Test and publish your model

  • Release version: Xanadu
  • Updated August 1, 2024
  • 2 minutes to read
    • Summarize
    Summarized using AI
    This content was generated using new OpenAI-powered functionality. Results are provided on an as is basis and are not guaranteed to be accurate or complete.

    Summary of Test and publish your model

    This phase enables you to evaluate your Natural Language Understanding (NLU) model's performance by testing it against a default or custom test set and then publishing it to make it accessible to other ServiceNow applications such as Virtual Agent. Testing helps identify areas for improvement by showing how well the model predicts intents from test utterances.

    Show full answer Show less

    Testing Your Model

    • Navigate to NLU Workbench > Models, select your model's application tab, and open the model.
    • Use the Test and publish your model card to run tests and view results.
    • Results include bar charts summarizing percentages of correct, correct among multiple, missed, and incorrect predictions, helping you understand model accuracy.
    • The detailed results tab lists each test utterance with its predicted intents for deeper analysis.
    • Testing requires the Multi-model Batch Testing feature, available through the NLU Workbench - Advanced Features application.
    • Confidence thresholds influence whether intents are predicted; adjusting these can optimize model behavior.

    Publishing Your Model

    • The Publish model button activates the current trained version of your model for use by other applications.
    • If your model is not trained, publishing is disabled—train your model first in the Build and Train phase.

    Advanced Testing with Multi-model Batch Testing

    This feature allows you to:

    • Test multiple models simultaneously.
    • Use different test sets beyond the default.
    • View aggregated test results for comprehensive evaluation.

    Access this via NLU Workbench > NLU Advanced Features > Multi-model Batch Testing.

    Practical Benefits for ServiceNow Customers

    • Ensures your NLU model accurately understands user intents before deployment.
    • Provides detailed insights to enhance model quality and user experience in Virtual Agent and AI Search.
    • Makes the latest trained model available for enterprise use once published.
    • Supports comparison between draft and published model versions to track improvements.

    Assess the performance of your NLU model to identify areas for improvement. Then publish your model to make it available to other applications such as Virtual Agent.

    Summary usage

    Test your Virtual Agent or AI Search model against its default test set to see how the model responds. Test results provide information you can use to improve your model.

    Note:
    Testing your model requires the Multi-model Batch Testing feature, available with the NLU Workbench - Advanced Features application from ServiceNow® Store. For more information, see Install NLU Workbench - Advanced Features.

    To test your model, navigate to NLU Workbench > Models. Select the tab for your model's application, then select the name of the model. In the Test and publish your model card, select View phase. Test and publish your model phase card

    Overview of testing and publishing your model

    The Test and publish your model phase opens in the Overview page by default. Buttons for Run new test and Publish model are located here.

    Test and publish your model overview page

    Overview provides information about a previous test run, with bar charts summarizing the test results.

    If you have earlier test runs, you can view those by selecting from the Test run date list.

    Test run date pulldown

    To drill down into the test results table, select the Detailed results tab. Each test utterance is listed in Detailed results, with its prediction.

    Understanding test results

    The test results show how your model responded to the utterances in the test set.

    Test results for a model test in the NLU Workbench.

    The bar chart shows the prediction percentages for correct, correct among multiple, missed, and incorrect:
    Percentage Description
    Correct The percentage of utterances for which your model correctly predicted the intent.

    When the model predicts no intent for utterances marked as Not relevant, that result is counted as Correct.
    Correct among multiple

    For utterances that had more than one intent predicted.

    The percentage of utterances for which the model correctly predicted the intent or intents, but also predicted intents that did not belong to the utterance.
    Missed The percentage of utterances for which your model did not predict an intent, even though there was an expected intent.
    Incorrect The percentage of utterances for which your model predicted an intent that was not correct.

    Testing can affect the model's confidence threshold. The confidence threshold determines how confident a model must be to predict an intent for an utterance. For more information on confidence thresholds, see NLU model settings.

    For information about utterances which should not have any intent predicted, see Irrelevance detection in NLU.

    Publish model

    The Publish model button makes the current version of the model available to other applications such as Virtual Agent.
    Note:
    If the model has not been trained, the Publish model button is unavailable. Return to the Build and train your model phase to train the model before publishing.

    For more information on publishing your model, see Publish your NLU model.

    Multi-model Batch Testing

    In the Test and publish your model phase, you test your model against its default test set. With Multi-model Batch Testing, you can test against other test sets, test multiple models at once, and see your test results. To use Multi-model Batch Testing, navigate to NLU Workbench > NLU Advanced Features > Multi-model Batch Testing.

    For more information, see Multi-model Batch Testing.

    For more information about test sets, see:

    For information about the process of testing, see Test your model.