Since we first started using computers, we have dreamt of interacting with them in a more natural, human-like way using spoken language. Advances in machine learning and artificial intelligence have spiked an interest in natural language processing, and building conversational interfaces no longer is a dream, but a reality every developer can use in their own applications. The problem of recognising spoken input has been largely solved, and the next challenge is how to build compelling user experiences that are modelled after natural human conversation. In this talk, I will give an introduction to the Google Assistant and how to use Actions on Google to build conversational interfaces for devices with and without screens. We will take a look at how API.AI makes implementing natural language understanding feasible even for non-linguists, and discuss how you can integrate agents with existing backends. In conversation, it's all about being natural, so we will also look at how to create agents that feel like real characters.

1. Extending the Google Assistant with
Actions on Google
+PeterFriese
@peterfriese
Developer Advocate, Google

2. “In the next 10 years, we’ll shift
to a world that is AI-first”
— Sundar Pichai, CEO Google

3. Computing
becomes
universally
available

4. Human
computer
interaction
becomes
more intuitive

5. Products with
AI will offer
more intelligent,
delightful
experiences

6. Mobile-first
changed our
world in less
than 10 years

7. The Google Assistant — A conversation between you and
Google that helps you get things done in your world.
Google Home/Mobile device — The surface to interact with
the Assistant.
Actions on Google — How developers can extend the
Assistant (via Assistant apps)

8. Sorry, but it looks like you missed the last session for
Android but you can try to catch it online after Google
I/O!
Sure! Here’s Google IO 17.
Ok Google, let’s talk to Google IO 17
Enter Earcon
Exit Earcon
Hey there. I have all kinds of info on I/O, from the
keynote to the sessions you can catch online. Which
of these do you want to know about?
When is the next Android talk?

10. How does it work?
Design Develop Discover

11. Assistant app
{ conversation
api request }
{ conversation
api response }
user input
app
response

12. Ok Google,
let’s talk to
Google IO 17
I want to know
about Machine
Learning
Sure, here’s
Google IO 17
Hi! Welcome
to Google IO!
The next ML
session is...
Speech to Text
NLP
Knowledge Graph
ML Ranking
User Profile
Handle
“action.intent.MAIN”
Speech to Text
Text to Speech
...
...
...

13. “
Intent Matching — Match and categorize user utterances to
an intent.
Entity Extraction — Identify key words and phrases spoken by
the user.
@

14. Ok Google,
let’s talk to
Google IO 17
I want to know
about Machine
Learning
Sure, here’s
Google IO 17
Hi! Welcome
to Google IO!
The next ML
session is...
Speech to Text
NLP
Knowledge Graph
ML Ranking
User Profile Convert
“action.intent.MAIN”
to WELCOME event
Speech to Text
Text to Speech
...
...
...
...
NLP:
Intent Matching
Entity Extraction

16. Intent
Triggered via a series of “user says” phrases or platform based events
Can collect entity values
Matched at every turn of conversation

17. Input and Output Contexts
You can require a context to be available before an Intent is enabled
Intent can set context to enable other Intents
Retain data across intents

18. Entities (slot-filling)
Values that we are trying to capture from the user phrases
Can specify a parameter name and a type of value
Values can be optional
Values can be a list of fixed values

19. System Entities
● Time
● Date
● Time Period
● Number
● Cardinal
● Ordinal
● Temperature
● Speed
● Volume
● Weight
● Age
● Currency
● Country
● Location
● Language
● Artist
● Music
● ….

20. Developer Entities

21. Specify follow up questions if a user doesn’t specify certain values
Read out in random order to make it more natural
Prompts

22. Fulfillment

24. Node.js Client library
https://github.com/actions-on-google/actions-on-google-nodejs
const App = require('actions-on-google').DialogflowApp;
exports.yourApp = (request, response) => {
const app = new App({request, response});
console.log('Request headers: ' + JSON.stringify(request.headers));
console.log('Request body: ' + JSON.stringify(request.body));
// Fulfill action business logic
function responseHandler (app) {
// Complete your fulfillment logic and send a response
app.ask('Hello, World!');
}
const actionMap = new Map();
actionMap.set('<Dialogflow_action_name>', responseHandler);
app.handleRequest(actionMap);
};

25. From ‘app’ to ‘experience’...

26. Create a persona.

27. Preserve and reinforce your persona by
engaging the user as a separate entity
from the Google Assistant.
Own it.
Hey! This is ___
Welcome to ___
Ready to play ____
Hi! ___ here.
Hello. I’m ___
Greetings, human.
Welcome back to ___
Hey again. ___
Let’s play ___
Here’s your ___
Brought to you by ___
Hi there, ___
Let’s get started.
Ready for your ___
___, here to…
Live from ___
This is ___
What’s up, ___
and more...

