31 января, семинар "День MIT в Яндексе"
Регина Барзилай "Извлечение информации из социальных медиа"
- Методы машинного обучения в применении к извлечению информации из сетевого пользовательского контента.
- Рассмотрение набора задач, связанных с извлечением информации, таких как анализ рецензий по составляющим и создание базы событий по твитам.
- Автоматическое построение контентной структуры документа на основе большого потока пользовательского контента с сильным шумом.
- Автоматическая агрегация содержимого рецензий и извлечении событий из потока сообщений в твиттере.
5. User-‐generated
Content
• Large
amounts
of
user-‐generated
content
• Increasingly
important
in
decision
making
• Time-‐consuming
to
read
it
all
NLP can help!
5
6. The
Power
of
Word
Counts
Simple
sta(s(cal
models
are
effec(ve
for
many
informa(on
extrac(on
tasks
• Bag-‐of-‐words
approaches
for
classifica(on
trading
financial
stocks
bank
cloudy
cold
storm
plants
6
7. The
Power
of
Word
Counts
Simple
sta(s(cal
models
are
effec(ve
for
many
Informa(on
Extrac(on
tasks
• Sequence
labeling
for
seman(c
role
labeling
the
earthquake
injured
three
people
NONE
EVENT
NONE
CASUALTIES
CASUALTIES
7
11. Our
Approach
• Model
document
structure
as
part
of
the
extrac(on
process
• Exploit
large
amount
of
raw
data
to
supplement
annota(ons
11
12. I
wandered
in
here
on
a
whim
with
a
friend
a
while
back,
completely
underdressed
but
on
the
lookout
for
This
is
a
fantas(c
restaurant
in
Cambridge.
The
decor,
So
so
so
good!
One
of
my
favorite
restaurants
in
a
good
meal.
When
we
went
around
the
side
to
the
music
and
smells
will
make
you
feel
like
you
are
in
Boston!
I
have
to
take
off
a
single
star,
because
there
The
suspense
factor
as
each
surprise
dish
was
entrance,
someone
called
to
us
from
the
roof-‐-‐"watch
another
world.
I
was
a
lille
skep(cal
when
I
heard
are
a
couple
dishes
I
didn't
enjoy,
but
if
you
go
and
delivered
added
to
the
whole
experience,
and
the
out,
there's
glass
on
the
floor,
I
dropped
a
light
bulb.
Turkish
Food,
but
those
feelings
were
quickly
order
well,
this
is
a
6
star
experience!
(ming
was
flawless-‐-‐we'd
finish
a
dish,
have
ample
Can
you
go
inside
and
grab
someone
for
me
and
ask
squashed
when
I
had
the
food.
Had
dinner
here
on
Friday
night
and
it
was
superb!!
A
new
favorite!
them
to
bring
a
screwdriver?"
Sure,
no
problem-‐-‐so
Oleana
serves
inspired,
well
prepared
food
from
the
I
(me
to
enjoy
it,
a
few
minutes
for
some
sips
of
wine
You
start
with
bread
and
the
most
delicious
olive
oil.
It
The
hosts
seated
us
almost
immediately
(even
and
conversa(on,
and
just
as
there
was
a
breath,
the
went
inside
and
told
the
hostess,
"Hey
your
best
possible
ingredients.
The
menu
is
well
priced
for
started
with
the
Fried
Mussels
with
Hot
Peppers
We
Cute
ambiance...great
for
in(mate
dinner
or
date
#1.
place.
Dim
ligh(ng,
nice
decor,
not
too
loud.
has
a
strong
olive
taste,
I
had
to
forcefully
stop
myself
maintenance
guy
needs
a
screwdriver
up
on
the
roof."
quality.
The
wine
list
is
very
food
friendly,
the
and
Turkish
Tarator
Sauce.
The
Mussels
were
fried
to
though
we
didn't
have
a
reserva(on
on
a
Friday
night).
dish
would
arrive.
The
whole
meal
felt
like
a
well-‐
next
Delicious.
Delighsul.
Worth
it.
from
finishing
our
basket.
orchestrated
performance
rather
than
just
a
meal.
