Shade or ignore

Is it useful to compare the values in this column?

Code sorts each column into numbers, labels, or text. Code does not send text columns to the model.

For a number column, the model tells what the numbers are. Amounts get shading. Code ignores identifiers, years, cutoffs, and records such as 12-4. It ignores ranks that count the rows (1, 2, 3 down the table), but shades other ranks, such as one outlet's ranking beside the table's own. A bracket such as 20–25 or 65+ is an amount when it is a value of the row, such as a median age. Code puts it on the scale at its midpoint. Brackets in the first column that name the rows (0–4, 5–9, 10–14) are like years: code ignores them.

For a label column, the model tells if the labels are grades in a fixed order. Only ordered labels get shading. Code ignores a column with only two labels, unless it shares a scale with a column that has more labels.

Decision rule

Shade a number column when P(amount) + P(bracket) is 0.7 or more. Shade a label column when P(ordered) is 0.7 or more and the order confidence is 0.7 or more. Also shade it when P(ordered) is 0.5 or more and the order confidence is 0.85 or more.

The questions

The model gets the question text and a short description of the table. It gives a probability for each possible answer. These are real questions from the test pages, with the recorded answers. Here, the question names the column by its header. The question that the extension sends points to the column in the table description.

Results

198 of 199 checks pass, on 32 tables. The failure is a known issue. Tables with a failure come first.

Each table shows the shading that the extension applies. The strip under the table shows all judgments on each column. Put the pointer on a result to outline its column. A red outline shows a failure.