Utterances are triggering incorrect topics
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
04-10-2025 03:44 AM
Hi all,
Context: Utterances are triggering incorrect topics when end users interact with the virtual agent.
Actions Taken:
- Submitted feedback via the NLU Workbench test panel, marking the utterance as not relevant to any model.
- Verified the feedback is present in the ml_labeled_data table.
- Added the utterance to the test set in the build and test module, selecting "not relevant" for the model.
- Reviewed all training utterances and found none similar.
- Trained the model, ran the batch test, and published the model, but the issue persists.
- Adjusted the confidence threshold.
- We are using Xanadu.
Question: What additional steps can I take to prevent this issue? I am encountering false positives that are affecting the model's accuracy.
Any guidance would be greatly appreciated.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
04-11-2025 09:39 AM
When I have this happen and after going through the steps you already stated, I go to the Intent add an Utterance, train and publish the NLU to force it to behave how I need.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Tuesday
I have found that creating an intent within the NLU that isn't associated with any topic can help with this.
For example, utterances like "My VDI isn't working" or "Can't login to VDI" was triggering the wrong intent (say it's called "AccessIssues", so I created the "VirtualDesktopIssues" intent with all these utterances (and more). I enabled that intent but didn't associate it to any topic, and now those utterances don't trigger the "AccessIssues" intent and now triggers the "VirtualDesktopIssues" intent, but because this intent isn't assocated with any topics, the NLU returns the fallback search results and present knowledge articles instead that match the utterance (which is what we've wanted it to do when a topic within our NLU doesn't exist).