Cognitive Systems Institute Speaker Series talk by Mona Diab from George Washington University on May 14, 2015 "Towards Building Effective Computational Sociopragmatics Models of Human Cognition."
1. Towards
building
effec2ve
computa2onal
sociopragma2cs
models
of
human
cogni2on
Mona
Diab
George
Washington
University
2. Acknowledgement
• Many
collaborators:
Dragomir
Radev,
Amjad
Abu
Jbara,
Pradeep
Dasigi,
Weiwei
Guo,
Owen
Rambow,
Julia
Hirschberg,
Kathy
Mckeown,
Mustafa
Mughazy,
Heba
Elfardy,
Vinod
Prabhakaran,
Greg
Werner,
Muhammad
Abdulmageed
• Research
supported
by
IARPA
SCIL
program
and
DARPA
DEFT
&
BOLT
programs
and
Google
Faculty
award
• Slides
adapted
from
several
publica2on
presenta2ons
3. What
is
sociopragma2cs?
• The
aspect
of
language
use
that
relates
to
everyday
social
prac.ces.
hVp://www.wordsense.eu
dic2onary
– What
are
social
prac2ces?
Well
…
from
our
language
focused
prism
J
• Interac2ons,
expressions
of
emo2ons/beliefs/opinions,
etc.
4.
Text
and
Social
Rela2ons
We
can
use
linguis2c
analysis
techniques
to
understand
the
implicit
rela2ons
that
develop
in
on-‐line
communi2es
Image
source:
clair.si.umich.edu
5. Overarching
Agenda
• Goal:
AVempt
to
mine
social
media
text
for
clues
and
cues
on
understanding
human
interac2ons
• How:
Iden2fy
interes2ng
sociolinguis2c
behaviors
and
correlate
them
with
linguis2c
usage
that
are
quan2fiable
devices
and
build
effec2ve
models
in
the
process
• Compare
these
devices
cross
linguis2cally
6.
Many
Different
Forms
of
Social
Media
• Communica2on
• Collabora2on
• Mul2media
• Reviews
&
opinions
7. Social
Media
Explosion
source:
www.internetworldstats.com
>3
billion
Internet
users
worldwide
>42.3%
popula2on
penetra2on
(>48%
in
the
MENA
region)
75%
of
them
used
“Social
Media”
8.
Text
in
Social
Media
Some
social
media
applica2ons
are
all
about
text
9.
Text
in
Social
Media
Even
the
ones
based
on
photos,
videos,
etc.
generate
a
lot
of
discussions
10.
Text
in
Social
Media
Huge
amount
of
text
exchanged
in
discussions
11. Do
you
s2ll
need
convincing
that
text
is
important!
Yeah
I
thought
not!
Just
checking
J
13. Approach
to
processing
social
construct
phenomena
(Direc.ve
from
the
IARPA
SCIL
Program)
• Iden2fy
language
uses
(LU)
per2nent
to
the
different
social
constructs
(SC)
• Correlate
the
LUs
with
Linguis2c
Construc2ons/
Cons2tuents
(LC)
14. Social
Construct:
Influencer
(inf)
• Language
Uses
– AVempt
to
Persuade
– Agreement/Disagreement
– Level
of
CommiVed
Belief
Influencers
15. Social
Construct:
Pursuit
of
Power
(PoP)
• Language
Uses
– AVempt
to
Persuade
– Agreement/Disagreement
– Level
of
CommiVed
Belief
– Nega2ve/Posi2ve
Aktude
– Who
is
talking
about
whom
– Dialog
PaVerns
(non
linguis2c)
Pursuit
of
Power
16. Social
Construct:
Subgroup
(Sub)
• Language
Uses
– Agreement/Disagreement
– Nega2ve/Posi2ve
Aktude
– Sarcasm
– Level
of
CommiVed
Belief
– Signed
Network
(non
linguis2c)
Mul2ple
Viewpoints