She
laughed,
"Oh...
that's
not
a
maintenance
guy,
includes
many
organic
and
bio-‐dynamic
wines,
and
iperfec(on
and
the
baler
was
very
light.
I
could
have
s
that's
the
owner."
Service
was
excellent
as
well
as
all
the
also
reasonably
priced.
It's
a
great
place
for
eaten
a
thousand
of
them,
but
I
love
fried
food.
The
#2.
The
food
was
amazing
(N.B.
the
bread
in
the
Vermont
Quail
was
very
tasty
as
well.
The
quail
was
I've
eaten
at
Oleana
several
(mes
and
the
food
is
recommenda(ons.
My
sugges(on
is
to
order
family
style,
and
based
basket
goes
really
well
with
the
warmed,
spiced
olive
for
the
food-‐-‐there's
not
enough
to
be
said.
I
realize
on
As
vegetarians.
A
vegetarian
tas(ng
menu
was
available
the
number
of
people
you
have,
order
main
entrees
And
from
that
moment
I
knew
this
place
was
special.
night
I
went
-‐
it
was
superb
and
plen(ful
(I
could
the
very
tender,
which
is
hard
to
do
because
those
lille
always
very
good.
The
service
is
typically
really
solid
-‐
oil
from
one
of
our
meze
plates),
We
had
the
olives
this
is
a
lille
cliched,
but
honestly,
I
haven't
tasted
w/
for
half
your
party
size
and
3-‐4
small
plates
for
the
food
like
this
since
living
in
the
bay
area,
with
crea(ve
a
good
sign
when
the
owner
of
a
restaurant
is
up
finish
it).
It's
not
guys
are
so
small.
The
last
starter
was
the
Sultans
I've
had
some
dinners
where
the
waitstaff
was
really
Started
with
the
Moroccan-‐style
Octopus
and
Fatoush.
za'atar
(a
lot
for
two
people,
$5),
quail
kebabs
top-‐notch
and
other
(mes
when
it's
been
good
but
Tasty,
light,
unique
flavors,
great
presenta(on.
other
half.
(ex.
Party
of
4
=
2
entrees
and
6-‐8
apps)
combina(ons
and
well-‐designed
dishes
and
contrasts
the
roof
changing
lightbulbs-‐-‐it's
clear
what
kind
of
on
Delight.
The
Tamarind
Glazed
Beef
was
so
tender.
The
(delicious,
tender,
spicy,
2
for
$13),
monkfish
curry
not
spectacular.
Everything
was
quickly
eaten
up
with
smiles.
(yummy,
$26)
and
a
tea-‐ser
dessert
with
sour
cherries
surprise
the
palele
and
are
a
delight
to
eat;
not
and
alen(on
goes
into
every
detail.
If
only
I
had
service
is
friendly,
prompt,
and
helpful.
The
space
that
care
The
Smokey
Eggplant
Puree
went
well
with
the
dish
and
The
entrees
really
are
to
die
for,
every
one
I've
had
tastes
a
lot
like
vermicelli/milk
dessert
(12).
simply
good,
but
joy-‐inducing.
The
highlight
of
the
known
then
how
that
would
translate
to
the
en(re
is
relaxing
and
casually
elegant.
I
was
there
on
a
was
even
beler
slathered
on
the
bread.
that
Either
way
the
hummus
and
falafel
small
plates
(meze)
For
entrees,
had
the
Beef
Kabob
special
(which
was
has
been
delicious.
Such
a
great
combo
of
flavors
and
night
was
a
dish
of
crab
cakes
with
asparagus-‐-‐they
experience,
and
especially
the
food.
Tuesday
evening
when
two
men
were
quietly
playing
spices,
they
some
seriously
ar(s(c
crea(ons!
The
best
lovely
world
music.
We
shared
an
entree
for
the
evening.
I
highly
suggest
just
SO
tasty.
Definitely
a
great
place
to
celebrate
the
hit!,
beef
with
delicious
flavors
and
cooked
are
#3.
