Get a table's contents from a Grobid <table> node — .tei_table

Grobid renders a table as <row>/<cell> elements, with a cols attribute on a cell that spans several columns (GroBid's own equivalent of an HTML colspan). Reads the header as one or more leading rows and every following row as data (see below), expanding a spanning cell's text across the columns it covers (the same grid-alignment approach .stat_table_parse () uses for JASP/jamovi colspans in R/stat-tables.R), so a header cell that spans two data columns still lines up with both of them rather than shifting everything after it.

Usage

.tei_table_contents(tab_node)

Arguments

tab_node: the <table> xml2 node (or NA/NULL when the table was not found)

Value

a list of character vectors, list(headers_row, ...data_rows) (the shape inst/schema/paper.json's table.contents declares), or NULL when the node is absent or has no rows

Details

Grobid's <row>/<cell> structure makes no header/data distinction at all (unlike HTML's <th>/<td>) — but a real table's header can genuinely span SEVERAL leading rows (a super-header naming a group of columns, a sub-header, then the actual column names — confirmed against a real corpus paper's regression table: "Association with mean reappraisal" / "" / "\(\beta\)01 (SE)", three leading rows before any data). Treating only the first <row> as the header (as an earlier version of this function did) reads the REAL column names as if they were a data row, and scans genuine data as if it were headers. The leading rows are instead scanned from the top: every row whose cells are MOSTLY non-numeric (prose/blank, the shape a header or super-header row has) is folded into the header, stopping at the first row that looks like real data (mostly numeric/bracket/dash cells) — mirroring .stat_table_parse()'s own "last distinct-label row is the real header" heuristic, adapted for Grobid's lack of <th>. Folded header rows are joined per column with " / " (so "Association with" / "mean reappraisal" / "\(\beta\)01 (SE)" becomes one combined column name), which keeps the super-header's grouping information rather than discarding it.

Ragged rows (a row with fewer/more cells than the header — GroBid's PDF-layout table detection is imperfect) are padded/truncated to the header's width rather than erroring, so one malformed row does not lose the rest of the table.