feat: Managed LU sample - updates to TodoBotWithLuis

[vishwacsena]

Description

Fixes #806

• Updated TodoBotWithLUIS to use Managed LU constructs
• Updated LU to the new @ entity definition syntax
• Configured all inputs to allow interruption with expressions controlling Value and AllowInterruption properties.

Documentation

Advanced dialog and language understand concepts

Conversations do not always progress is a linear fashion. Users might want to over specify information or present information out of order or would like to make corrections etc.

Bot Framework composer relies on the advanced dialog capabilities offered by Adaptive dialogs as well as relies on managing one or more backing LUIS applications to support these advanced scenarios.

This topic looks at each one of these in a bit more detail. For this doc, snippets/ concepts covered here are available in the TestManagedLU sample.

Flexible entity extraction

Let's take this simple example -

user: Hi
Bot: Hello, what is your name?

To this prompt from the bot, the user can say one of the following

user: vishwac
user: my name is vishwac
user: you can call me v

In several of these examples, we need the bot to not just verbatim take everything the user said as input but detect vishwac or v as the user name.

Within composer, you can achieve this by setting the recognizer on the dialog to LUIS and subsequently using LU notation to define user response to the specific what is your name input.

Take a look at the input and all its configured properties in the TestManagedLU under UserProfile dialog, BeginDialog trigger, AskForName input.

The input has the following configuration -

    "property": "user.name"
    "value": "=coalesce(@userName, @personName)"
    "allowInterruptions": "!@userName && !@personName"

And the following Expected response LU configuration

- my name is {@userName = vishwac}
- I'm {@userName = tom}
- you can call me {@userName = chris}
- I'm {@userName = scott} and I'm {@userAge = 36} years old
> add few patterns
- my name is {@userName}

> add entities
@ prebuilt personName hasRoles userName

Two key properties of interest are value and allowInterruptions. The expression specified in value property will be evaluated on every single time user responds to the specific input.

In this case, the expression =coalesce(@userName, @personName) attempts to take the first non null entity value userName or personName and assigns it to user.name. The input will issue a prompt if the property user.name is null even after the value assignment unless always prompt evaluates to true.

The next property of interest is allowInterruptions. This is set to the following expression - !@userName && !@personName. This literally means what this expression reads - allow an interruption if we did not find a value for entity userName or entity personName

Notice that you can just focus on things the user can say to respond to this specific input in the Expected replies. With these capabilities, you get to provide labelled examples of the entity and use it no matter where/ how it was expressed in the user input.

If a specific user input does not work, simply try adding that utterance to the Expected response.

Out of order entity extraction

Consider this example

user: add an item to the list
bot: sure, what is the title? 
user: buy milk
bot: ok. pick the list type - todo | shopping
user: shopping list
bot: ok. i've added that.

The user could have pre-emptively answered multiple questions in the same resposne. Here is an example

user: add an item to the list
bot: sure, what is the title?
user: add buy milk to the shopping list
bot: ok. I've added that.

You can see how this is all wired up in the TestManagedLU under Add dialog, BeginDialog trigger, AskForTitle and AskForListType inputs.

By including the value property on each of these inputs, we can pick up any entities recognized by the recognizer even if it was specified out of order.

Interruption

Interruptions can be handled at two levels - locally within a dialog as well as re-routing as a global interruption. By default adapive dialog does this for any inputs -

1. On every user response to an input action's prompt,
2. Run the recognizer configured on the parent adaptive dialog that holds the input action
3. Evaluate the allowInterruption expression.
   a. If it evaluates to true, evaluate the triggers that are tied to the parent adaptive dialog that holds the input action. If any triggers match, execute the actions associated with that trigger and then issue a re-prompt when the input action resumes.
   b. If it evaluates to false, evaluate the value property and assign it as a value to the property. If null run the internal entity recognizer for that input action (e.g. number recognizer for number input etc) to resolve a value for that input action.

Handling interruptions locally

With this, you can add contextual responses to inputs via OnIntent triggers within a dialog. Consider this example:

user: hi
bot: hello, what is your name?
user: why do you need my name?
bot: I need your name to address you correctly. 
bot: what is your name?
user: I will not give you my name
bot: Ok. You can say "My name is <your name>" to re-introduce yourself to me. 
bot: I have your name as "Human"
bot: what is your age? 

See TestManagedLU under UserProfile dialog, Why, NoValue or Cancel triggers.

Handling interruptions globally

Adaptive dialogs have a consultation mechanism which propagates a user message up the parent dialogs until a dialog has a trigger that fires. If no dialog's triggers fired upon consultation then the active input action gets the user utterance back for its own processing. Consider this example -

user: hi
bot: hello, what is your name? 
user: what can you do?
bot: I'm a demo bot. I can manage todo or shopping lists. 
bot: what is your name? 

Notice that the bot understood interruption and presented the help response. See TestManagedLU UserProfile as well as Help dialogs.

[github-actions]

Coverage remained the same at 40.962% when pulling 98e7e79 on vishwac/todoWithLuisUpdates into 7e53610 on master.

[xieofxie]

how about add score requirements to intents of main dialog? Otherwise todo like 'play basketball' will be routed to Delete

[vishwacsena]

Yes. The original sample had this and I missed this in the rewrite. I'll update the PR.

stale