Sara Nash and Urmi Majumder, Principal Consultants at Enterprise Knowledge, presented on April 19, 2023 at KM World in Washington D.C. on the topic of Scaling Knowledge Graph Architectures with AI.
In this presentation, Sara and Urmi defined a Knowledge Graph architecture and reviewed how AI can support the creation and growth of Knowledge Graphs. Drawing from their experience in designing enterprise Knowledge Graphs based on knowledge embedded in unstructured content, Sara and Urmi defined approaches for entity and relationship extraction depending on Enterprise AI maturity and highlighted other key considerations to incorporate AI capabilities into the development of a Knowledge Graph.
View presentation below in order to learn about how:
Assess entity and relationship extraction readiness according to EK’s Extraction Maturity Spectrum and Relationship Extraction Maturity Spectrum.
Utilize knowledge extraction from content to gather important insights into organizational data.
Extract knowledge with three approaches:
RedEx Rule, Auto-Classification Rule, Custom ML Model
Examine key factors such as how to leverage SMEs, iterate AI processes, define use cases, and invest in establishing robust AI models.
3. ⬢ Serves as implementation lead for knowledge graphs - ranging
from early design and prototyping to enterprise solutions
⬢ Expert in Knowledge Graph and Semantic Technologies
⬢ Established standards in design and delivery of semantic
recommender solutions
SARA
PRINCIPAL CONSULTANT, ENTERPRISE KNOWLEDGE
NASH
URMI
PRINCIPAL CONSULTANT, ENTERPRISE KNOWLEDGE
MAJUMDER
⬢ Expert in system architecture, design, and implementation
of semantic enterprise solutions
⬢ Leads the development of technical solutions in support of
a wide variety of both federal and commercial clients
ENTERPRISE KNOWLEDGE
4. ENTERPRISE KNOWLEDGE
A knowledge graph is a specialized
graph of the things we want to
describe and how they are related.
● Standardize data entities and
enrich data with context.
● Can be expanded by leveraging
various approaches, including
AI-driven entity and
relationship extraction.
Most of the data in our world is
unstructured.
Unstructured data:
1) Has no metadata
2) Can’t be captured neatly in
structured formats like XML, JSON,
or relational databases; and,
3) Lacks standardization, which
prevents establishing uniform
processes for analysis.
The Challenge The Knowledge Graph
5. Creating Knowledge from Unstructured Content
There is a vast amount of information embedded in documents, reports, records, process flow
diagrams, and more. There is an opportunity to extract this knowledge and make
meaningful connections to accelerate knowledge discovery across teams.
Extract Knowledge from Content Organize Knowledge in Logical Structure Get Insights from Data
ABC
Material
XYZ
Product
Process 1
Process Step
E1
Experiment
F2
Material
isInput
isOutput
E0
Experiment
isInput isInput
creates
What Materials were used
to make XYZ?
● ABC was used in Process 1
What are the experiments
in which ABC was used?
● E0
● E1
Which experiment of ABC
was used to make F2?
● E1
6. ENTERPRISE KNOWLEDGE
Natural Language
Processing (NLP)
model leveraged to
build a Knowledge
Graph (KG) for
providing coherent
and relevant learning
content
recommendations.
ML model was used to
facilitate KG
generating dynamic
automated
regulatory reporting,
and expediting
research and
publication processes.
Learning Enablement Safety Standards Regulatory Reporting and More
Discovery Analysis Research
Machine Learning
(ML) model was
trained to allow KG to
facilitate thorough
analyses of possible
risks, and help
planners plan the
best safety measures
for mitigation.
● Product Marketing
● E-Commerce
● Content Cleanup
● Data Discovery in
Research
Top Graph Use Cases: https://enterprise-knowledge.com/top-graph-use-cases-and-enterprise-applications-with-real-world-examples/
Success Stories
7. Source Data and
Content
Taxonomy/
Schema Storage
Entity and
Relationship
Extraction
Enterprise
Content and Data
Dedicated
Taxonomy/Ontology
Management System
Auto-tagging and/or
Extraction of Key
Knowledge
Enriched Content
Storage
Persistent Graph
Storage
Data Orchestration
Front End
Visualization /
UI
API
AI
Search
Chatbots/
Q&A
Data Visualization
and Reporting
Recommender
Systems
Solutions Architecture for Scalable Knowledge Graphs
8. Source Data and
Content
Taxonomy/
Schema Storage
Enterprise
Content and Data
Dedicated
Taxonomy/Ontology
Management System
Enriched Content
Storage
Persistent Graph
Storage
Data Orchestration
Front End
Visualization /
UI
API
Search
Chatbots/
Q&A
Data Visualization
and Reporting
Recommender
Systems
AI accelerates extracting entities and relationships at
scale from unstructured enterprise data. This is
increasingly possible due to advances in the natural
language processing space.
