Using Product Data to Drive Chatbot Dialogs - GS1 2019

GS1 Connect 2019 | June 19-21
Conversational Commerce:
Using Product Information
to Drive ChatBot Dialogs
Seth Earley, Founder and CEO
Dave Skrobela, Client Partner
#convcomm

Copyright © 2019 Earley Information Science, Inc. All Rights Reserved. 2
1994
YEAR FOUNDED.
Boston
HEADQUARTERED.
50+
SPECIALISTS & GROWING.
Earley Information Science accelerates information innovation.
Our proven methodologies are designed specifically to address product and customer data;
content assets and corporate knowledge bases.
We help leading brands organize, operationalize, and optimize their information
to drive measurable business results.
We make information more
useable, findable, and valuable.

© 2019 GS1 US All Rights Reserved
Key Takeaways
• Chatbots are a channel
• Good news – use your existing product data
• Bad news – you will need more attributes
• Early stage of the marketplace = great opportunity to
lead with a functional bot
• Knowledge engineering will have ROI throughout the
organization
• Need to experiment and build capability long term
3

Use Cases
• Product selector bot
• Configuration bot
• Pricing
• Availability
• FAQ
• Troubleshooting
4

What Is Conversational Commerce?
5
Bot-based
interaction
Product
Content
Shopping
Conversational commerce is
about delivering convenience,
personalization, and decision
support while people are on the
go, with only partial attention to
spare.
- Chris Messina, Uber

Benefits of a conversational user interface…
• “Reduction of friction”
• “Extreme personalization”
• “Hyper contextualization”
• “Synchronized stream of interactions”
• “Natural interactions”
• “Zero barrier to adoption”
• “Starting conversations rather than downloading
apps”
source: Chris Messina
6

… however achieving benefits is not trivial
Benefit* What it means Requirement Implication
Reduction of friction Messaging is part of daily
user experience
Brand needs to engage with the
user
Need to embrace micro interactions and
retain identity
Extreme
personalization
Understanding of
preferences, demographics,
interaction history
Need to address privacy concerns
and align preferences with a
meaningful experience
Challenge of personalization across any
channel is still there – what is a
personalized experience?
Hyper
contextualization
Mobile context provides
multiple signals about user
intent
What do user task signals tell us
about moment to moment needs?
Though mobile context is gained, context
of web vehicle is lost
Synchronized stream of
interactions
Messaging apps synch across
devices
Need to respect preferences Understanding of device context needed
Natural interactions Language interaction
provides additional signals
for intent
Requires domain specific tuning
around products, services and
offerings
Taxonomies, ontologies, curated content is
even more important
Zero barrier to
adoption
Native use in common
applications
Nuances of chat experience
require thoughtful design
Due to varying context, rich experience
requires new UX approaches
Starting conversations
rather than
downloading apps
Inviting bot to discussion
begins process
How do you engage the user to
initiate outside of branded
context?
Gaining and retaining attention requires
creative engagement, high value
experience
*source: Chris Messina
7

Digital Assistants:
Why You Should Care
75% of homes will have one smart speaker by 2020 – Gartner
Over half of consumers expect their digital assistants to help make
retail purchases within the next 5 years.
8
Microsoft April 2019 Voice Report
https://chatbotsmagazine.com/the-complete-guide-to-conversational-commerce-e47059293efa
https://about.ads.microsoft.com/en-us/insights/2019-voice-report
“How are people using digital assistants?”
Searching for a product or service 52%

EIS KNOWLEDGE ENGINEERING
Product Bot Maturity (relative)
Configuration Bot
KE for content, bot services and context switching, with an ontology
management hub. Scalable. Reusable. Portable.
2
CPQ,
Transactions
Simple, Discrete
Tasks
MONETIZATION
&
ENGAGMENT
Expert Assistant Bot
Situational advisory and conversational commerce. Leverage
knowledge sources and eCatalogs across channels with a hybrid UX.
Expertise &
Commerce
3
Personal Concierge Bot
Tap into customer / user profiles, transactions and behavioral context to
personalize recommendations and carry on persistent conversations..
Personalized
Interactions
4
Intelligent Bot
Continuous feedback and machine learning on what worked (conversions), what
didn’t (dialog gaps), and emerging trend alerts. Integration of IOT signals / skills.
Continuous
Learning & IOT
5
Simple Product Retrieval Bot (Predefined Dialogs)
The typical “hard-coded” approach to chat bots and VAs – no knowledge
base or IA to leverage.
1
9

Inaccurate Data + Wrong Answers = Confusion
Why Chatbots Fail

Why Chatbots Fail
11

Product Data for Conversational Commerce
There’s No AI without IA

Chat and Voice = Text
• Voice interactions are converted to text
• Chat interactions are text
• Text variations are resolved to an intent
13
Answers the question: “What does the user want?”