(Subgroups)
17. LUs
in
our
approach
• AVempt
to
Persuade
(Inf,
PoP)
• Agreement/Disagreement
(Inf,
PoP,
Sub)
• Level
of
CommiVed
Belief
(Inf,
PoP)
• Nega2ve/posi2ve
aktude
(Sub,
PoP)
• Sarcasm
(Sub)
• Who
is
talking
about
whom
(PoP)
• Dialog
PaVerns
(PoP)
• Signed
Network
(Sub)
Do
not
depend
on
linguis6c
analysis
Rely
on
linguis6c
analysis
18. Cross
language
comparison:
Generaliza2ons
• In
general
similar
LU
level
devices
cross
linguis2cally
• AVempt
to
persuade
– Claim:
grounding
in
experience,
commonly
respected
sources
– Argumenta2on:
evidence
and
support
from
other
discussants
• Agreement/Disagreement
– Shared
opinion
(explicit
expression),
shared
perspec2ve
(implicit
aktude)
• Level
of
CommiVed
Belief
– CommiVed:
The
sun
will
rise
tomorrow
– Non
commiVed:
John
may
believe
that
the
moon
is
made
of
cheese
19. Generaliza2ons
• -‐ve/+ve
aktude
– Nega2ve
language
– Sen2ment/word
polarity
• Who
is
talking
about
whom
– Use
of
men2ons
and
their
frequency
20. But
how
do
they
differ
in
their
linguis2c
expression?
• Arabic
vs.
English
social
media
use
different
linguis2c
cons2tuents
(LC)
to
exhibit
language
use
21. Focus
of
this
talk
Influencers
Pursuit
of
Power
Disputed
Topics
Mul2ple
Viewpoints
(Subgroups)
23. Example
The
new
immigra2on
law
is
good.
Illegal
immigra2on
is
bad.
Peter
I
totally
disagree
with
you.
This
law
is
blatant
racism.
Mary
Have
you
read
all
what
Peter
wrote?
He
is
correct.
Illegal
immigra2on
is
bad
and
must
be
stopped.
John
You
are
clueless,
Peter.
Stop
suppor2ng
racism.
Alexander
Peter
John
Support
the
new
law
Against
the
new
law
Mary
Alexander
27. 1
-‐
Thread
Parsing
The
new
immigra2on
law
is
good.
Illegal
immigra2on
is
bad.
Peter
I
totally
disagree
with
you.
This
law
is
blatant
racism.
Mary
Have
you
read
all
what
Peter
wrote?
He
is
correct.
Illegal
immigra2on
is
bad
and
must
be
stopped.
John
You
are
clueless,
Peter.
Stop
suppor2ng
racism.
Alexander
P1
P2
P3
P4
D1
D2
D3
D4
Iden2fy
Posts,
Discussants,
and
the
reply
structure
of
the
discussion
thread
29. 2
-‐
Iden2fy
Opinion
Words*
The
new
immigra2on
law
is
good+.
Illegal
immigra2on
is
bad-‐.
Peter
I
totally
disagree-‐
with
you.
This
law
is
blatant-‐
racism-‐.
Mary
Have
you
read
all
what
Peter
wrote?
He
is
correct+.
Illegal
immigra2on
is
bad-‐
and
must
be
stopped.
John
You
are
clueless-‐,
Peter.
Stop
suppor2ng
racism.
Alexander
P1
P2
P3
P4
D1
D2
D3
D4
*Iden2fying
opinion
words
using
Opinion
Finder
with
an
extended
lexicon
(implemented
using
random
walks
–
Hassan
&
Radev,
2011)
31. 3-‐
Iden2fy
Candidate
Targets
of
Opinion
Target
Discussant
(
e.g.
you,
Peter)`
Topic/En1ty
(e.g.
The
new
immigra2on
Law,
Illegal
Immigra2on)
32. Candidate
Targets
3-‐
Iden2fy
Candidate
Targets
of
Opinion
The
new
immigra2on
law
is
good+.
Illegal
immigra2on
is
bad-‐.
Peter
I
totally
disagree-‐
with
you.
This
law
is
blatant-‐
racism-‐.
Mary
Have
you
read
all
what
Peter
wrote?
He
is
correct+.
