A new research poster from the School of Underlying Conditions asks whose names a transcription model can give away once it has learned to read a community's correspondence, and finds the answer was settled by the community's own writing habits long before training began.
A handwriting recogniser trained on a community letters archive remembered only the organisers who signed in full
When a community asks for its letters to be made readable by machine, the harder question is who else the machine will then be able to read. The Names That Were Signed, a new research poster from the School of Underlying Conditions, takes up that question inside one support collective’s archive, at the collective’s invitation and under its oversight throughout.
The poster is led by Dr Brynjar Obiora, Senior Research Fellow, with Dr Mirela Hanke, Postdoctoral Fellow and Deputy Convenor of the Living Dashboard, and Dr Iben Chikere, Lecturer, both of whom bring their own schools’ habits of attention to the work.
The collective decided a long time ago who could be named in its letters and who could not. Our part was to make sure the model honoured a decision that had already been made, and made well.
— Dr Brynjar Obiora, Senior Research Fellow, School of Underlying Conditions
Bringing privacy evaluation, archival stewardship and the long-run reading of handwritten records into a single line of inquiry, the team tested what a trained recogniser retains of the people whose words it learned from. The answer turned on form rather than rarity: names signed in full and repeated on the newsletter’s closing page carried nearly all of the exposure, while members who wrote as initials remained out of reach.
The University sees in the poster a careful extension of its growing practice of custody-first research, and a fitting place for its earth sciences school to bring a surveyor’s patience to a boundary drawn by others.
Dr Hanke described the collaboration as “a reminder that the most important design decisions in an archive are usually made by the people who write it”.
The full poster is available from the University’s research repository under an open licence, doi:10.5555/slop.juvgwh.