How to reduce the number of false positive matches in my topics?

What are the "false positives"?

Sometimes you might encounter results in your topics that do not belong there. We call these types of results "false positives".

Let's say you are tracking "Account Cancellations" and you have two results in your topic:

  • "I would like to cancel my account, please." (CORRECT)
  • "Can we please cancel our call and reschedule for another time?" (FALSE POSITIVE)

The first result is correct (a true positive) as it relates to account cancellation. The second result is a false positive because it refers to cancelling a call, rather than an account subscription.

Main causes

There is a variety of reasons why you might see false positives in your topics or subtopics:

  1. Dataset is not clean
  2. Keyword criteria is too broad
  3. User Intent prediction inaccuracies

We will explore each of these causes and ways to reduce false positives.

1. Dataset is not clean

The foundation of accurate text analysis is a clean dataset that contains only the results that reflect the voice of your customers. 

Your dataset is not clean if your false-positive results match on unwanted "spammy" messages such as email newsletters, solicitations, transactional automated messages or agent messages.

How to fix it?

2. Keyword criteria is too broad

Playvox Customer AI will return all results that match the keyword criteria set by you. If you addressed the data cleanliness issues but still seeing false-positive results in your topic, it means that the keyword criteria is too broad.

How to fix it?

  • Use more specific keywords (e.g. instead "cancel" try "cancel account") (learn more)
  • Consider using "exact match" instead of "broad match" when adding keywords (learn more)
  • Use keyword exclusions to remove unwanted results

3. User Intent prediction inaccuracies

If you are using one of our User Intents as your topic criteria and seeing false-positive results, it might mean that the User Intent model needs more training to properly calibrate.

How to fix it?

  • Use the User Intent Training tool to increase the accuracy of the User Intent model
  • Consider adding keyword criteria to your topics to further narrow down your results (learn more)

Bonus: manually remove false positives

If you have exhausted all the options above, consider manually removing unwanted mentions from your topics and subtopics. You can learn more about how to do this here.

If you are struggling with managing false positives, please do not hesitate to contact support so we can investigate and your particular case and offer tailored tips.