We’re seeing a scoring drift on a topic - ‘order cancellation’ - after a small keyword update. Added “cancel order” and “refund request” to the phrase list. Precision hasn’t changed much, but recall dropped almost 15 percent. It’s weird, because the new keywords are pretty specific.
The flow is simple - using Zoom Contact Center’s native speech analytics. Not doing anything fancy with the SDK. Did see a similar discussion about SIP trunk jitter affecting precision, but the audio quality looks okay. Is there a setting I am missing, or a recalibration step after adding new keywords?
The API call to /api/v2/analytics/topics/{topicId}/scores seems to return consistent results, but the overall model’s recall is down. Does Zoom’s scoring algorithm penalize short phrase matches?
The scoring drift is probably connected to the minimum confidence score setting for topic detection. We had a similar case last year - someone in the community found that the default value is too high for new keywords - you should lower it and monitor the impact on both precision and recall. It’s a common issue when adding phrases, especially specific ones.
That’s right about the confidence score - it’s often the culprit, and the documentation does state, quote, “Adjusting the minimum confidence score allows you to balance precision and recall based on your specific needs” - you’ll find that in the Zoom Contact Center Speech Analytics guide. But quick one - are you sure the new phrases aren’t triggering false positives on other topics? It’s like adding a new street to a map - you need to check it doesn’t accidentally connect to the wrong destination.
Fun one today. That’s right about the confidence score - it’s often the first thing to check. But we’ve hit a similar drift, and it wasn’t confidence, it was actually the phrase weighting. Zoom Contact Center’s phrase weighting is…subtle.
Here’s how we approached it:
Understanding Phrase Weighting in Zoom Contact Center
Default Weighting: Every phrase you add initially gets a weight of 1. This means it’s equal in importance to every other phrase.
Impact of New Phrases: Adding highly specific phrases - like “cancel order” and “refund request” - can inadvertently dilute the importance of broader phrases you already have. Recall will drop because the system needs all conditions to be met for that specific phrase to fire.
The Fix: Increase the weight of your core phrases.
Implementation Details
You adjust phrase weighting within the Topic Detection configuration. It’s not exposed through the API directly - you must do it in the Zoom Contact Center UI.
Navigate to Speech Analytics → Topic Detection.
Select your ‘order cancellation’ topic.
Edit the phrases. You’ll see a Weight column.
For your existing, broader phrases (e.g., “order problem”, “issue with my order”), increase the weight to 2 or even 3.
Leave the new, more specific phrases at 1 for now.
Save your changes.
We found that increasing the core phrase weights helped restore recall without significantly impacting precision.
A couple things to watch:
Over-Weighting: Don’t go crazy with the weights. Weights greater than 3 can cause other unexpected behaviors.
Monitor Performance: It’s an iterative process. Monitor the scoring over the next 24-48 hours and adjust as needed. A slight increase in precision loss is usually acceptable if it significantly improves recall.
Phrase Overlap: Be sure the phrases aren’t accidentally overlapping in meaning. Zoom Contact Center isn’t always great at disambiguation.
fun one today. Phrase weighting - yeah, that’s where the REAL time sink is. We saw similar drift after a keyword add, and it wasn’t the confidence score at all. Zoom’s weighting algorithm is… opaque.
Here’s what bit us: new phrases → higher initial weight → disproportionate influence on topic scoring → recall looks good, but precision tanks. Lowered the weight on “cancel order” to 0.5 and “refund request” to 0.7. Monitor the API call duration - those weighting updates hit the analytics endpoint HARD.