A thin scan is more dangerous than a wrong one
Better Translate Team ·
You can argue with a wrong translation. You cannot order a dish you were never
shown.
Why a short result is invisible
When a scan drops lines rather than mangling them, the output looks perfect. The English is
clean, the prices line up, nothing is obviously broken. There is no error to notice, because the
error is an absence.
We measure this separately in our own benchmark for exactly that reason. Completeness is its own
number, scored apart from whether the English reads well, because a configuration that writes
beautiful English about half a menu is not a good result.
How to catch it
- Count sections, not dishes. If the board has four headings and your result
has two, retake the photograph.
- Look at the original beside each line. It stays there so you can compare the
shape of the result against the shape of the page.
- Use the overlay. Overlay mode keeps each translation where the dish sat on
the page, so a gap in the middle is visible in a way a written-out list never is.
- Retake before you retry. Glare and angle move the result further than
anything happening on our end.
The general lesson
When you are judging anything that extracts or summarises, ask what it does with the parts it is
unsure about. Dropping them scores well on every measure that only looks at the output, and it is
the failure a person is least equipped to catch.