How is voice the same as text interaction?
14
A user’s speech (“utterance”) is translated to text via speech recognition
Text is submitted as a search or as a trigger for a chatbot interaction
Hi, I’m
looking for
green
peppers…
Search
Hi, I’m looking for
green peppers…
I can help you
with that…
Green peppers
Chat
 Speech to text conversion
 Text submitted as query
 Need to anticipate user
intent from signal (text
search)
 Speech to text conversion
 Text submitted as trigger
 Need to derive user intent
from signal (text or data
retrieval)

Product Selection Bot Requirements
15
• Conversational product selection is the foundation of conversational
commerce for product manufacturers, retailers and distributors.
• Conversational product selection is similar to site search, but requires
additional elements:
• conversational interface (e.g., Amazon Lex)
• scripting rules
• enhanced product metadata for categories, attributes, synonyms, etc
• Enhanced product data must help to narrow and refine search results,
disambiguate terms, and adjust for variations in natural language.

Product Selection Bot (Example 1)
16
Example 1
General search term, where search results
span more than one product category, but
less than six.
Script: “Do any of these categories
describe the product you are looking for?
[List all by descending frequency:
<category label: short plural proper case>]”

17
Example 2
General search term, where term
matches two redirects.
Script: “Are you looking for <redirect
entity 1: short plural proper case> or
<redirect entity 2: short plural proper
case>?”

18
Example 3:
Search term matches thesaurus entry, which
expands the search.
Script: “We have ## <thesaurus preferred term
singular lower case> items.”

19
Example 4:
Narrow results from more than 5 product
matches in a single category
Script: “We have ## <thesaurus preferred
term singular lower case> items. In what
<Stock Keeping Condition> would you like
that < thesaurus preferred term singular >?”
“[List in descending frequency <attribute
value>]?”

20
Example 5:
Search term matches attribute value for more
than 5 items within a single category. User is
directed to a product listing page.
Script: “We have ## <category lower case short
singular> items of the <matching attribute
value>. Would you like to see them? {Link}”

© 2019 GS1 US All Rights Reserved 21
Incorrect:
Correct:
Example 6:
Open ended prompt for clarification
within a category.
Script: “What kind of <category label
singular> are you looking for?”

Category Metadata Example
22
Long Label Proper: HVAC Motors & Actuators
Short Label Proper: Motors & Actuators
Long Plural Clause Proper: HVAC Motors and Actuators
Short Plural Clause Proper: Motors and Actuators
Long Singular Clause Proper: HVAC Motor or Actuator
Short Singular Clause Proper: Motor or Actuator
Long Product Adjective Plural Proper: HVAC Motor and Actuator Products
Short Product Adjective Plural Proper: Motor and Actuator Products
Long Product Adjective Singular Proper: HVAC Motor or Actuator Product
Short Product Adjective Singular Proper: Motor or Actuator Product
Plus lower case variations…

Information Retrieval Continuum
23
BASIC
SEARCH ENGINE
KNOWLEDGE
PORTAL
VIRTUAL
AGENT
INTELLIGENT
ASSISTANT
KNOWLEDGE
BASE
Any text
Multiple sources
Keyword or full text
query
None necessary, but
Improves with metadata
Search box,
documents list
Search
Multiple sources, separate
taxonomies and schemas
Full text query or
Faceted exploration
Taxonomies, clustering,
classification
Role-Based
Search, classification,
databases
Domain specific ontologies
Highly curated sources
Query, explore facets
Offers related info
Conversational
NLP, search, classification
Process engines
Dynamic info enrichment
improves with interaction
Implicit query / Recommends
based on users’ history
Conversational, retains
context, personalized
NLP, search, classification
Machine Learning
Ontologies, clustering,
classification, NLP
Ontologies, clustering,
classification, NLP,
personalization
SEARCH
INTERACTION
INFORMATION
ARCHITECTURE
USER
EXPERIENCE
ENABLING
TECHNOLOGY
Increasing functionality

