ai, nlu, chatbots - aspect · services. what is? activate. whatis_monumentme_connect_se rvices...
TRANSCRIPT
1
©2017 Aspect Software, Inc. All rights reserved
WHAT NEXT?10 STEPS TO BUILDING YOUR BOT
TOBIAS GOEBEL | DIRECTOR, EMERGING TECHNOLOGIESDR. LISA MICHAUD | DIRECTOR ENGINEERING, NATURAL LANGUAGE
2
WHAT IS THE BOT’S ROLE? SET YOUR GOALS
1
3
©2017 Aspect Software, Inc. All rights reserved
DIFFERENT KINDS OF CUSTOMER-FACING BOTS
• Marketing• Engage the customer to increase mind share and brand recognition• Goal: enhance market standing and create MQLs
• Sales• Lead to a sale or near-sale situation• Goal: increase the topline revenue
• Service & Support• Traditionally done via IVR, Web• Goal: contact deflection to reduce cost, free up agents for more
meaningful work
4
©2017 Aspect Software, Inc. All rights reserved
INFORMATIONAL VS TRANSACTIONAL
General Information CRM Information/Transactions
Ente
rpris
e Fi
rew
all
CRMDB
Appointment Scheduling
Data Lookup
PaymentsFAQsFAQs
5
PICK A CHANNEL
2
6
©2017 Aspect Software, Inc. All rights reserved
DIFFERENT CHANNELS HAVE DIFFERENT CHARACTERISTICS
Most personal channel
Concise, 160 characters
Text only
Works on every cell phone
Rich media
Still "new" to some
Requires smartphone
Traditional home for “virtual assistants”
Needs meaningful integration into existing environment
Think hard about the gain
App fatigue won’t go away just because you have chat now
SMS Messenger & Co. Web Chat In-App Chat
7
©2017 Aspect Software, Inc. All rights reserved
PILOTS AND NEW CHANNELS
• Untapped channel: won’t conflict or compete with any existing programs in your organization
• Starting with a pilot: easier to “sell” internally
• Low-risk commercial engagement, while establishing the technology needed for full roll-out later
• Your team will “forget” that it was ”just” a pilot foundation is laid for moving forward
8
CREATE A CONVERSATIONAL ARCHITECTURE
3
9
©2017 Aspect Software, Inc. All rights reserved
GRAPHICAL UI VS. CONVERSATIONAL UI
GUI CUIInteraction-Based: User “manipulates” screen Dialog-Based: 2 interlocutors take turnsScreen-Based: new page replaces old Story-Based: dialogue flows in one direction, with
time; previous interaction remains visibleUnambiguous: what you click is what you get Error-Prone: misunderstandings, ambiguities,
corrections, confirmations, …“High Bandwidth”: visual channel allows higher information throughput
“Low Bandwidth”: gradual information discovery
10
©2017 Aspect Software, Inc. All rights reserved
THE PECULIAR PRONOUN…
• CUIs use conversational markers to efficiently refer to previously mentioned concepts
What is the Concierge Service?
A premium service with exclusive information and assistance. It includes point of interest
search, restaurant/hotel reservations, cultural event schedules, and more…
How much is it?
How much is what?
11
©2017 Aspect Software, Inc. All rights reserved
GUIS ARE STRUCTURED THROUGH SITE MAPS, OR INFORMATION ARCHITECTURES (IA)
12
©2017 Aspect Software, Inc. All rights reserved
CUIS NEED A CONVERSATIONAL ARCHITECTURE (CA)
• Define the scope: Start by collecting the top N questions
• Gather your content
• Define hierarchies within content, and contextual dependencies• For every concept introduced in an answer, expect a follow-up question…
13
©2017 Aspect Software, Inc. All rights reserved
Connect
Busi
ness
What is?
Monument Car “Driver Connect”
Wha
t is?
Wha
t is?
Serv
ices
Wha
t is?
Activ
ate
WH
ATIS
_MO
NU
MEN
TME_
CO
NN
ECT_
SER
VIC
ES
List
HO
WTO
_AC
TIVA
TE_S
ERVI
CES
LIST
_SER
VIC
ES_F
OR
MYC
AR
WH
ATIS
_MO
NU
MEN
TME_
CO
NN
ECT
WH
ATIS
_MO
NU
MEN
TME
Tech
nica
l R
equi
rem
ent
WH
ATIS
_CO
NN
ECT_
REQ
S
WH
ATIS
_MO
NU
MEN
TME_
CO
NN
ECT_
BUSI
NES
S
Adap
ter
Wha
t is?
