Reviews are read for the gap, and a gap a model was asked to notice is an impression. This computes them instead, in R, from the extraction table and the claims: a declared category nobody studied, a dimension with no variation, an empty cell in a crosstab, a claim resting on one study that nobody has tried to replicate, a disagreement nothing in the table explains. Every row is a fact about the corpus that can be checked by counting.
Arguments
- claims
A
gr_claims()result.- extraction
The
gr_extract()result the claims came from. Only its$fieldsis used, and only to learn which categories were declared; without it a category nobody studied is indistinguishable from one nobody thought of.- max_cells
Cap on reported empty combinations, which grow as the product of two fields' sizes.
- min_reported
A field missing for more than this fraction of studies is reported as not reported.
Details
No model is called. gr_synthesise() takes the result as gaps = and hands
it to the closing section with an instruction to state these and no others,
which is what keeps the written gap list attached to the table.
Examples
tab <- data.frame(document = c("a.pdf", "b.pdf"), status = "ok",
duplicate_of = NA_character_, n_filled = 1L, n_unverified = 0L,
conflicts = NA_character_, design = c("cohort", "cohort"),
stringsAsFactors = FALSE)
cl <- gr_mock_client(function(messages, params) paste0(
'{"claims":[{"claim":"One cohort found it.","kind":"finding","supported_by":[1],',
'"contradicted_by":[],"moderator":null,"scope":"one study"}]}'))
cm <- gr_claims(tab, question = "Does it work?", client = cl)
gr_gaps(cm)
#> <gr_gaps> 2 gap(s) over 2 study/studies
#> no schema given, so a category nobody studied cannot be told from one
#> nobody asked about; pass `extraction =` for those
#> no variation 1
#> unreplicated 1