This article details a few best practices that can be adhered to for building sound NLU models.

Starting your NLU Journey

ServiceNow provides complete out-of-box NLU models for you to use with your Virtual Agent. When creating your own NLU model, here are some tips and best practices to consider that can help steer you on the right path in your model-building journey.

1. Review release notes and product documentation around the latest NLU features and product offerings. You won't know how to improve your NLU model if you don't know about the latest features and tools available to you. Check out our documentation and release notes (as of the San Diego release) below.
   NLU documentation: Link
   Release Notes: Link
2. Leverage AI features to help in the model-building process by using available platform tools to assess and finalize your intents​
   
   • In the documentation and release notes above, you'll see various tools to help you tune your NLU model, including:
     
     • Intent Discovery gives you top intent clusters against historical data​
       
     • Topic Recommendation helps enable off-the-shelf VA topics and NLU intents​
       
     • Conflict Resolution identifies overlapping samples in an existing model, potentially exposing ambiguous intents​
       
   • Skip building intents that can be self-solved with AI​
     
     • Leverage AI Search capabilities​ to intelligently self-serve users via AI Search
       
     • Understand the knowledge gaps in your knowledge base​ and address them for a better search experience
       
     • Choose NLU Entities over separate intents and topics for the same or better user experience​
3. Adhere to recommended guide rails​ as outlined below in this article​.
4. Use vocabulary items and sources to teach your model. Refer to the 'Using Vocabulary in your Models' sub section under 'Adhere to specific guide rails...' section below for some great tips on using vocabulary
5. Use an iterative approach to test and tune your models prior to the final Publish of your model. Leverage features such as Model Optimize and Expert Feedback Loop to effortlessly tune your model during final publish.
   
   • Refer to the 'Additional NLU Resources' section at the bottom of this article for links to other community articles on NLU testing and tuning capabilities.

Adhere to specific guide rails during your NLU journey

Here is a graphical illustration of the NLU implementation life cycle:

As you start building out the model, below are some tips and best practices that help build a clear and optimally performing model.

Specific NLU Model Building Tips

• * No misspellings in model train samples
• 
  
  • Address acronyms and industry specific terms through vocabulary
• 
• Avoid ambiguity in your model
• 
• * * Handle overlap between intents
  
  • 
    
  • 
    
  • 
    
  • 
    
    • Avoid samples with inadequate context
      
  • 
    
  • 
    
  • 
    
  • 
    
  • * Avoid combining samples into a single utterance

• Create Balanced Intents
  
  • Follow intent structure recommendations

• * Follow Intent size best practices
• 
  
  • Follow best practices for train utterances
• 
  
  • Add variety to your intent samples

• Using Vocabulary in your model

This section builds on NLU Best Practice - Using Vocabulary & Vocabulary Sources to provide additional tips and guidance for when and how to use vocabulary in your models.

Different Vocabulary Types in NLU Models

• * Specific Vocabulary Tips and Usage Examples
• 
  
  • * Choosing between Regular and Pattern Vocabulary
• 
• 
• * * Understand when vocabulary is needed
• 
• 
  
  • * Vocabulary List - How-tos
    
  • 
    
    • Benefits of a List Vocabulary Source

Additional NLU Related Resources:

Additional NLU troubleshooting KBs:

Conversational Interfaces Academy on NLU Tuning