I've written a blog post The Data Citation Corpus revisited (with doi:10.59350/wvwva-v7125) that highlights some problems with the corpus. The key points are:
- The same repository can appear under different names (e.g., “Figshare” and “figshare”), which will affect metrics of repository citation
- Some repositories are "branded" rather than being distinct entities. For example "Taylor & Francis" is actually a branded instance of Figshare https://tandf.figshare.com/
- The vast majority of data citations are unique, that is, an item is cited just once. At least some of these are not citations in the sense of reusing existing data, but rather the link between a paper and supplementary data it has published. It might be useful to distinguish between these to events (citation of existing data versus publication of new data)
- The data that is most highly cited is almost certainly not data. For example "LY294002" has the most citations and is a protein inhibitor. The algorithms used to identify data citations are picking up acronyms for chemicals and cells, not data.
There is obviously scope for some careful data curation to fix these problems.
I've written a blog post The Data Citation Corpus revisited (with doi:10.59350/wvwva-v7125) that highlights some problems with the corpus. The key points are:
There is obviously scope for some careful data curation to fix these problems.