The
service
was
excellent-‐-‐perfectly
(med
plates,
a
small
poached
quail
egg
on
each
cake,
combined
had
a
special
event
or
just
when
you
need
a
par(cular
perfectly
med-‐rare
tender),
Cod,
and
Lamb.
Everything
small
plates
are
the
Sultans
Delight
(fall
apart
lamb
with
a
lemon
flavor
from
a
lille
juice
and
zest
it
Once
out
on
the
pa(o,
we
waited
for
only
a
few
the
Azuluna
Pork,
Crispy
Pea
Paella,
Fried
Fiddleheads
etc.
pick-‐me-‐up.
Not
the
fussiest
or
the
fanciest
food
(this
was
tasty
yet
light,
with
delicate
complimen(ng
and
unreal
baba),
the
spiced
carrots
(seems
simple,
almost
made
a
meringue
that
was
incredible.
This
is
minutes
by
the
fountain
before
being
seated.
I
had
a
I
was
visi(ng
the
Boston
area
with
my
family,
we
and
Paprika
Sauce.
The
pork
was
just
as
tender
as
the
As
is
a
compliment
from
me!)
nor
the
most
elegant
flavors.
but
they
are
amazing!),
the
falafel....ok,
there
are
a
Ana
Sortun
was
there!
She's
awesome
and
she
was
#4.
without
a
doubt
the
best
single
dish
I
have
had
at
any
of
wine
and
was
just
enjoying
the
pa(o,
the
brought
along
our
children
(8&10).
While
I
would
not
and
the
accompaniments
went
so
well
with
the
glass
beef
ambiance
(wish
it
was
a
lille
more
quiet),,
but
clearly
bunch
that
are
great,
but
those
are
some
favorites!
hands-‐on-‐-‐adding
things
to
dishes,
etc.
restaurant
in
Boston,
and
I'd
go
back
to
Oleana
just
bread,
and
the
otherworldly
feel
of
the
place.
You
call
Oleana
family-‐friendly,
they
were
accommoda(ng
The
meal
was
very
flavorful
and
seasoned
to
for
meat.
a
very
special
place
for
a
great
meal!
Plus
you
NEED
to
Dessert
was
the
winner-‐Passion
fruit
this.
Or
any
other
of
the
dishes
we
had
that
night,
can't
be
stressed
out
here;
it's
designed
perfectly
to
our
children.
I
should
note
one
of
my
children
is
a
of
perfec(on.
I
would
skip
the
deviled
eggs,
mul(ple
people
talked
frankly.
Some
were
beler
than
others,
but
each
was
almost
a
Shangri-‐La
of
spaces,
and
all
in
the
middle
selec(ve
eater,
but
they
are
both
used
to
ea(ng
be
highly
get
the
Baked
Alaska
(YUM).
It
should
be
a
Bisteeya....goodness
what
is
this??
IT
was
to
DIE
for
I
would
go
back-‐-‐I
s(ll
like
the
hummus
at
Sofra
and
requirement
for
going
to
the
place.
and
a
perfect
ending
to
the
meal.
I
think
we
literally
them
up
to
me
before
I
went,
really
not
that
exci(ng.
unique
and
in
some
way
surprising
and
fun.
of
Inman
square.
Unexpected.
Impeccable.
in
high-‐end
restaurants
and
a
very
well
behaved
in
I
wish
I
could
have
golen
dessert,
but
I
was
so
full.
I
wasn't
sure
that
Oleana
would
be
more
impressive-‐-‐
ate
this
in
2
seconds
and
contemplated
ordering
a
The
chick
peas
in
vermicelli,
also
not
great,
in
fact
I
it
was
excellent.
but
nice
restaurants
(or
so
we
are
told).
did
see
them
bring
some
out
and
they
looked
really
did
not
like
it
with
the
fake
orange
taste.
It
Absolutely
fantas(c
experience
all
around,
and
so
far
this
point
we
were
so
confident
in
the
holis(c
At
wonderful
and
decadent.
I
had
the
Sangria
to
drink,
(I
would
give
the
food
5
stars;
the
overall
ambiance
-‐
second.
