Part 3 of this webinar series deconstructs the underlying technology - from the search index to latent semantic indexing/analysis.

1. Demys&fying
Technology
Assisted
Review
Part
3:
Deconstruc&ng
the
Technology
Sonya
L.
Sigler

2. Agenda
 Review/Overview
 Underlying
Search
Technology
 dtSearch
 Lucene
(open
source)
 Others
–
My
SQL,
etc.
 Underlying
StaCsCcal
Based
Technology
 Rules
Based
Technology
(LinguisCc
or
StaCsCcal)
 Bayesian
ProbabilisCc
Technologies
 Latent
SemanCc
Indexing
 Q
&
A
Demys&fying
Technology
Assisted
Review

3. Review/Overview
-­‐
Search
&
Review
Spectrum
Linear
Review
Culling
IteraCve
search
Review
Accelerated
Review
Email
Threading
Near
Duplicate
DetecCon
Automated
Review
Per
CA
-­‐
Clustering
Relevance
Ranking
Document
CategorizaCon
(Supervised)
Machine
Learning
Cost
Latent
SemanCc
Indexing
(staCsCcal
probability)
PaRern
Analysis
Sampling
Data
for
High
Precision
and
Recall
Rates
Organiza3on
Commitment
Demys&fying
Technology
Assisted
Review

4. Underlying
Technologies
Rules
Based
Systems
dtSearch
Key
word
Search
Ontologies
Lucene
Other
Search
Engines
LinguisCc
–
word
based
StaCsCcal
-­‐
#s
based
Bayesian
ClassificaCon
Support
Vector
Models
Latent
SemanCc
Indexing
Demys&fying
Technology
Assisted
Review

5. Database
NormalizaCon
From:
Nuala
Coogan
Nuala@SFLData.com
Subject:
EDI
Summit
–
Florida
Date:
October
3,
2012
10:11:21
AM
PDT
To:
Sigler
L.
Sonya
Sonya@sigler.name
From:
Nuala
Coogan
Subject:
EDI
Summit
–
Florida
Date:
10/03/12
To:
Sonya
Sigler
Demys&fying
Technology
Assisted
Review

6. TokenizaCon
 Words,
Phrases,
Symbols
 Mostly
at
the
word
level
 Numbers
 PunctuaCon
 Meaningful
Elements
or
Pieces
–>
Tokens
 Parsing
and
Text
Mining
 Treatment
of
ContracCons,
Hyphenated
words,
EmoCcons
and
Larger
Constructs
(like
urls)
 Look-­‐up
tables
Demys&fying
Technology
Assisted
Review

7. LinguisCc
Based
Technologies
Keyword
Sample
Ontology
Sample
Simple:
q
((+(std:%CapacityReports_%
std:%DINCapacity_
%)
(std:%ACMEEPPlant_%
std:%ProductName_%))
"legal
systems"
OR
legalsystems
(+(std:%ACMEPNPlant_%
std:%ProductName_%)
+
"Mike
Custodian”
(std:%ProducCveCapability_%
std:
%CapacityReports_%))
(+(std:%CapacityCreep_%
std:%OperaConsImprovement_%
std:
Medium:
%CapacityExpansion_%
std:%CapacityRestoraCon_
mail(custodian@domain.com)
AND
"legal
systems”
%)
+(std:%ACMEPNPlant_%
std:%ProductName_
%))
(+(std:%EquipmentReplacement_%
std:
(Custodian
w/3
(Mike
OR
Michael
OR
M))
%FinishingColumn_%)
+(std:%ACMEPNPlant_%
std:
%ProductName_%))
(std:%Audit_%
actor:%Audit_
Complex:
%)
(+(std:%SeRlementNegoCaCons_%
std:
%ContractNegoCaCons_%
)
+(actor:
(privilege
OR
privileged
OR
legally
OR
"work
%ACMEOutsideCounsel_%
std:
product")
NOT
w/35
(((original
OR
intended
OR
%ACMEOutsideCounsel_%
actor:%ACME
designated
OR
named)
w/3
(recipient
OR
UBOutsideCounsel_%
std:
recipients
OR
addressee
OR
addressees
OR
%AcmeSubOutsideCounsel_%
actor:%AcmeSub_%
solely))
OR
("message
in
error")
OR
("received
in
std:%AcmeSub_%))
(std:%FTC_%
actor:%FTC_%)
((+subject:%ProductName_%
+(std:swap
error")
OR
("named
above")
OR
((electronic
or
std:"supply
agreement"
std:"exchange
agreement"
email
or
e-­‐mail)
w/3
(message
or
transmission))
std:"agree
to
exchange"))
std:"name
OR
("confidenCality
noCce"))
Demys&fying
Technology
Assisted
Review

