Skip to main content
Back to NLP Topics The platform continuously validates your NLP training and surfaces recommendations to proactively fix issues before they affect production accuracy.

​
Validation Types

Go to Natural Language > Training > Intents to see recommendations per intent.
ValidationSeverityDescription
Untrained IntentsWarningIntent has no utterances.
Inadequate Training UtterancesWarningIntent has too few utterances.
Utterance Does Not Qualify Any Intent (false negative)WarningUtterance can’t predict any intent, including its own.
Utterance Predicts Wrong Intent (false positive)WarningUtterance predicts a different intent than trained.
Utterance Predicts Intent with Low ConfidenceWarningLow confidence score for the utterance.
Incorrect PatternsErrorPattern has invalid syntax.
Wrong Entity AnnotationsErrorEntity annotated with an incompatible value type.
Short UtteranceWarningUtterance has ≤ 2 words.
Severity:
  • Error — definite problem impacting accuracy. High severity; must fix.
  • Warning — issue that can be resolved; moderate severity.

​
View Recommendations

  1. Go to Natural Language > Training > Intents tab.
  2. Hover over the Error or Warning icon on an intent to see the recommendation.
  3. Click the intent to see the Bulb icon summary for all issues.

​
Validate Model

Click Validate Model (top-right on the Training page) to open the Recommendations Summary.

​
Recommendations Summary

Shows all current issues with suggested actions.
  • All tab — all recommendations.
  • Error tab — error-type recommendations only.
  • Warning tab — warning-type recommendations only.

​
Filter Recommendations

  1. Hover over a recommendation and click Filter.
  2. The Training panel shows only intents with that issue type.
  3. Click +Utterance, +Pattern, or +Rule to fix directly from the panel.
  4. As issues are resolved, counts update automatically.
  5. Click CLEAR FILTER to reset.
Only one filter can be active at a time. The filter applies to intent-level, utterance, rule, and pattern recommendations.

​
Refresh Recommendations

Click the Refresh icon to update recommendations after:
  • Modifying training data.
  • Adding new intents or utterances.
  • Fixing a recommendation to verify it’s resolved.
If untrained utterances exist:
  • Train & Regenerate — trains the model with untrained utterances, then regenerates recommendations.
  • Regenerate — generates recommendations from the current (already trained) model without retraining.

​
Validation Options

From the Validate Model dropdown:
  • Confusion Matrix — visualize utterance-to-task match performance. See Model Validation.
  • K-fold Cross-Validation — estimate ML model generalization performance. See Model Validation.