Virtual Assistants Require Domain Modeling
and Knowledge Base Development
24
“But even those personalities
required proficiency in other
facets of the technology such as
an expertly developed domain
model”
“Because intelligent virtual
assistants are focused within a
domain model, they benefit from a
clearly defined knowledge base and
are able to go much deeper and
stay within those bounds…”
“…domain models and ontologies
are important”
Source: Analyst Gigaom Research https://gigaom.com/2014/09/01/the-next-step-for-intelligent-virtual-assistants-its-time-to-consolidate/

The Role of Ontology
25
Bot dialog snippets,
product relationships
and additional
product metadata
are all managed as
reusable elements in
an ontology

A Sample “Beef Bot” Interaction on Mobile
26
Did you know there are almost a hundred
different cuts of beef you could try? I'm the Beef
Bot, and I can help you choose the best cuts and
suggest recipes, courtesy of
BeefItsWhatsForDinner.com.
Now, we can either check out the Beef Cuts
or explore some Recipe Suggestions.
Which would you prefer?
OK, will you be cooking on a stovetop, in
the oven or outdoor grilling?
Grilling
Let's focus on your preferred Grilling method.
Would you like to try grilling on a
barbeque, indirect grilling or
rotisserie grilling?
BBQ please
Lets see the cuts
I can show you Lean Beef Cuts or cuts
with more generous Marbling -
which will it be?
Something lean
Here are your Beef Cuts. You can click through
one to get Recipe Suggestions.

A Sample “Beef Bot” Interaction on Mobile
27
Are we done for now, or would you like to try
out a different Cooking Method?
Thanks I’m done

Building on Existing Investments
in Product Data
28
Extend product taxonomies with additional semantics, digital assets
and metadata as ontology facets for conversational commerce

Example Ontology for BEEF
Extend your product ontologies with
additional semantics, digital assets
and metadata as ontology facets for
conversational commerce
29
Beef Cuts Internal content Repository of beef
cuts extracted and persisted from beef.org
Preferred Cooking Method Facets
Dialog snippets to support dialog structure
Dialogs  All dialogs with native parents
Content Access  API Access to external CMS

User invokes BEEF BOT Did you know…?
Ok. Go ahead.
OK, will you be cooking on a stovetop, in the oven or outdoor grilling?
1
2
outdoor grilling
3
1. Grilling on BBQ
2. Indirect Grilling
3. Rotisserie Grilling
Rotisserie grilling
Now, we can either check out some beef cuts, or explore some
recipe suggestions. Which would you prefer?
4
Beef cuts
I can show you lean beef cuts or cuts with more generous marbling
– which will it be?
5
something lean
Here are your beef cuts. Please click thru one to get recipe
suggestions
EIS BEEF Bot Dialog Flow
30

Architecture for Conversational Commerce
CONTEXTUALIZED USER EXPERIENCE
Context Aware Information Architecture
Content Model Taxonomy Metadata
Structured
(Operational) Data
Unstructured
(Big) Data
Information Infrastructure
Marketing
Data
User
Data
Product
Data
Historical
Data
Operating
Content
Information Management Platforms
PIM DAM CMS ECM CRM ERP
Customer
Personalization
Content
Publishing
Site
Merchandizing
Product Info.
Management
Digital Commerce
Business
Intelligence
Knowledge
Management
Enterprise
Search
Content
Management
Digital
Workplace
31

