Genesys Cloud - Main

 View Only

Sign Up

  • 1.  Overlapping Topics and accurate hit counts in Genesys Intent Miner

    Posted 22 days ago

    We currently have around 35 Topics created based on intents, but we're running into challenges with overlapping coverage - many phrases and utterances seem to trigger multiple topics. Because of this, we're unable to get an exclusive count of hits per Topic.

    We're looking for best practices or configuration suggestions on how to:

    1. Segregate Programs based on major call drivers so that intent clusters are more distinct.

    2. Use quotes or specific word filters to narrow down phrases.

    3. Explore time-based operators or any other options that can help refine the Topics and improve data accuracy.

    Has anyone faced a similar issue, or can suggest approaches to minimize overlaps and improve exclusivity in topic hits?

    Thanks in advance for your help!


    #Reporting/Analytics
    #WEM-Quality,WFM,Gamification,etc

    ------------------------------
    Sravanthi T
    ------------------------------


  • 2.  RE: Overlapping Topics and accurate hit counts in Genesys Intent Miner

    Posted 22 days ago

    Hello Sravanthi, 

    I might have some ideas that could help you. Remember that NLU is fundamentally a classification problem where the model assigns probabilities to each intent. The system requires a minimum confidence score, usually around 40%, to classify an input. That means there must be a review of the full probability distribution when troubleshooting, not just the top-scoring intent. 

    Some basic strategies to minimize intent overlap would be intent refinement and technical configuration. Ill start with intent refinement:

    1. Define clear, non-overlapping intents with distinct purposes
    2. Establish clear boundaries between topics
    3. consider merging similar topics or refining their definitions
    4. review and potentially consolidate your 35 topics into broader, more distinct categories. 

    Next I will talk about technical configuration. Use specific filters and operators. This means implementing exact phrase matching quotes by adding mandatory keywords and phrases. I would also consider using exclusion rules to prevent cross-triggering. I also recommend using time-based parameters, you can start with setting specific time windows for certain topics and use temporal context to differentiate similar intents. 

    Helpful resource center doc: best-practices-to-build-and-test-your-natural-language-understanding

    Cheers, 



    ------------------------------
    Cameron
    Online Community Manager/Moderator
    ------------------------------

    Original Message


Related Content