The train and test sub-commands in the NLU.DevOps CLI tool both accept an --model-settings parameter, which allows the user to configure entity types, as well as other NLU provider-specific features such as builtin intents and phrase lists.
The --model-settings for LUIS is either an exported LUIS model or a JSON object with two properties, appTemplate and prebuiltEntityTypes. The latter allows you to specify a partial LUIS JSON model in appTemplate with identical schema to the LUIS import JSON schema. You can also supply mappings from the from user supplied names for entity types to LUIS prebuilt entity type names in the prebuiltEntityTypes property.
To configure an entity type as a closed list from LUIS, add the following to the appTemplate property of your LUIS settings JSON (i.e., the --model-settings file supplied to the train command):
{
"appTemplate": {
"closedLists": [
{
"name": "Genre",
"subLists": [
{
"canonicalForm": "country",
"list": []
},
{
"canonicalForm": "hip hop",
"list": []
},
{
"canonicalForm": "jazz",
"list": []
}
]
}
]
}
}To configure an entity type as a prebuilt domain entity from LUIS, add the following to the appTemplate property of your LUIS settings JSON (i.e., the --model-settings file supplied to the train command):
{
"appTemplate": {
"entities": [
{
"name": "Genre",
"inherits": {
"domain_name": "Music",
"model_name": "Genre"
}
}
]
}
}An alternative to closed lists for specifying entity types is to create a "simple" entity (i.e., an item in the entities property of the LUIS app JSON), and configure a phrase list that captures an non-exhaustive list of valid matches. To achieve this, add the following to the appTemplate property of your LUIS settings JSON (i.e., the --model-settings file supplied to the train command):
{
"appTemplate": {
"entities": [
{
"name": "Genre"
}
],
"model_features": [
{
"name": "genre",
"mode": false,
"words": "hip hop,jazz,pop,country",
"activated": true
}
]
}
}Prebuilt entity types in LUIS are handled in a slightly different way than other entity types. Specifically, users cannot supply a custom name for prebuilt entity types, and the type label for recognized prebuilt entities in LUIS is prefixed with builtin.. E.g., if you add the personName prebuilt entity to a LUIS app, labeled entities returned from the model will have type builtin.personName. The expectation of the utterances model is that users supply their own name for all entity types, including those that will be configured as prebuilt entity types on LUIS. To overcome the limitation on LUIS, we've added a settings property called prebuiltEntityTypes. This property includes all mappings from user supplied names for entity types to the LUIS prebuilt entity type name.
E.g., for the following training utterances:
[
{
"text": "say hi to bill",
"intent": "SendMessage",
"entities": [
{
"entityType": "Recipient",
"matchText": "bill",
"matchIndex": 0
}
]
},
{
"text": "tell anne thanks",
"intent": "SendMessage",
"entities": [
{
"entityType": "Recipient",
"matchText": "anne",
"matchIndex": 0
}
]
},
{
"text": "say goodbye to mary",
"intent": "SendMessage",
"entities": [
{
"entityType": "Recipient",
"matchText": "mary",
"matchIndex": 0
}
]
}
]To configure the Recipient entity type as a LUIS prebuilt entity, the --model-settings file supplied to the train command should look like:
{
"prebuiltEntityTypes": {
"Recipient": "personName"
},
"appTemplate": {
"prebuiltEntities": [
{
"name": "personName"
}
]
}
}During testing, to ensure that the entity types returned from LUIS are remapped back to the user-supplied entity type name, you must also supply the LUIS settings file above to the test command via the --model-settings option.
In order to label a LUIS role on an entity in a generic utterance, you can label the entityType using the role name, but you must also map each role name to the entity type in the [roles] object in the LUIS model settings. For example, if you have the following utterance with a role-labeled entity:
{
"text": "play two songs from my play list",
"intent": "PlayMusic",
"entities": [
{
"matchText": "two",
"entityType": "songCount"
}
]
}Then you can also map the songCount role to the number defined entity type (or any other LUIS entity type supporting roles) using the following configuration in the model settings:
{
"roles": {
"songCount": "number"
}
}This will result in the labeled entity being sent to LUIS as follows:
{
"startPos": 5,
"endPos": 7,
"entity": "number",
"role": "songCount"
}In both LUIS v2 and v3, if a role is recognized, the entityType property will use the role name instead of the entity type.
To configure a builtin intent for LUIS, add the following to the appTemplate property of your LUIS settings JSON (i.e., the --model-settings file supplied to the train command):
{
"appTemplate": {
"intents": [
{
"name": "Skip",
"inherits": {
"domain_name": "Music",
"model_name": "SkipForward"
}
}
],
"utterances": [
{
"text": "next song",
"intent": "Skip",
"entities": []
}
]
}
}Note that for LUIS, you will also need to include at least one utterance for each configured intent, including builtin intents. If you do not wish to include utterances in the generic utterances file supplied to the train command, it's fine to add these to the appTemplate section as well. We've done this in the sample above.
(Optional) The mapping from user-supplied entity type names to LUIS prebuilt entity names.
See Configuring prebuilt entities for more information.
(Optional) The partial LUIS import JSON that will be merged into the model before importing.
Note, if you plan to train and test with entities, you must supply entity type configurations through this property. See Configuring LUIS entities for more information.