copies.
Just a couple days later, I got a response saying if I can narrow down the list, they will get it ready for me. I picked the first two boxes and listed all the items that seemed interesting, including all the letters from Beatrice’s mom to herself in 1936 from New York to England. Beatrice’s mother was May Barton Lamberton. She was an influential literary critic and children’s book author and publisher. The more I find out about this family, the more intrigued I am. I will have to continue to publish about my research, there is far too much to document here.
The library told me they only had access to the letters at the moment and charged me 50 pounds (about $67). A few days later I had letters written by May Barton Lamberton on my computer.
The only problem then was figuring out how to read her cursive with some shorthand mixed in. My husband sent me to the website Transkribus. This site will allow you to use public ai models to read text and convert to digital. I tried a few with varying results. One kept turning the language into German which was annoying. Finally I landed on one meant for English handwriting in the correct time period. It is amazing how well trained these models are.
Here you can see the numbered lines the AI model has assigned to each row.
I am going through all 44 pages of letters now with that language model. First it determines the lines. Then it can distinguish words. These models are usually word-based and will do it’s best to find a word that matches. I have found that it is 90-95% accurate. I have to go through and edit each one as I go. I find that as I do this, the ai is training me to understand the handwriting. By the end of it, I will be an expert in May Barton Lamberton’s handwriting I suppose.
Once I finish this batch of letters, I am planning on requesting more things from that collection at the University of Birmingham. I am so grateful for their promptness in responding to me. I am also grateful for the hard work Transkribus and the model creators have put into making this technology available to everyone. This work would be impossible for me to do without it.
