Recognising Text, Recognising Processes - eXplainable Automatic Text Recognition (X - ATR) for Scottish Spiritualist Newspapers (1869 - 1882)

NVIVO labelled dataset of Automatic Text Recognition (ATR) provider public-facing material, consensually web crawled. This dataset was used to correlated a more qualitative reiew of ATR processes against technical CER (Character Error Rate) results, as we required an interpretive approach to fully understand the technology's local application. Therefore, each trial was correlated against a thematic content analysis of latent themes in ATR providers’ public-facing material: general publicity, model cards, README files and walkthroughs. Conducted in NVIVO, thematic labels were converted into coverage percentages, normalised based on the word count of each ATR’s documentation, and corresponded with our technical error rates. Particular concern was paid to over-reaching language about model efficiencies, as well as any reason for NLS implementation gaps such as: cost, heavy processing, or tools being reliant on closed systems. Affordances were also highlighted, including references to the wider ATR user community or a model’s multilingual capability.