Roadmap to a Digital Foundation
Implementation

Bot workstreams
• Content type and
variable definitions
• Classification schema
design
• Feature engineering
• Vocabulary
development
• Associative
relationship mapping
• Deconstruction and
componentization of
FAQ’s,
troubleshooting
guides, reference
materials, e-learning
modules, etc.
• Content refactoring
and component
tagging
• Integration of
component models
with user experience
• Crowdsourcing of
phrase variations for
intent triggers
(utterances)
• Classification of
intent using
customer issue and
query data
• Entity extraction
training and tuning
• Escalation and handoff
model
• Feedback workflow
design: utterance,
intent and knowledge
• Success metric design
• Governance and
accountability model
• Speech to text conversion
• Text mining/ analytics on
call logs / support content
• Search analytics
• User journey mapping
• Scenarios and use cases
• Identification of repeatable,
unambiguous processes
• Deconstruction of
user journeys into
dialogue components
• Precoordinated intent
design
• Disambiguation
models
• Intent entity
extraction
• dialogue context
tagging model
PROCESS
ANALYSIS
DIALOGUE DESIGN &
INTENT
CLASSIFICATION
CONTENT ANALYSIS,
DOMAIN MODELING
& ONTOLOGY
DESIGN
COMPONENTIZATION
OF KNOWLEDGE
CONTENT
TRAINING DATA
CORPUS
DEVELOPMENT
HYBRID LEARNING &
CONTINUOUS
IMPROVEMENT
MODEL CREATION
33

Standardized/Normalized Content is
Portable and Reusable
Standardized domain
specific schemas for
reuse
Field 1
Field 2
Field n
…
Field 1
Field 2
Field 3
Field n
…
ELearning, FAQ’s,
Troubleshooting charts,
support articles
Componentized
content
Tagging for ingestion
Componentized content
can be repurposed across
tools and technologies
Improved CSR
Information Access
Faster time to value for all
information access scenarios
Portability across AI and
Chatbot systems
Improved customer
self service
Metrics aligned with
specific content
performance
COMPONENTIZATION
OF KNOWLEDGE
CONTENT
34

Key To Success:
Program Not Project
• Begin with single category and simple use
case
• Enhance metadata for conversational retrieval
• Focus on simple, unambiguous scenarios
• Use voice to navigate catalog
• Example: find information about products,
categories or brands: price, availability, FAQs
35

Complex Advisory/ Diagnosis
Product Support
Product Configuration
Judgment Based
Domain
Complexity
Transaction Support Knowledge Retrieval
Information/
status inquiries/
order processing
Task/dialogue Complexity
Task Complexity vs. Domain Complexity
“Helper bots”
“Configuration bots”
“Transaction bots”
Don’t start here
High domain complexity/
High task complexity
36

Good news: you will use your product attributes*
Bad news: you will need more product attributes
Refactoring Product Attributes
*Assuming they are harmonized, normalized and optimized

How good is your current product data?
A. It’s in excellent shape
B. It’s not great but in decent shape
C. It’s a mess
38

System is comprised of 5 major parts:
How to build a conversation bot?
1. Product data repository optimized for
conversation
2. Chat platform capable of classifying intents using
phrase variations for training data
3. Dialog management interface
4. Mechanism for handing off to human
5. Performance metrics for ongoing improvement
39

Conclusion

The Future of Conversational Agents
• Conversational agents will mature and evolve
• Begin to prepare your product data for voice and chat
access
• Experiment with the various chat frameworks
• Build PoC’s and minimum viable products to build internal
capabilities
• There is no magic – chatbots are a channel to data, content
and knowledge
• Training data and curated content are the secret sauce
41

Vendor Hype and Market Confusion
• AI vendors are overselling capabilities
• Don’t buy what you don’t understand
• You will always need curated, high quality data
no matter how good the algorithm
• In fact, the data is more important than the
algorithm
42

Appendix

Additional Resources
• Allstate’s ABIe project case study
http://www.earley.com/knowledge/case-
studies/allstate%E2%80%99s-intelligent-
agent-reduces-call-center-traffic-and-
provides-help
• Earley Executive Roundtable Understanding
virtual agents – what's needed to make them
a reality? http://info.earley.com/roundtable-
intelligent-virtual-agents-reality
• Vendor Landscape: Knowledge Management
For Customer Engagement
https://www.forrester.com/report/Vendor+La
ndscape+Knowledge+Management+For+Cust
omer+Engagement/-/E-RES119672
• Making Intelligent Virtual Assistants a Reality
http://info.earley.com/make-intelligent-
virtual-assistant-reality-whitepaper
• Cognitive Search – The Next Generation of
Information Access
http://www.earley.com/blog/cognitive-
search-next-generation-information-access
• Earley Executive Roundtable - Training the
Robots: Evolving Virtual Assistants and the
Human Machine Partnership
http://info.earley.com/roundtable-virtual-
assistant-human-machine-partnership
• Follow the twitter hashtag #convcomm
44