8. Search
Engines
 dtSearch
 dtSearch
Corp.,
founded
1991
 Incorporated
into
Symantec’s
Norton
Navigator
 SDKs
available,
most
license
off
the
shelf
 hRp://support.dtsearch.com/faq/search.html
 Lucene
 Open
source
-­‐
hRp://lucene.apache.org/core/
 Doug
Cukng,
1999,
Part
of
Apache
projects
in
2001
 APIs,
Customizable
 Other
–
My
SQL,
SQL,
(DBMS,
RDBMS)
Demys&fying
Technology
Assisted
Review

9. dtSearch
 RelaCvity,
Concordance,
Viewpoint,
others
 Single
User
desktop
license
$199
 LiRle
CustomizaCon
–
more
similariCes
across
apps
 Includes
Boolean
operators
 Includes
Proximity
searching
 Includes
Fuzzy
Searching
 Alphabet
-­‐>
Alphaqet,
alpphabet,
alpkaqet
Demys&fying
Technology
Assisted
Review

10. Lucene
 Clearwell,
Intella,
Cataphora,
SHIFT,
others
 Open
Source
Tool
–
meant
to
be
customized
 LiRle
SimilariCes
Across
Apps
 Know
your
defaults!
 Includes
Boolean
Operators
 Includes
Proximity
Searching
Demys&fying
Technology
Assisted
Review

11. dtSearch
–
Fuzzy
Searching
 Degrees
of
Fuzziness
 1-­‐10;
dtSearch
uses
1-­‐3
 Marked
by
use
of
%
symbol
 InserCon:
co%t
→
coat
 DeleCon:
coat
→
co%t
 SubsCtuCon:
coat
→
cost
 TransposiCon
cots
→
cost
 Fuzziness
Degrees
 Alphabet
–
Alphaqet,
Alpkaqet
Demys&fying
Technology
Assisted
Review

12. Boolean
Operators
–
AND,
OR
,
NOT
dtSearch
Lucene
 Search
for
 Depends
on
 MulCple
words
customizaCon
treated
as
a
phrase
 OR
 ANY
–
treats
word
 AND
list
as
separated
by
OR
 Know
your
defaults
 ALL
–
treats
word
list
 Spell
out
variaCons
as
separated
by
AND
Demys&fying
Technology
Assisted
Review

13. Proximity
dtSearch
Lucene
 Pre/post
 w/
order
doesn’t
 w/
order
doesn’t
maRer
maRer
 House
white
 No
pre
usage
 White
house
 Pre/
finds
first
word
prior
to
second
word
 White
house
Demys&fying
Technology
Assisted
Review

14. Punctua&on
dtSearch
Lucene
 LeRers
 All
punctuaCon
 Space
treated
as
a
word
 Ignored
break
 Hyphens
 %
-­‐
fuzzy
searching
 _
-­‐
ignored
Demys&fying
Technology
Assisted
Review

15. dtSearch
Hyphen
Example
Demys&fying
Technology
Assisted
Review

16. Noise
Words
–
Unindexed,
Ignored
dtSearch
Lucene
 Unindexed,
Can
 Ignores
*
in
quotes
create
Custom
Index
 (Quality
Control*)
=
 Many,
but
a
few
Quality
Control
but
examples:
Do,
not,
nothing
else
for,
your,
only,
under,
made,
way
 Know
defualts
Demys&fying
Technology
Assisted
Review