Req
uirm
ents
WH
ATIS
_AD
APTE
R
Com
patib
le
mod
els
WH
ATIS
_AD
APTE
R_R
EQS
LIST
_CAR
S_AD
APTE
RC
OM
PATI
BLE
Serv
ices
WH
ATIS
_MO
NU
MEN
TME_
CO
NN
ECT_
BUSI
NES
S_SE
RVI
CES
Activ
ate
Car
HO
WTO
_AC
TIVA
TEC
AR_C
ON
NEC
T
Lock
ing
WH
ATIS
_REM
OTE
LOC
KIN
G
Empt
y R
emot
e D
ata
HO
WTO
_FIX
EMPT
YREM
OTE
DAT
A
Vehi
cle
Loca
tion
Not
W
orki
ngH
OW
TO_F
IXVE
HIC
LELO
C
Allo
w C
o-U
ser
HO
WTO
_ALL
OW
OTH
ERSA
CC
ESS_
CO
NN
ECT
Rem
ove
Co-
Use
rH
OW
TO_R
EMO
VEC
OU
SER
_CO
NN
ECT
= answer mentions concepts/terms that user might have follow-up questions on, and therefore needs to remain the context for resolving anaphora
14
DESIGN DIALOGUE FLOW AND STORYBOARDS
4
15
©2017 Aspect Software, Inc. All rights reserved
DETAILED FLOW AND MESSAGE DESIGN
https://www.aspect.com/globalassets/microsite/nlu-lab /images/ChatbotDesignTemplate.pdf
16
©2017 Aspect Software, Inc. All rights reserved
THE FIRST RULE OF MESSAGE DESIGN
For every message you design, anticipateevery possible way someone could respond
17
©2017 Aspect Software, Inc. All rights reserved
SOME TIPS…
o Design for how humans talk, not how you WANT them to talk
o Design with context in mind
o Don’t expect your user to know exactly what they want
o Don’t use rhetorical questions
…and savings?
Checkout
That’s not what I meant vs. go back
Battery
Interested in our newsletter? Click here:
aspct.co/Fg45Si9
Where can I get one?
Yes!
18
DESIGN THE INTEGRATIONS
5
19
©2017 Aspect Software, Inc. All rights reserved
REUSE EXISTING WORK
• Your IVR, Web, mobile presence should already be built on Web Services, exposing similar if not same functions
• FAQs: Integrate with KB, or start with custom data base/structure
20
©2017 Aspect Software, Inc. All rights reserved
20
COLLECT CHAT DATA
6
21
©2017 Aspect Software, Inc. All rights reserved
VARIATIONS ON A QUESTION
Start with your top N questions:
Q: “How do I free up space?”
A: If an update fails because you need more space, you can
manually remove content from your device by clicking on ‘Settings.’
22
©2017 Aspect Software, Inc. All rights reserved
VARIATIONS ON A QUESTION"I don't have enough space to update my iphone"
"It says I have not enough space on my device"
"I don't have enough room for the update"
"I need more space"
"How do I free up space?"
"My memory is full"
"My disk is full"
"I ran out of space"
"The space is insufficient"
23
©2017 Aspect Software, Inc. All rights reserved
VARIATIONS ON A QUESTION
• Age
• Region
• Socioeconomic group
• Language background
• Investment in the task
24
©2017 Aspect Software, Inc. All rights reserved
CROWD-SOURCING THE COLLECTION
Internal “Users” Amazon Mechanical Turk
25
©2017 Aspect Software, Inc. All rights reserved
CROWD-SOURCING THE COLLECTION
ANSWERS:
• How do I remove files?
• How can I delete files?
• How to remove files
• Removing files
PROVIDE VARIATIONS TO:
How do I delete files?
26
©2017 Aspect Software, Inc. All rights reserved
CROWD-SOURCING THE COLLECTIONMEMORY
MANAGEMENT
$400
If an update fails because you need
more space, you can manually remove
content from your device by clicking on
‘Settings.’
“I need more space”
“How do I free up space?”
“My memory is full”
“my update failed because of space”
“How do I remove content from the phone”
“Where do I go to clear out my device?”