Illegal
immigra2on
is
bad-‐
and
must
be
stopped.
John
You
are
clueless-‐,
Peter.
Stop
suppor2ng
racism.
Alexander
P1
P2
P3
P4
D1
D2
D3
D4
All
discussants
are
candidate
Targets
33. Candidate
Targets
3-‐
Iden2fy
Candidate
Targets
of
Opinion
The
new
immigra2on
law
is
good+.
Illegal
immigra2on
is
bad-‐.
Peter
I
totally
disagree-‐
with
you.
This
law
is
blatant-‐
racism-‐.
Mary
Have
you
read
all
what
Peter
wrote?
He
is
correct
+.
Illegal
immigra2on
is
bad-‐
and
must
be
stopped.
John
You
are
clueless-‐,
Peter.
Stop
suppor2ng
racism.
Alexander
P1
P2
P3
P4
D1
D2
D3
D4
D1
D1
D1
Iden2fy
discussant
men2ons
(2pp
or
name)
in
the
discussion
D2
34. Candidate
Targets
3-‐
Iden2fy
Candidate
Targets
of
Opinion
The
new
immigra2on
law
is
good+.
Illegal
immigra2on
is
bad-‐.
Peter
I
totally
disagree-‐
with
you.
This
law
is
blatant-‐
racism-‐.
Mary
Have
you
read
all
what
Peter
wrote?
He
is
correct
+.
Illegal
immigra2on
is
bad-‐
and
must
be
stopped.
John
You
are
clueless-‐,
Peter.
Stop
suppor2ng
racism.
Alexander
P1
P2
P3
P4
D1
D2
D3
D4
D1
D1
D1
D1
Peter
Iden2fy
anaphoric
men2ons
of
discussants
D2
35. Candidate
Targets
3-‐
Iden2fy
Candidate
Targets
of
Opinion
The
new
immigra1on
law
is
good+.
Illegal
immigra1on
is
bad-‐.
Peter
I
totally
disagree-‐
with
you.
This
law
is
blatant-‐
racism-‐.
Mary
Have
you
read
all
what
Peter
wrote?
He
is
correct
+.
Illegal
immigra1on
is
bad-‐
and
must
be
stopped.
John
You
are
clueless-‐,
Peter.
Stop
suppor2ng
racism.
Alexander
P1
P2
P3
P4
D1
D2
D3
D4
D1
D1
D1
D1
Peter
Topic1
Topic1
Topic2
Topic2
D2
Topic
1
Topic
2
36. 3-‐
Iden2fy
Candidate
Targets
of
Opinion
• Techniques
used
to
iden2fy
topical
targets
– Named
En2ty
Recogni2on
– Noun
phrase
chunking
40. Candidate
Targets
4-‐
Opinion-‐Target
Pairing
The
new
immigra1on
law
is
good+.
Illegal
immigra1on
is
bad-‐.
Peter
I
totally
disagree-‐
with
you.
This
law
is
blatant-‐
racism-‐.
Mary
Read
all
what
Peter
wrote.
He
is
correct+.
Illegal
immigra1on
is
bad-‐
and
must
be
stopped.
John
You
are
clueless-‐,
Peter.
Stop
suppor2ng
racism.
Alexander
P1
P2
P3
P4
D1
D2
D3
D4
D1
D1
D1
D1
Peter
Topic1
Topic1
Topic2
Topic2
Topic
1
Topic
2
41. 4-‐
Opinion-‐Target
Pairing
• Language
Uses
(LUs)
present
in
this
step:
– Targeted
sen2ment
toward
other
discussants
(2nd
person)
– Targeted
Sen2ment
toward
topic
men2ons
(3rd
person)
I
totally
disagree
-‐
with
you.
This
law
is
blatant
-‐
racism
-‐.
42. 4-‐
Opinion-‐Target
Pairing
• LU
details
– Rule-‐based
detec2on
of
sen2ment
targets
(we’ve
also
been
experimen2ng
with
supervised
target
detec2on
methods)
– Discussant
targets
are
iden2fied
by
2nd
person
pronouns
(you,
your,
yourself,
etc.)
and
by
username
men2ons
(casper3912,
etc.)