28. <speak> <!-- Must be at the start of the string -->
<say-as interpret-as="characters">SSML</say-as>
<break time="3s"/>.
<audio src="https://example.com/file.mp3"></audio>
<prosody rate="slow" pitch="-2st">Can you hear me?</prosody>
<emphasis level="moderate">This is important</emphasis>
<sub alias="World Wide Web">WWW</sub>.
<p><s>This is one.</s><s>This is two.</s></p>
</speak> <!-- Must be at the end of the string -->
Reinforce it with SSML
Speech Synthesis Markup language
“S S M L”
[3 second pause]
[audio file plays]
“Can you hear me?”
“This is important”
World Wide Web
[two sentences]

29. // This function is used to handle the welcome intent
// In Dialogflow, the Default Welcome Intent ('input.welcome' action)
// In Actions SDK, the 'actions.intent.MAIN' intent
function welcome (app) {
if (app.getLastSeen()) {
app.ask(`Hey you're back...`);
} else {
app.ask('Welcome to Number Genie!...');
}
}
Reinforce it with persistence
https://developers.google.com/actions/assistant/best-practices

30. Design for multiple surfaces.

31. Support different surface capabilities
https://developers.google.com/actions/assistant/surface-capabilities
AUDIO_OUTPUT SCREEN_OUTPUT

32. function simpleResponse (app) {
app.ask({
speech: 'Howdy! I can tell you fun facts about ' +
'almost any number, like 42. What do you have in mind?',
displayText: 'Howdy! I can tell you fun facts about ' +
'almost any number. What do you have in mind?'
});
}
Support speech and display text
https://developers.google.com/actions/assistant/responses
Chat text should be a subset of audio

33. function suggestionChips (app) {
app.ask(app.buildRichResponse()
.addSimpleResponse('Howdy! I can tell you fun facts ‘ +
‘about almost any number. What number do you have ‘ +
‘in mind?')
.addSuggestions(['0', '42', '100', 'Never mind'])
);
}
Guide the user (suggestion chips)
https://developers.google.com/actions/assistant/responses
Chip text should be under 25 characters

34. function basicCard (app) {
app.ask(app.buildRichResponse()
.addSimpleResponse('Math and prime numbers it is!')
.addBasicCard(
app.buildBasicCard(`42 is an even composite number. It ` +
`is composed of three distinct prime numbers multiplied together. It ` +
`has a total of eight divisors. 42 is an abundant number, because the ` +
`sum of its proper divisors 54 is greater than itself. To count from ` +
`1 to 42 would take you about twenty-one…`)
.setTitle('Math & prime numbers')
.addButton('Read more')
.setImage('https://example.google.com/42.png', 'Image alternate text')
)
);
}
Display basic cards
https://developers.google.com/actions/assistant/responses

35. Lists and carousels for selection
https://developers.google.com/actions/assistant/responses
Used for easy selection
<10 items
Used for comparison
<30 items

36. Know the user.

37. Updates

38. Google Home Mobile Device
NAME Registered device user’s full name Registered device user’s full name
DEVICE_COARSE_LOCATION Zip code and city N/A
DEVICE_PRECISE_LOCATION Coordinates and street address Coordinates
let permission = app.SupportedPermissions.DEVICE_COARSE_LOCATION;
app.askForPermission('To find bookstores near you', permission);
Q: "Recommend me a local bookstore"
A: "To find bookstores near you, I'll just need to get your zip code from Google. Is that okay?"
Ask for information
https://developers.google.com/actions/assistant/helpers#user_information

39. Link an account (OAuth2)
https://developers.google.com/actions/identity/
Seamless account linking with Google Sign-in

40. Transact with the user
https://developers.google.com/actions/identity/
Build orders
Use Google provided payment
instrument
Use your payment processor
(Stripe, Braintree, Vantiv, more
coming)
Update order status

42. Reach users...

43. In Dialogue Discovery:
Explicit Triggering (of Actions)
Ok Google, ask Google I/O 17 about what was announced
Trigger Phrase App Name
Developer Specified
Action
Preposition
Action Phrase
Developer Specified
Partner Examples: Let’s speak to Domino’s
At Akinator
Ask Dr. Doggy if dogs can eat chocolate
Ok Google, let’s talk to Google I/O 17
Trigger Phrase App Name
Developer Specified

44. Hey Google, I want to play a game.Hey Google, tell me a joke.
In Dialogue Discovery:
Implicit Triggering
Hey Google, I want to work out. Hey Google, what’s the surf report

45. Discovery in the Google
Assistant Directory

46. Link to your
Assistant App
From anywhere
Share through social media.
Promote through your own site or apps.
Encourage press to drive traffic
to your Assistant app.

47. What makes a great app

48. Where do I start?

49. Actions Console
console.actions.google.com

50. 0. Put away the code
1. Draft a persona
2. Write some dialog

51. DIALOG + STRUCTURE

52. g.co/dev/ActionsDesign

53. developers.google.com/actions

54. Action Templates

55. Actions on Google Developer Community
https://g.co/actionsdev
Actions on Google Developers
https://developers.google.com/actions
+PeterFriese
@peterfriese
Developer Advocate