“JEOPARDY” APPROACH
27
©2017 Aspect Software, Inc. All rights reserved
27
PICK A PLATFORM AND DEVELOPMENT APPROACH
7
28
©2017 Aspect Software, Inc. All rights reserved
TASKS FOR THE BOT
Task #1: determine an INTENT
I want to book a stay.
Can I check my reservation?
Task #2: extract DATA
Check-in on Friday February 5
I want a king-sized bed
29
©2017 Aspect Software, Inc. All rights reserved
VARIATIONS ON A QUESTION
transferdata
transfer data
insert object
30
©2017 Aspect Software, Inc. All rights reserved
A TOP-DOWN VS. BOTTOM-UP APPROACH
RULES
DATA
A machine-learning approach requires a large amount of classified data from which to derive a classifier that can classify new data the same way. It relies on that data capturing the diversity one expects to see in future data.
Botto
m-u
p
31
©2017 Aspect Software, Inc. All rights reserved
A TOP-DOWN VS. BOTTOM-UP APPROACH
RULES
DATA
A linguistic approach starts with intuitions about how we should classify the input and leverages a full linguistic analysis to empower the classifications. It can then be augmented and tuned when more data exist.
Top-down
32
©2017 Aspect Software, Inc. All rights reserved
Botto
m-u
pTop-dow
n• A Big Data approach –
requires large amounts of classified data.
• Relies on that data capturing the diversity one expects to see in future data.
• Result: a “black box” classifier.
• When an intent is misclassified, you need to add training data to convincethe classifier to do otherwise.
• A Sparse Data approach –does not require as much data to get started.
• Leverages linguistic abstractions to capture expected diversity.
• Result: a fully inspectable classifier.
• When an intent is misclassified, you can directly modify the logic to tell it to do otherwise.
33
©2017 Aspect Software, Inc. All rights reserved
SEMANTIC APPROACHES TO INTENT CLASSIFICATION
I want to transfer my data
How do I move my files?
I want to transfer that to my file
34
©2017 Aspect Software, Inc. All rights reserved
34
IMPLEMENT THE DIALOGUE
8
35
©2017 Aspect Software, Inc. All rights reserved
"I don't have enough space to update my iphone”,DELETE"It says I have not enough space on my device”, DELETE"I don't have enough room for the update”, DELETE"I need more space”, DELETE"How do I free up space”, DELETE"My memory is full”, DELETE"My disk is full”, DELETE"I ran out of space”, DELETE"The space is insufficient”, DELETE
IMPLEMENTING THE CHATBOT: MACHINE LEARNING
36
©2017 Aspect Software, Inc. All rights reserved
IMPLEMENTING THE CHATBOT: CLASSIFICATION RULES"I don't have enough space to update my iphone"
"It says I have not enough space on my device"
"I don't have enough room for the update"
"I need more space"
"How do I free up space"
"My memory is full"
"My disk is full"
"I ran out of space"
"The space is insufficient"
37
©2017 Aspect Software, Inc. All rights reserved
IMPLEMENTING THE CHATBOT: CLASSIFICATION RULES"I don't have enough space to update my iphone"
"It says I have not enough space on my device"
"I don't have enough room for the update"
"I need more space"
"How do I free up space"
"My memory is full"
"My disk is full"
"I ran out of space"
"The space is insufficient"
38
©2017 Aspect Software, Inc. All rights reserved
38
INTERNAL TESTING AND REVISION OF YOUR USE CASE DETECTION
9
39
©2017 Aspect Software, Inc. All rights reserved
INTERNAL TESTING AND REVISION
Collect Example Sentences
EngineerNLU
Test NLU
Identify Sources of Error
Internal Testers
QA Service
Actual Users
40
©2017 Aspect Software, Inc. All rights reserved
40
EARLY DEPLOYMENT AND REVISIONS
10
41
©2017 Aspect Software, Inc. All rights reserved
EARLY DEPLOYMENT AND REVISIONS
Collect Example Sentences
EngineerNLU
Test NLU
Identify Sources of Error
What percentage of incoming requests are in-scope?Do you need to add more use cases?
How many use cases can be handled without human assistance?Does the intent classification need additional work?Is the response text being understood clearly?
What are the most common use cases?What language patterns are you observing?How can these observations help future projects?
42
©2017 Aspect Software, Inc. All rights reserved
THANK [email protected] / @[email protected] / @LISANMICHAUD
NLU Lab: www.aspect.ai