Intent Discovery

Use the Intent Discovery application to help identify opportunities for incident deflection. For example, you can use it to identify which Virtual Agent conversations to activate next.

Summary usage

For applications that consume NLU, such as Virtual Agent and AI Search, Intent Discovery helps you to better understand which prebuilt intents you can benefit from, and which custom intents would be useful to create.

Intent Discovery provides an analysis that you run on historic incident data or other task data. You can also group the run’s remaining records into different clusters so you can manually add utterances to NLU intents. In addition, you can use specific clusters to create new intents in a model.

In this example scenario, you're using Intent Discovery to identify the top intents in your instance, and how much coverage they can provide across your historic incident records.

Installation

Intent Discovery is available from the ServiceNow Store. For more information, see Install Intent Discovery.

After Intent Discovery is installed and activated, it appears under All > NLU Workbench > NLU Advanced Features.

Note:

Although organized under NLU Advanced Features in the navigation pane, Intent Discovery is a separate application that is not included when installing NLU Workbench - Advanced Features.

Intent Discovery report details

• When Taxonomy is selected, the generated report contains intent recommendations against the selected taxonomy. A taxonomy is a prebuilt library of intents in a specific domain. While you don't have access to the underlying intents, when you run Intent Discovery against a specific taxonomy, data that maps to any intent in the taxonomy will be identified.
• Unmatched records are the utterances which couldn't match to any intent in the taxonomy.
• Recommended intents are the intents which are found from utterances that data was run on.
• The percentage of Unmatched records (clustered) are the records that aren't classified (records that don't belong to any of the recommended intents).
• The percentage of unmatched records and the number of recommended intents don't need to match. It's a coincidence if they match.

Creating an Intent Discovery report

1. Using the admin or nlu_admin role, navigate to All > NLU Workbench > NLU Advanced Features > Intent Discovery.

2. Select either Run analysis or Find recommendations.

Figure 1. Intent Discovery landing page

Running an analysis on the report

1. For this example report, you configure the following fields on the Intent Discovery > Create new screen.

• Data Source: Select the Incident (incident) table.
• Filter by: [Created] [on] [This quarter]
• Field to analyze: Short description (short_description). You choose Short description because it's a highly used string field that references words that can help the system identify an intent.
• Taxonomy: Select ITSM. This field tells the system to run classification processing on your ITSM incident records. It has 3 options: Classification, ITSM, or blank, which defaults as Classification.
• Cluster unmapped utterances by keywords... : Select the check box. When you check this box, the system groups your incident records that weren't classified into clusters.
• Report name: The field automatically defaults to Incident <month/day/year>. You can edit the name if you prefer. In this example scenario, you enter Incident 12/16/2020 - SF Test.

2. Select Run analysis.

Figure 2. Selecting data sources in Intent Discovery for a run analysis

Result: Your report appears on the Intent Discovery screen, showing its status as the analysis begins. The subsequent status values appear in the following order during the analysis: Preparing to run, Work in progress, Clustering, and Done. This can take from 5 minutes to 30 minutes to complete. The fewer the records you have in a cluster, the less time it takes. Turning clustering off can also speed up the process.

Figure 3. An ongoing run analysis

When the analysis is complete, the column values on the screen appear, with the Status column value set to Done, as shown in the image below.

Note:

If you want to delete the report and start over, point to the right of the Status column to invoke the Delete report icon.

3. Select the Name of your report.

Figure 4. A completed run analysis

Result: The screen refreshes, showing the analyzed incident records and the remaining incident records that were not classified.

Importing recommended intents to new or existing custom models

Before importing intents to an NLU model, ensure that you are in the same application scope as the model. For more information, see Select an application from the application picker.

1. On the Records covered by recommendations section of the screen, select the caret icon on a recommended intent you want to add to a custom model.

Figure 5. Reviewing a recommended intent

Result: The details of the recommended intent appear so you can review them, as shown in the image below.

2. Select Add to Model.

Figure 6. Adding a recommended intent to a model

3. On the Select a destination model screen that appears, choose a model you want to add the recommended intent to. If you can't find an appropriate model, create a new one, return to the report, and add the new model.

Note:

The model you choose must have the same application scope as your current scope.

4. Select Save.

Figure 7. Saving a recommended intent to a model

Result: A banner appears on the screen, confirming the intent is added to the target model.

Figure 8. Confirmation of adding a recommended intent to a target model

The recommended intent also appears on the Model screen of the target model, as shown in the image below.

Figure 9. View a recommended intent in the target model

Adding clustered utterances to an intent and its model

1. On the Remaining records section of the intent discovery records screen, select and open a cluster of utterance and short description data that you want to add to an intent and its associated model.

As you continue to build out new intents from these clusters, you can click the Ignore icon to remove any unwanted intents from the report.

There's also a Show Additional filter you can use to show or hide the added intents, and the ignored intents as well.

2. Select Add to intent.

Figure 10. Adding a cluster to an intent

3. In the Add this cluster to an intent and model screen, select an intent and model pair you want to associate to this cluster.

Figure 11. Adding a cluster to an intent and model

4. Enter a few utterance examples into the open text field. Select Add each time you complete your entry to save it in the system. Use the pencil icon or the trash can icon respectively to edit or delete your entry.

5. Select Save.

Figure 12. Adding paraphrased utterances to an intent

Result: The records screen appears, showing a banner confirming you added two new utterances to the target intent and its associated model. The model and intent pair appears in the Added To column, as shown in the image below.

Figure 13. Confirmation of adding paraphrased utterances to an intent

Use the Show Additional filter if you want to show or hide the clusters that have added intents, and the clusters that are ignored.

Figure 14. Viewing or hiding clusters and ignored clusters

Running another analysis on your Intent Discovery report

1. Select Run Again.

Figure 15. Selecting the version of the analysis to run

Result: The new run begins. When it's in progress, the option to cancel the run appears, as shown in the image below.

Figure 16. The Cancel Run option

When the run is complete, a new banner appears that states you have a new version of the report.

2. Select the new version, then select Run Again.

Figure 17. Selecting the new version of the report

Result: The time stamp you selected for the most recent run appears in the Run date column of the Intent Discovery screen.

Figure 18. View the new time stamp of the Intent Discovery report

• Install Intent Discovery
  You can install the Intent Discovery application (sn_nlu_discovery) if you have the admin role.