Natural Language Understanding of Virtual Agent responses
- UpdatedFeb 1, 2024
- 3 minutes to read
- Washington DC
- Virtual Agent
Virtual Agent (VA) uses the Natural Language Understanding (NLU) service to understand user input. Use the NLU Prediction tab to see how well NLU predicts intents, and to improve the intents so NLU makes better predictions.
The NLU Prediction tab on the Conversational Analytics Dashboard shows how well NLU is understanding user input in chat conversations. Virtual Agent comes with NLU models, but you can use the Activate the NLU Workbench to modify or create new models.

To use the NLU Prediction tab, you must have the Chat Analytics Viewer (chat_analytics_viewer) role.
NLU Prediction graph
- Correct Predictions—NLU predicted multiple intents and showed them to users. They selected one.
- Incorrect Predictions—NLU predicted multiple intents and showed them to users. They indicated that none of them were what they were looking for.
- Auto Selected Predictions—NLU predicted a single intent based on chat input from the user. Sometimes, NLU returns multiple topics, each from a separate intent, and the user selects one.
The example graph only shows Auto Selected Predictions. On 2021-01-27, one intent was auto-selected, and on 2021-01-31, two intents were auto-selected.
- Did the NLU prediction model determine an intent?
- Did the predicted intent match the topic bound in the model to the intent?
- What's the prediction score and did the dashboard show multiple options to the end user to specify the correct intent, or did the dashboard just display the topic associated with one intent.
- After showing the user multiple intents, did the user select one?
- Correct Predictions—1, 2, 3, and 4 are true.
- Incorrect Predictions—4 was false.
- Auto Selected Predictions—3 was auto-selected, and there was no 4.
It's possible for intents to appear in several categories. For example, the Activate Account (Ben) intent appears in the Correct Predictions and the Auto Selected Predictions columns because NLU correctly predicted the intent and only presented a single response to the user.
Improving NLU predictions
Clicking anywhere on the graph opens the Model Performance page. It shows a summary of intents predicted or not predicted over the date range specified on the graph.

See Activate the NLU Workbench to see how to use the NLU Workbench to improve NLU predictions.