50. English
Data
• 117
Discussions
• Short
threads
• short
posts
• Human
annota2on
• More
formal
• 12
Polls
+
Discussions
• Long
threads
• Long
and
short
posts
• Data
self-‐labeled
• Less
formal
• 30
debates
• Long
threads
• Long
and
short
posts
• Data
self-‐labeled
• Less
formal
52. Arabic
Data
• Forum
for
2
sided
self
labeled
poli2cal
debates
www.naqeshny.com
• 36
debates
comprising
711
posts
corresponding
to
326
users
•
• The
average
number
of
posts
per
discussion
19.75
and
average
number
of
discussants
per
topic
13.08
54. Evalua2on
Metrics
2. Entropy
3. F-‐Measure
where
P(I,
j)
is
the
probability
of
finding
an
element
from
the
category
i
in
the
cluster
j,
nj
is
the
number
of
items
in
cluster
j,
and
n
the
total
number
of
items
in
the
distribu2on.
55. Baselines
• Interac2on
Graph
Clustering
(GC)
– Nodes:
Par2cipants
– Edges:
interac2ons
(connect
two
par2cipants
if
they
exchange
posts)
• Text
Classifica2on
(TC)
– Build
TF-‐IDF
vectors
for
each
par2cipant
(using
all
his/
her
posts)
– Cluster
the
vector
space
60. Comparison
to
baselines
Our System
English
Results
Arabic
Results
Method
P
E
Signed
Network
0.71
0.68
Our
System
0.67
0.72
61. Wikipedia
Poli1cal
Forum
Create
debate
Purity
0.66
0.61
0.64
Entropy
0.55
0.80
0.68
F-‐measure
0.61
0.56
0.60
English
Results
62. Wikipedia
Poli1cal
Forum
Create
debate
Purity
0.66
0.61
0.64
Entropy
0.55
0.80
0.68
F-‐measure
0.61
0.56
0.60
English
Results
Best
performing
63. Wikipedia
Poli1cal
Forum
Create
debate
Purity
0.66
0.61
0.64
Entropy
0.55
0.80
0.68
F-‐measure
0.61
0.56
0.60
English
Results
Best
Performing
&
Worst
Performing
64. Component
Evalua2on
Our
System
No
Topical
Targets
No
Discussant
Targets
No
Sen1ment
No
Interac1on
No
Anaphora
Resolu1on
No
Named
En1ty
Recog.
No
NP
Chunking
65. Component
Evalua2on
Our
System
No
Topical
Targets
No
Discussant
Targets
No
Sen1ment
No
Interac1on
No
Anaphora
Resolu1on
No
Named
En1ty
Recog.
No
NP
Chunking
Not really a linguistic feature
66. Component
Evalua2on
Our
System
No
Topical
Targets
No
Discussant
Targets
No
Sen1ment
No
Interac1on
No
Anaphora
Resolu1on
No
Named
En1ty
Recog.
No
NP
Chunking
More of a linguistic feature!
67. Deeper
look
at
Agreement/
Disagreement
• So
far
we
employed
shared/divergent
opinion
in
the
form
of
explicit
polarity
indicators
– Sen2ment
polarity
towards
other
discussants
• A:
So,
no
maHer
how
much
faith
you
have,
one
of
you
MUST
be
wrong!
(nega.ve)
• B:
You
are
a
scien.st?!
May
I
ask
in
which
field?
(nega.ve)
– Sen2ment
polarity
towards
an
en.ty
• A:
Here
is
an
excellent
verse
from
the
Bible..
(posi.ve)
• B:
The
Bible
rightly
says
that...
(posi.ve)
68. Implicit
Opinion/Perspec2ve
• Observa2on:
People
sharing
similar
beliefs/perspec2ve
tend
to
use
the
same
evidence
to
support
their
point
– Believers:
faith,
peace,
love,
ci2ng
verses
from
the
Bible...
– Atheists:
reason,
science,
aVack
on
the
“logical”
flaws
in
Bible...
