np.argmax over a 3-by-5 grid of logits (3 reviews, 5 classes). axis=-1 scans across each row for one prediction per review; axis=0 scans down each column for the winning review per class; no axis flattens everything to a single scalar. The axis you name is the axis that disappears from the output.

The axis you name is the one that collapses

A grid of scores — 3 reviews × 5 classes. argmax always finds the biggest; the axis you hand it decides across what — and that axis vanishes from the answer.