How is voice different from chat interaction?
45
Voice search: fragments of a question
Chat: More natural language of interaction
Hi, I need a widget…
I can help you with that…
What type would you like?
category, category,
category…
Widget…
Chat
• Ambiguous query
• Fragment of
communication
• Expectation of list of
results (or facets to
further refine)
• Conversational query
• (More) complete
question or query
• Expectation of an
answer
category
category
category
Search
Hi, I need
to purchase
a widget …

How is voice interaction different from
text interaction?
46
No visual clues
Need to rely on working memory
“Hi, I need
to purchase
a widget
…”
“I can help you with that.
What kind of widget would
you like?
I can get you category,
category, category,
category, category,
category, category…”
“Umm…
What was
the middle
thing?”

Clearly this won’t work…
47
“Hi, I need
to purchase
a widget …”
“I can help you with that.
What kind of widget would
you like?
I can get you, category,
category, category, category,
category, category , category,
category, category , category,
category, category ,
category…
“Umm…
What was
the middle
thing?”

Hybrid voice and text
48
“Hi, I need
to purchase
a widget …”
Here are your choices:
category, category ,
category, category,
category , category,
category…
“I can help you with
that.
What kind of widget
would you like?

Seth Earley
Founder and CEO
Earley Information Science
Seth@earley.com
781-820-8080
Thank You
Eli Cooley
Senior Consultant
Earley Information Science
Eli.Cooley@earley.com
312-371-8232
Come see us
at Booth #26
49

Antitrust Caution
GS1 US is committed to complying fully with antitrust laws.
We ask and expect everyone to refrain from discussing prices, margins, discounts,
suppliers, the timing of price changes, marketing or product plans, or other competitively
sensitive topics.
If anyone has concerns about the propriety of a discussion, please inform a
GS1 US® representative as soon as possible.
Please remember to make your own business decisions and that all GS1 Standards are
voluntary and not mandatory.
Please review the complete GS1 US antitrust policy at:
www.gs1us.org/gs1-us-antitrust-compliance-policy
50

Legal Disclosure
GS1 US, Inc. is providing this presentation, as is, as a service to interested parties. GS1 US MAKES NO
REPRESENTATIONS IN THIS REGARD AND DISCLAIMS ALL WARRANTIES, EITHER EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO, ANY WARRANTY OF ACCURACY OR RELIABILITY OF ANY
CONTENT, NONINFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
GS1 US shall not be liable for any consequential, special, indirect, incidental, liquidated, exemplary, or punitive
damages of any kind or nature whatsoever, or any lost income or profits, under any theory of liability, arising
out of the use of this presentation or any content herein, even if advised of the possibility of such loss or
damage or if such loss or damage could have been reasonably foreseen.
*GS1 US employees are not representatives or agents of the U.S. FDA, and the content of this
presentation has not been reviewed, approved, or authorized by the U.S. FDA.
*If applicable
51

Trademark Notices
DataBar®, EAN®, EPC®, EPCglobal®, GDSN®, GS1 Global Registry®, GTIN®,
and Global Trade Item Number® are registered trademarks of GS1 AISBL.
GS1 US® and design is a registered trademark of GS1 US, Inc. Trademarks
appearing in this presentation are owned by GS1 US, Inc. unless otherwise
noted, and may not be used without the permission of GS1 US, Inc.
The letters “U.P.C.” are used solely as an abbreviation for the “Universal
Product Code” which is a product identification system. They do not refer to
the UPC, which is a federally registered certification mark of the International
Association of Plumbing and Mechanical Officials (IAPMO) to certify
compliance with a Uniform Plumbing Code as authorized by IAPMO.
52

Using Product Data to Drive Chatbot Dialogs - GS1 2019

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Using Product Data to Drive Chatbot Dialogs - GS1 2019

Similar to Using Product Data to Drive Chatbot Dialogs - GS1 2019 (20)

More from Earley Information Science

More from Earley Information Science (20)

Recently uploaded

Recently uploaded (20)

Using Product Data to Drive Chatbot Dialogs - GS1 2019