• However
it
is
not
always
explicit
(using
similar
words
and
similar
aktudes)
• Peter:
God
is
the
creator
of
mankind
• Mary:
The
belief
in
an
ul2mate
divine
being
has
sustained
me
over
the
years
– Not
necessarily
posi2ve/nega2ve
– High
dimensional
similarity
between
both
sentences
is
low!
– BUT
we
know
Mary
and
Peter
share
the
same
perspec1ve
and
will
tend
to
be
in
agreement
with
each
other
69. Modeling
of
implicit
agreement/
disagreement
• Implicit
agreement
or
disagreement
(perspec2ve)
–
using
text
similarity
to
help
iden2fy
subgroups
• Perspec2ve
modeling
is
used
to
complement
explicit
aktude
• Perspec2ve
granularity
has
to
be
collected
on
the
level
of
a
thread
rather
than
a
single
post
• Hence
we
summarize
all
the
posts
in
the
thread.
70. Our
Model
• Explicit
high
dimensional
aktude
toward
other
discussants
and
en22es
• Modeling
shared
perspec2ve
among
discussants
over
threads
using
textual
similarity
on
the
post
level
in
the
latent
space
71. Extrac2ng
implicit
perspec2ve
• Run
Latent
Dirichlet
Alloca2on(LDA)
on
the
thread
• Extract
the
topic
distribu2on
of
each
post
• Aggregate
the
distribu2ons
of
all
posts
between
each
pair
of
discussants
72. FEATURE
REPRESENTATION:
ATTITUDE
PROFILES
• Vector
Representa2on
• Explicit
aktude
towards
other
discussants
and
En22es
A
B
C
E1
E2
A
0
0
0
1
1
2
0
1
1
1
0
1
0
0
0
B
…
C
-‐-‐
73. FEATURE
REPRESENTATION:
ATTITUDE
PROFILES
• Vector
Representa2on
• Implicit
agreement
with
other
discussants
A
B
C
E1
E2
A
B
C
A
0
0
0
1
1
2
0
1
1
1
0
1
0
0
0
1
1
1
1
0
0.5
0.5
0
0
B
…
C
-‐-‐
1
1
1
74. Data
• English
– Create
Debate
(CD)
• www.createdebate.com
• Deba2ng
on
a
certain
topic
• Sides
are
explicitly
indicated
by
discussants
in
a
poll
Informal
language
– Wikipedia
Discussion
Forum
(WIKI)
• en.wikipedia.org
• Groups
labels
are
manually
annotated
• Formal
language,
not
much
nega2ve
polarity
• Arabic
– www.naqeshny.com
– Self
labeled
poli2cal
debates
75. Experimental
Condi2ons
• Clustering
algorithm
– S-‐Link
#
of
clusters
by
rule
of
thumb
=
√n/2
• Evalua2on
Metrics
– Purity,
Entropy,
F-‐measure
• Baseline
– RAND-‐BASE:
Assign
discussants
to
clusters
randomly
– SWD-‐BASE:
Calculate
surface
word
distribu2on,
as
a
simpler
form
of
perspec2ve
77. Observa2ons
Condi1on
Wiki
CD
Purity
Entropy
Fmeasure
Purity
Entropy
Fmeasure
RAND-‐BASE
0.675
0.563
0.652
0.399
0.966
0.41
SWD-‐BASE
0.772
0.475
0.646
0.452
0.932
0.432
SD
0.834
0.360
0.667
0.824
0.394
0.596
SE
0.827
0.383
0.655
0.793
0.422
0.582
SD+SE
0.835
0.362
0.665
0.82
0.385
0.604
PERS
0.853
0.321
0.699
0.787
0.399
0.589
SD+PERS
0.853
0.320
0.698
0.849
0.333
0.615
SE+PERS
0.853
0.321
0.702
0.789
0.399
0.591
SD+SE+PERS
0.857
0.310
0.703
0.861
0.315
0.625
Best
Performance
is
when
we
combine
