In the third and last RESILIENCE training prototype, automatic keyword tagging of multimedia data was the focus. Read here the reflections of librarian René van der Werf, one of the participants.
The training was preceded by an introduction session, informing the participants of what they could expect and how they could prepare for the training. The participants had various background. By no means everyone is a librarian like me, there is also a chemist, an archaeologist, a musicologist, and so on. The training seems to be interesting to a wide range of professionals.
AI tools will be demonstrated inspiring us to use them ourselves and participants will be allowed to contribute their own content in the form of images, texts, even manuscripts. Unfortunately for the musicologists: sound files were not eligible.
The actual training could be described as an information avalanche. By the end of the day, I was overwhelmed with all the information, sometimes leading to square eyes and a heavy head. However, the information was interesting to me, so there will be a lot to digest.
One of the topics in the course was “Object Detection”. We were given quite a lot of information about the technical background. We experienced ‘first hand’ that AI by no means always works flawlessly: a picture of a statue brought in by one of the participants was really “recognized” as a particular work of art from a particular place, but … both “facts” were wrong. Here, for me, is a positive point of the training: it gave me a realistic picture: AI can do a lot, but for now it cannot be applied without human control.
Another topic that came up was LLMs, Large Language Models (such as ChatGPT). Here an idea came up that seems interesting to me for researchers and librarians: you ‘feed’ a chatbot with papers from a particular scientist and then you ask the chatbot a question along the lines of “What is this scientist’s view on …” This way you don’t have to study the scientist’s papers and you can get right to your question. I know: this raises all sorts of questions: will we be able to research ourselves later, and what about the scientist’s copyrights? Of course, we need to think about these questions, but they were not part of this training
One of the things I enjoyed most was the topic of Handwritten Character Recognition. Here we watched in real time with a trainer who previewed how she was training an AI tool to convert handwritten texts into plain, typed text. This is about religious texts in ancient handwritings, made accessible for research in this way. A great result, no doubt! However, also labor intensive: almost every sentence produced by AI had to be corrected by the trainer.
The training offered me unexpected insights into a hitherto little-known world that certainly offers opportunities for librarians. However, much will most probably be needed before I will fully rely on these tools in my daily work.
René van der Werf, librarian at the Theological University of Apeldoorn