this video shows how to use natural language understanding workbench to create an nlu model for virtual agent an nlu model Maps users phrases or utterances to specific things they want to do or intents like changing their password or ordering a new laptop the model also identifies details of an utterance or entities like a specific laptop brand or time or location virtual agent uses those intents and entities to initiate the actions that users request creating a model in the nlu workbench has three phases first you build and train your model to recognize the utterances we expect to get from users next you test and refine it to make sure it's working when it's ready you publish it to start using it with virtual agent and finally when it's in use you tune in to recognize users intents better there are a few different ways to build your model you can start with a pre-built model and tailor the intents and utterances to your business and nlu comes with pre-built models for a variety of domains like it service management human resources and customer service management you can also create a model by importing data from a CSV file or start from a blank model to help get started you can use the intent Discovery application to analyze user chat or incident data and identify the top intents to add to your model here's the model we'll use for our demo we'll open it and start the first phase build and train these are the intents in this model we create a set of utterances for each intent phrases that users might actually enter in Virtual agent here are the utterances for the resolve incidence intent we'll want at least 15 distinct utterances for each intent while we're doing this we can try specific utterances to see how our model handles them the intense page shows you which intents need more utterances and whether there are any conflicts between intents as you update the data for your model be sure to train to incorporate your changes into the model you'll also create a test set a set of test utterances to help you see how your model is performing when you've built and trained your model you'll go on to the next phase test and publish we'll test our model using our test set the test score shows how well your model handles the test utterances you'll use the test results to refine your model and then train and test it again when you're satisfied with the test score you'll publish the model to make it available to Virtual agent finally you'll use Virtual agent designer to map each intent in your model to a specific virtual agent topic now you're ready to start using virtual agent with your nlu model when your model is in use you can tune it using the expert feedback loop in the nlu advanced features to continuously improve its performance every month natural language understanding workbench lets you review a subset of the utterances users entered and the intent that your model predicted you can confirm correct predictions and specify the correct intent when the prediction is wrong you can easily incorporate that feedback directly into your model and test set that's how natural language understanding workbench for virtual agent helps you build train and test your nlu model and tune it when it's in use for more information see our product documentation or knowledge base or ask a question in the servicenow community