17. Stemming
v.
Wild
Cards
Stemming
Wild
Cards
 SyntacCc
VariaCons
 Strings
of
characters
 Replacements
for
beginning,
 Regular
Verbs
parts,
or
endings
 Irregular
Verbs
 Lucene
-­‐
*
 dtSearch
performs
 dtSearch
-­‐
?
For
single
character,
*
for
any
#
of
poorly
with
irregular
characters
verbs
 Time
consuming
 Spelling
out
recommended
 Wild
cards
in
quotes
Demys&fying
Technology
Assisted
Review

18. Stemming
v.
Wild
Cards
Example
Stemming
Wild
Cards
Catch
–
Lucene
Catch*
Catch~
-­‐
dtSearch
 Catch
 Catch
 Catches
 Catches
 Catching
 Catching
 Catcher
 Catcher
 Catch1234
–
not
in
 Caught
-­‐
not
in
dtSearch
stemming
Demys&fying
Technology
Assisted
Review

19. StaCsCcal
Technologies
 Rules
Based
 Bayesian
ClassificaCon
 Vector
Space
Modeling
 Latent
SemanCc
Indexing
Demys&fying
Technology
Assisted
Review

20. StaCsCcal
Based
Technologies
 Concept
-­‐
Clustering
 Machine
 Unsupervised
 Quickly
understand
data
 Uncontrolled
Clusters
Demys&fying
Technology
Assisted
Review

21. StaCsCcal
Based
Technologies
 Concept
-­‐
Categoriza&on
 User
Created
 Supervised
 Control
Topics
 Time
Consuming
Demys&fying
Technology
Assisted
Review

22. StaCsCcal
Based
Technologies
Rules
Based
Systems
 If..
Then…
 If
email
=
person
1
to
person
2
then
return
it
 If
email
=
person
1
or
person
2
then
return
it
 ArCficial
Intelligence
Systems
 EnCty
extracCon
(&
dicConaries)
 Time
consuming
 Mirror
human
thinking
 Case,
subject
maRer
 Transparent
System
Demys&fying
Technology
Assisted
Review

23. StaCsCcal
Based
Technologies
 Bayesian
ClassificaCon
 ProbabilisCc
 Co-­‐occurrence
 Frequency
 Spam
Filters
 Viagra
 Concepts
 Words,
phrases
Demys&fying
Technology
Assisted
Review

24. Bayesian
 Bayesian
illustraCon
 Baseball,
glove,
diamond,
bats,
hit,
home
run
 Diamond,
pendant,
jewelry
 Co-­‐occurrence
 Local
–
within
a
document
 Global
–
across
document
populaCon
 Frequency
–
how
ozen
does
it
appear
 WeighCng
–
uniqueness
counts
Demys&fying
Technology
Assisted
Review

25. StaCsCcal
Based
Technologies
 Vector
Space
Modeling
 Latent
Seman&c
Indexing/Analysis
 Words
 Phrases,
Concepts
 Tables
 Algebraic
equaCons
represenCng
docs
 WeighCng
Algorithms
Demys&fying
Technology
Assisted
Review

26. Latent
SemanCc
Indexing
Example
 Exclude
Noise
Words
 The,
and,
or,
etc.
 Vector
Space
Modeling
 Build
Document
Profile
 Diamond
 Base,
ball
 Necklace,
pendant
 Diamond
Saw
Demys&fying
Technology
Assisted
Review

27. MathemaCcal
FoundaCon
 Tables
built
with
0s,
1s
 Yes
it
has
that
word
or
phrase
 No
it
doesn’t
Demys&fying
Technology
Assisted
Review

28. Simple
Matrix
with
WeighCng
Demys&fying
Technology
Assisted
Review

29. Weighted
by
Document
(not
just
type)
Demys&fying
Technology
Assisted
Review

30. Defensibility
Report
 Document,
Document,
Document
 Transparency
 Workflow
 What
Was
Considered,
By
Whom?
 QC
Process
 Metrics
Demys&fying
Technology
Assisted
Review

31. Q&A - Thank you!
Post
your
ques&ons
to
the
presenter
in
the
chat
secCon
Sonya
L.
Sigler
Vice
President,
Product
Strategy
&
Consul&ng
SFL
Data
415-­‐321-­‐8385
sonya@sfldata.com
www.sfldata.com
Demys&fying
Technology
Assisted
Review