Reviewing and Correcting Individual AI Classifications¶

Editing and Refining Categories in Text Analytics covers fixing the category set itself - renaming, merging, splitting. This is the other half of quality control: a single response that's been assigned to the wrong category, even though the category set itself is fine. That's a per-response fix, not a category-set fix, and it works differently.

When to Reach for This¶

A response-level correction is the right tool when the categories themselves are working well overall, but one specific response clearly landed in the wrong bucket - a borderline answer that could reasonably go either way, or a genuine misread of what the respondent meant. If you're instead noticing a pattern across many responses, that's usually a sign the category set itself needs adjusting - see Editing and Refining Categories in Text Analytics for that case.

Finding Responses Worth Reviewing¶

Rather than reading every classified response, start with the ones flagged with lower confidence - see Understanding AI Confidence Scores in Text Analytics for how to sort or filter your results that way. This is the fastest path to catching the responses most likely to need a correction.

Making a Correction¶

From your classified results, open a specific response and reassign it to a different category from your existing set. The original verbatim text stays exactly as the respondent wrote it - a correction only changes which category the response is filed under, never the response itself.

Correct After Your Categories Are Final, Not Before¶

This is the one sequencing detail worth knowing: re-running classification for a question - which happens whenever you edit your category set and classify again (see Editing and Refining Categories in Text Analytics) - reclassifies every response in that question from scratch, including ones you've already manually corrected. A correction made before that kind of re-run doesn't carry forward automatically.

The practical takeaway: finish shaping your category set and run classification until you're not planning further category edits, and do your manual response-level review and correction after that point. Corrections made once you're done editing categories are safe from being reset by routine use of the product going forward - it's specifically a fresh classification run that resets them, not normal review or browsing.

Spot-Checking Even High-Confidence Results¶

Confidence scores are a guide for where to focus, not a guarantee that everything else is right. For anything going into a report you'll stand behind publicly, a quick skim across a sample of high-confidence responses is still worth the few minutes it takes - not because they're likely wrong, but because "likely" isn't the same as "certain."

FAQ¶

Can I undo a manual correction?
Yes - a corrected response can be reassigned again, either back to its original category or to a third option, the same way you made the first correction.

Does correcting one response affect how similar responses get classified in the future?
No - a correction applies to that specific response only. It doesn't retrain or adjust how the category set is applied to other responses, past or future.

Can more than one person review and correct classifications on the same survey?
Classification results are shared at the survey level, so anyone with access to the survey can review and correct responses - corrections made by one person are visible to others with access.

Does correcting a classification cost credits?
No - manual correction is a free action. Credits apply to the initial category generation and classification runs, not to reviewing or correcting their output afterward.

Is there a way to see which responses I've already reviewed?
Check your results view for whether reviewed or manually corrected responses are visually distinguished from ones still carrying their original AI-assigned category.

Once you've reviewed what needs correcting, see Exporting Classified Responses with Confidence Scores to bring the finished results into a report or hand them to someone else.