explicit
aktude
(SD
Sen2ment
toward
other
discussants,
SE
Sen2ment
toward
En22es)
with
implicit
perspec2ve
(PERS),
regardless
of
genre
78. Observa2ons
Condi1on
Wiki
CD
Purity
Entropy
Fmeasure
Purity
Entropy
Fmeasure
RAND-‐BASE
0.675
0.563
0.652
0.399
0.966
0.41
SWD-‐BASE
0.772
0.475
0.646
0.452
0.932
0.432
SD
0.834
0.360
0.667
0.824
0.394
0.596
SE
0.827
0.383
0.655
0.793
0.422
0.582
SD+SE
0.835
0.362
0.665
0.82
0.385
0.604
PERS
0.853
0.321
0.699
0.787
0.399
0.589
SD+PERS
0.853
0.320
0.698
0.849
0.333
0.615
SE+PERS
0.853
0.321
0.702
0.789
0.399
0.591
SD+SE+PERS
0.857
0.310
0.703
0.861
0.315
0.625
Wiki
seems
to
gain
more
from
implicit
perspec2ve
compared
to
CD
Explicit
Aktude
is
a
beVer
feature
for
CD:
people
express
their
sen2ments
openly,
while
in
Wiki
people
are
more
constrained
and
subtle
in
their
expressions
79. Observa2ons
Condi1on
Wiki
CD
Purity
Entropy
Fmeasure
Purity
Entropy
Fmeasure
RAND-‐BASE
0.675
0.563
0.652
0.399
0.966
0.41
SWD-‐BASE
0.772
0.475
0.646
0.452
0.932
0.432
SD
0.834
0.360
0.667
0.824
0.394
0.596
SE
0.827
0.383
0.655
0.793
0.422
0.582
SD+SE
0.835
0.362
0.665
0.82
0.385
0.604
PERS
0.853
0.321
0.699
0.787
0.399
0.589
SD+PERS
0.853
0.320
0.698
0.849
0.333
0.615
SE+PERS
0.853
0.321
0.702
0.789
0.399
0.591
SD+SE+PERS
0.857
0.310
0.703
0.861
0.315
0.625
BeVer
results
obtained
on
the
same
data
set
from
the
previous
results
for
Wiki
(P
0.66,
E
0.55)
CD
(P
0.64,
E
0.68)
80. Arabic
Results
Using
EM
Purity
Entropy
F-‐measure
Signed
Network
BASELINE
0.71
0.68
0.67
Explicit
Aktude
0.67
0.72
0.65
Implicit/Perspec2ve
0.64
0.74
0.65
Our
System
(combined)
0.77
0.50
0.76
81. Arabic
Results
Using
EM
Purity
Entropy
F-‐measure
Signed
Network
BASELINE
0.71
0.68
0.67
Explicit
Aktude
0.67
0.72
0.65
Implicit/Perspec2ve
0.64
0.74
0.65
Our
System
(combined)
0.77
0.50
0.76
Significant
improvement
over
baseline
82. Arabic
Results
Using
EM
Purity
Entropy
F-‐measure
Signed
Network
BASELINE
0.71
0.68
0.67
Explicit
Aktude
0.67
0.72
0.65
Implicit/Perspec2ve
0.64
0.74
0.65
Our
System
(combined)
0.77
0.50
0.76
Significant
improvement
over
baseline
Complementarity
between
Explicit
aktude
and
Perspec2ve
83. Conclusions
• We
can
successfully
model
sociopragma2c
phenomena
– Golden
rule
of
computer
science
(divide
and
conquer)
Form
subgroups
J
• There
is
significant
room
for
improvement
• It
takes
a
large
team
of
computer
scien2sts
and
significant
collabora2on
with
the
humani2es
to
get
this
program
going
84. Where
are
we
now?
• Extensive
work
on
Sen2ment
and
Emo2on
Intensity
characteriza2on/detec2on
• Work
on
Rumor
Detec2on
• Work
on
Level
of
CommiVed
Belief
Tagging
(check
us
out
at
*SEM
2015,
and
EXPROM
2015)
• Work
on
Ideological
Perspec2ve
Detec2on
(check
us
out
at
*SEM
2015)