Don't bury your research data in #PDF files. https://www.linkedin.com/posts/jonippolito_archives-collections-data-activity-7305940793683267584-_aHF
Put raw data in data repositories and use open formats to save them and make them reusable. #CSV / TSV are the perfect format for tabulated data.
www.linkedin.com#archives #collections #data #digitalcuration #digitization… | Jon IppolitoCan AI save PDF from itself? Probably not, so stop using PDF to archive your documents.
PDFs are not easily accessible, hide their dependencies and security vulnerabilities, and rely on a 700-page spec developed by a proprietary vendor rather than an open standards body.
Most importantly, they're a bitch to extract data from. (Have you tried copy-pasting text from a two-column PDF?) Along with bitmapped PDFs that are basically images, there are even searchable PDFs whose creators have accidentally or deliberately mixed up the letters in the font table to make them non-copiable. These failure points are why Arxiv started giving readers the option to read papers in HTML.
This sounds like a good job for AI, and in fact Mistral just launched an API for converting PDF to the simple, readable Markdown format. Unfortunately, independent tests have showed poor results https://lnkd.in/eyktiSaa. When you upload a PDF as background for your Custom GPT or NotebookLM podcast, this poor accessibility just compounds potential errors at inference time.
But it's a lot worse when a PDF is consulted in a life-or-death situation. Optical LLMs have been caught mismatching column headings and lines of data. Imagine a doctor who uses AI to consult a PDF table of drug protocols and prescribes 500mg of a potentially toxic chemotherapy agent instead of 50mg.
Especially in science, the value of information is the ability to re-use it. Yet archivists in the digital era have spent way too much time figuring out how to get data into a repository compared to getting it out. As I've said before, a collection of PDFs isn't an archive; it's a tomb.
#Archives #Collections #Data #DigitalCuration #Digitization #DigitalPreservation #AIethics #GenerativeAI #GenAI #LLM #MachineLearning #OCR
the web.
Dark mode.
