Reading Daniel Hook’s preprint on how machines encounter the scholarly record (and his earlier work with Simon Porter on bibliometric archaeology) made me think about the metadata gaps and other issues in content created in the 1990s and 2000s.
Anyone who worked with academic SGML/XML/HTML in this period will know how many ‘sins’ were committed to get content online. For example, I remember checking parts of the British Journal of Cancer archive following its transfer to a new publisher a couple of weeks before launch and realising that most of the author affiliations were missing and weren’t in the data we had received from the previous publisher.
Take A novel TP53 splicing mutation in a Li-Fraumeni syndrome family: a patient with Wilms’ tumour is not a mutation carrier, published in BJC in 1998. The paper has eight authors and according to the PDF there are four affiliations:
JM Varley¹, G McGown¹, M Thorncroft¹, GRM White¹, KJ Tricker², AM Kelsey³, JM Birch² and DGR Evans¹˒⁴.
¹ CRC Department of Cancer Genetics, Paterson Institute for Cancer Research, Manchester ² CRC Paediatric and Familial Cancer Research Group, Royal Manchester Children’s Hospital ³ Department of Histopathology, Royal Manchester Children’s Hospital ⁴ Department of Medical Genetics, St Mary’s Hospital
The archive had to go live without the data, and the current Nature and PubMed records still have all eight authors but only one affiliation.
A lot of the content created in this period is probably still quite messy. I say ‘sins’, but at the time content conversion and subsequent checking had a lot of manual steps and the standards we have now weren’t in place. Things did get missed, which were usually minor, but occasionally errors like misplaced decimal points in drug doses resulted in calls to Electronic Production for urgent fixes. Many of us made manual fixes to websites to get things to work (or appease a difficult author) that will have been lost long ago during later updates. Publication dates versus online publication dates were often confused, especially in archive transfers, and anyone who wrangled a conference book into an online format back in the day will have a wry smile right now about what was done to make these work. Group authorship standards were still developing, and we often omitted elements simply to get the content to process and go live (MEDLINE/PubMed included only the first author’s affiliation from 1988 to 2013).
Seems like a huge opportunity for Digital Science, Scite and similar companies to use AI to identify these problems and enhance their data. Getting publishers or PMC to fix the underlying records is a different kettle of fish. Some handle this kind of work really well, but for others the content is old, the cost/benefit is small, and there are probably various hard-coded workarounds in production systems that no one wants to touch.
Or perhaps the models will eventually sort this out and provide data that matches the PDFs more closely than the publisher websites? Asked about the affiliations for the BJC paper, ChatGPT-6 Medium flags the discrepancy between the PDF and the journal website. Grok 4.7 Expert’s response matches the PMC landing page and returns only one affiliation. Claude Opus 5.5 Medium says it can’t reach PMC, so it tries to reconstruct the affiliations from other papers and gets six of the eight roughly right (YMMV).
I am sure anyone around from that era will have a few war stories to tell about what did or didn’t make it into the XML archive.