It's
light,
tart,
fluffy,
creamy,
and
thirst
it
reminded
me
of
one
of
those
chocolate
oranges
you
quality
of
the
restaurant
that
we
decided
to
trust
the
Given
that
I
live
in
the
San
Francisco
Bay
Area,
I'm
was
refreshing
because
it
was
humid
outside,
but
it
read:
it
can
be
prely
loud
-‐
and
waitstaff
variability
quenching
all
at
the
same
(me.
hlp://condensr.com
Great
place
for
a
special
occasion
meal
or
a
night
out
best
overall
restaurant
experience
I've
had
in
the
the
knocked
the
overall
experience
to
a
4).
can
break
apart.
The
black
eyed
pea
soup
was
nothing
Boston
area.
These
people
understand
that
a
meal
is
and
go
all
out.
We
ordered
a
tas(ng
menu,
and
chef
spoiled
by
excellent
vegetarian
friendly
restaurants.
not
the
best
in
the
world.
I
think
I
might
try
was
with
tourists.
Needless
to
say
Oleana
became
a
favorite
in
one
special.
more
than
just
about
the
food,
it's
about
the
service,
other
mezos/appe(zers
that
sounded
good
from
I
have
a
spot
to
return
to
in
the
Boston
area
that
two
Now
something
new
if
I
ever
go
back
again.
12
evening...the
food
is
just
very
unique.
I
love
have
great
the
wine,
the
scenery,
and
on
top
of
all
that,
the
the
specials
menu.
If
you're
here,
I
*highly*
meets
expecta(ons.
For
dessert,
please
get
the
baked
Alaska,
it
was
recommend
this.
You
probably
won't
end
up
spending
The
servers
are
great,
so
nice
and
knowledgeable
flavors
without
feeling
like
I
gained
10
pounds
ea(ng
a
flavors
and
combina(ons
of
culinary
delights.
Oleana
unbelievable!
any
more
than
if
you
had
ordered
individually,
but
about
the
menu.
It
really
means
a
lot
to
me
to
see
wonderful
dinner.
I'll
definitely
be
back
soon!
turned
an
otherwise
ordinary
night
into
an
experience
I
won't
forget,
and
I
can't
wait
to
return.
you'll
taste
some
incredible
things
you
might
not
have
someone
get
excited
about
a
menu.
thought
to
get.
13. Mo(va(ng
Example
Aspect
Snippets
stylish decor
atmosphere
awesome art
loved it!
food
tasty calzones!
fast and friendly
service
impatient waiters
Importance
of
Context:
...
by
local
ar(sts.
{
Ordered
chicken
food
parm
and
loved
it!
Friend
had
the
veal.
The
service
was
...
13
14. Mul(-‐Aspect
Summariza(on
Content
Topic
Model
I
ordered
lunch
from
them
the
other
day
and
I
was
pleasantly
surprised.
Our
waiter
dazzled
me
with
his
blue
eyes
and
genuine
smile,
and
all
the
waiters
were
extremely
professional
and
efficient.
Sequence
Labeling
Task
I
ordered
lunch
from
them
the
other
day
and
I
was
[FOOD
pleasantly
surprised].
Our
waiter
dazzled
me
with
his
blue
eyes
and
genuine
smile,
and
all
the
waiters
were
[SERVICE
extremely
professional
and
efficient].
14
16. Approach
Overview
words
labels
Task
Labels:
Observed
I
had
the
shrimp
salad
and
was
[FOOD
pleasantly
surprised].
The
[ATMOSPHERE
decor
was
tasteful]
and
staff
was
[SERVICE
extremely
professional
and
efficient].
16
18. Approach
Overview
• Jointly
learn
structure
and
task
parameters
– Topics
are
latent
variables
shaped
by
task
• Principled
way
to
incorporate
unlabeled
data
– More
unlabeled
data,
beler
performance
18
28. Informa(on
Extrac(on
Goal:
Extract
phrases
from
review
text
in
pre-‐specified
categories
Input:
User-‐generated
review
text,
labeled
training
data
Output:
Labeled
phrases
in
each
category
FOOD
I
came
here
with
my
husband
for
the
tas(ng
menu,
and
we
SERVICE
were
not
disappointed.