Entity and
Relationship
Extraction
Auto-tagging and/or
Extraction of Key
Knowledge
AI
Solutions Architecture for Scalable Knowledge Graphs
9. Transformational
Institutional
Unaware of
how AI is being
adopted across
organizations.
Hopeful about
the promise of
AI and its
impact on
business.
Some AI/ML
models are in
use for specific
use cases.
AI solutions are
supporting
shared use
cases across the
organization.
AI is part of
business DNA,
transforming
infrastructure
and processes
to improve
efficiency while
optimizing
costs.
Operational
Experimental
Not Ready
(Pre-AI)
ENTERPRISE KNOWLEDGE
AI Maturity Spectrum for the Enterprise
10. …with
Automated
Monitoring
and Retraining
Entity Extraction Maturity Spectrum
Regular
Expression
Based
(RegEx)
Auto
Classification
Custom ML
Model …for
Entity
Extraction
…with Active
Learning
Definition
Taxonomy driven
categorization of
content
Definition
Traditional supervised
learning approach for
text classification
Considerations
Highly dependent on
training data
Definition
Model is re-trained
periodically based on
human feedback
Considerations
Increased text
classification accuracy
Definition
Model is automatically
re-trained, tested, and
deployed
Considerations
Recommended for
large orgs with
established DataOps
processes
Definition
Use patterns of
characters and
operators to match
text
Considerations
Requires explicit
definition of rules,
and may lead to false
positives
Considerations
Limited to the
terms defined in the
taxonomy
Transformational
Institutional
Operational
Experimental
Not Ready
(Pre-AI)
11. RegEx Rule: Knowledge Graph
University is a Service because
[*]University is a Service.
This content on EK’s site is rich with knowledge
that can be extracted through different
approaches.
Auto-Classification Rule:
● EKGU is a synonym for Enterprise
Knowledge Graph University and this
article is about EKGU.
● Information Analyst is a Role, so
Information Analyst may take EKGU
Custom ML Model
● SPARQL and SHACL are
frameworks
● Taxonomy, ontology, and Knowledge
Graphs are semantic models
● Graph database is a tool
Entity Extraction in Action
12. …with
Automated
Monitoring
and Retraining
Source
Schema
Based
Rule Based
Custom ML
Model …for
Relationship
Extraction
…with Active
Learning
Definition
Custom rule set,
borrowing relationships
from standard formats
Considerations
Relies on maintaining
rules and may lead to
false positives
Definition
Traditional supervised
learning for classifying
text between two
entities
Considerations
Highly dependent on
training data
Relationship Extraction Maturity Spectrum
Definition
Model is automatically
re-trained, tested, and
deployed
Considerations
Recommended for
large orgs with
established DataOps
processes
Definition
Model is re-trained
periodically based on
human feedback
Considerations
Increased text
classification accuracy
Definition
Exploit the schema
(JSON, XML, etc.) of the
source system
Considerations
Requires explicit
mapping in data source
between entities to
assign relationships
Transformational
Institutional
Operational
Experimental
Not Ready
(Pre-AI)
13. AI Maturity Spectrum for the Enterprise Revisited
Transformational
Institutional
Operational
Experimental
Not Ready
(Pre-AI)
ENTERPRISE KNOWLEDGE
Use pattern
matching for
deterministic
entity and
relationship
extraction
Design a
starter
taxonomy &
use it for
taxonomy
driven graph
instantiation
Use pre-trained
ML models for
probabilistic
entity and
relationship
extraction
Fine-tune
pretrained ML
models based
on SME
feedback to
update graph
Monitor
information
extraction
quality to
automatically
retrain ML
model
14. Be iterative Start small, then iteratively refine and expand AI
integration.
Involve your SMEs SMEs can help validate AI performance and provide
feedback, ensuring accurate and relevant
improvements.
Define consumer-
facing use cases
Use cases should address pain points or challenges
faced by your target audience.
Invest in quality Invest in establishing robust AI models and
structured content and data that align with your use
cases.
In order to best incorporate AI capabilities into your Knowledge Graph
pipeline, there are several key factors to consider:
Key Considerations
15. ENTERPRISE KNOWLEDGE
Any Questions?
Thank you for listening.
We are happy to take any
questions at this time.
Sara Nash
snash@enterprise-knowledge.com
www.linkedin.com/in/sara-g-nash/
Urmi Majumder
umajumder@enterprise-knowledge.com
www.linkedin.com/in/urmim/
How prepared is your organization
for AI? Take EK’s AI Maturity
Assessment:
https://s.enterprise-knowledge.com/ekaiassessment