Negative feedback that does not point at a location cannot localise the problem

What it is

Eight hundred words, one thumb down, no span marked. It might be a number in paragraph three, the tone, the missing conclusion. People training or changing the product face a piece-level −1 and can only guess. Unlocated negative feedback cannot localise: it says “this pass was bad,” not “where it broke.”

Thumbs not meaning true/false is polarity. Here polarity is already negative; what is missing is a coordinate.

Why it happens

Localisation maps dislike onto a span. One bit for the whole piece makes the search space every sentence × every failure type. At the moment of the down click people are often staring at one place; if the UI does not take that focus, the focus evaporates and the coarsest label remains. Downstream, whether human review or a loss, can only spread the penalty across the piece, hitting right sentences with wrong ones.

Select-then-down, “problem here” beside a sentence, field-level marks, all take a coordinate in the same act as the down. With no coordinate, the information in a negative is close to “don’t show this pass again,” barely actionable for improvement.

Studying it

The same erroneous output: piece-level down versus forced select-then-down versus sentence-side marks. Have another group (or a model) revise from the feedback. Dependent variables: share of true errors hit, share of correct sentences wrongly changed, time to fix. Independent variables: whether a location is required, grain (sentence / paragraph / field).

The lead measure is “share of penalty landing on the span that was actually wrong.” Counting feedback items will treat unlocated downs as equally useful.

Where it stops holding

When failure is piece-wide (wrong genre entirely, a refusal that should have been refused), piece-level negative already localises. Output short enough to be one or two sentences makes location almost equal to the piece. If users refuse to select, keep the polarity but do not pretend localisation happened. Edits carry location by themselves — a richer signal on another card.

Applying it

• Default negative feedback to “which stretch,” with select supported; allow skip, but do not put skipped samples into location-type improvement.
• Prefer “problem here” beside a sentence or field over one down at the end of the piece.
• Open review queues on the selection, rather than throwing a piece-level −1 at reviewers to go find the fault.
• Check: take texts with gold error spans, give only piece-level downs, see whether fixers hit them. If not, unlocated feedback has no localising power. Add select and measure again; hits should rise.

Related

• Same group: L3.13.1 Thumbs up and down collect satisfaction, not correctness · L3.13.3 People who submit feedback are a self-selected few; extreme experiences are over-represented · L3.13.4 Edits are implicit feedback with more information than an explicit rating · L3.13.5 If feedback produces no visible change, submission decays toward zero
• Nearby: L3.12 Editing and Taking Over Generated Content · L3.03 Hallucination and the Fact-Checking Burden
• Search terms: span-level feedback · unlocated downvote · error localisation