May 2024 -- Data Cleansing and the Encounter with Records

Much of May has been working with historical maritime records rather than producing artwork directly. I miss drawing.

I've been experimenting with extracting information from Lloyd's historical casualty reports, many of which exist only as scanned PDFs as well as downloading complications of records. Parsing these documents proved unexpectedly difficult. Tables break during extraction, columns merge, dates shift position, and OCR errors introduced strange distortions into vessel names and locations. I kinda like the distortion, so am capturing it for later. But whats starting to become clear is two kinds of AI use here - trying to make art producing engines and augmenting my own workflow. It's frustrating, theres loads of back and forth and the things you really want the machine to do, clean up files, re-format, tabulate; it feels wilfully unable to do. Theres a point in reformatting where I just give up, only to have to start again and try and figure out where I am. There are of course commercial data sets but mostly either available at commercial prices or requiring a level of programming I'm not able to do.

This process is slow and frequently frustrating and everything I try doing turns out much more expensive that I can support. At the moment theres a lot of specialist products - mostly adapted from OCR for document processing.

The reports themselves are highly compressed forms of narrative. Catastrophic events are reduced to administrative language:

"Reported missing."
"Presumed lost."
"Grounded on unknown obstruction."

The emotional weight exists indirectly, carried through repetition rather than description.

By the end of the month, the work had shifted away from image production toward an engagement with records as generative material. The act of cleaning data began to feel less technical and more interpretative deciding what survives translation from archive to dataset.