We
got
to
sit
at
the
chef’s
table,
which
ATMOSPHERE
overlooked
the
kitchen.
The
service
was
polite
and
PRICE
knowledgeable,
the
atmosphere
was
elegant
and
energePc
OVERALL
and
the
food
was
wonderfully
creaPve
and
delicious.
28
29. Systems
• NoCM:
Just
the
CRF,
no
content
model
• IndepCM:
Es(mate
content
model
parameters
first,
then
use
them
in
the
CRF.
• JointCM:
Es(mate
content
and
CRF
parameters
jointly
using
EM
29
31. Impact
of
Unlabeled
Data
Setup:
Using
the
Amazon
corpus,
fix
the
amount
of
labeled
data,
vary
the
amount
of
unlabeled
data
50
47.3
47.8
44
41,5
38
0
6
300
12
600
Number
of
Unlabeled
Reviews
31
32. Mul(-‐Aspect
Sen(ment
Ranking
Task:
Predict
sen(ment
(1-‐10)
for
each
aspect
Aspect
Rating
picture
9.0
audio
9.5
extra
7.0
Approach:
• Same
objec(ve
as
summariza(on
• Different
E-‐
and
M-‐Steps
[See
paper]
32
35. Agree
to
Disagree
#1
The
fried
oysters
were
very
good
The
casish
tasted
dry
and
bland
and
boring
The
star
of
the
plate
was
the
grits
#2
The
gnocchi
with
mushrooms
was
outstanding
The
casish
approaches
perfec(on
The
shrimp
and
grits
are
nothing
less
than
spectacular
35
36. Review
Aggrega(on
• Hundreds
of
reviews
for
each
product
• Opinions
vary
widely
→ Need
to
aggregate
sta(s(cs
• Histograms
show
sen(ment
distribu(on,
but
it’s
not
enough
36
37. Aspect-‐based
Analysis
Prior
work:
Use
a
set
of
predefined
domain-‐specific
product
aspects
(e.g.,
Snyder
and
Barzilay
2007)
→
Coarse
level
analysis
37
38. Informa(ve
Aggrega(on
Useful
informa(on:
– What’s
the
best
dish
at
this
restaurant?
– What
do
people
dislike
about
this
restaurant?
– Which
dishes
do
people
disagree
about?
38
39. Informa(ve
Aggrega(on
Aggrega(on
of
product-‐specific
aspects
Japanese
Restaurant
We
had
a
great
Pme
last
Wow,
I
can’t
believe
I
have
such
mixed
things
night
at
this
restaurant.
how
much
this
place
has
t o
s a y
a b o u t
t h i s
T h e
s u s h i
w a s
s o
changed!
They
used
to
restaurant.
On
one
incredibly
fresh.
We
had
be
mediocre,
but
now
hand,
their
sushi
is
a
bad
experience
at
the
they
never
fail
to
amaze.
unquesPonably
the
best
b a r ,
t h o u g h .
M y
We
started
off
at
the
bar
in
the
city.
On
the
other,
chocolate
marPni
was
with
awesome
sake
the
atmosphere
isn’t
absolutely
terrible.
We
bombs.
When
we
got
to
that
great.
Plus,
their
will
be
back,
but
we’ll
our
table,
the
sushi
was
drinks
are
completely
skip
the
drinks.
fantasPc.
watered
down.
Sushi
100%
posiPve
Chicken
33%
posiPve
Relevant
aspects
User
sen(ment
39
40. Corpus-‐driven
Aspect
Defini(on
Define
aspects
dynamically
based
on
reviews
Japanese
Restaurant
Bakery
We
had
a
great
Pme
Wow,
I
can’t
believe
I
have
such
mixed
We
had
a
great
Pme
Wow,
I
can’t
believe
I
have
such
mixed
l a s t
n i g h t
a t
t h i s
how
much
this
place
things
to
say
about
this
l a s t
n i g h t
a t
t h i s
how
much
this
place
things
to
say
about
this
restaurant.
The
sushi
has
changed!
They
restaurant.
On
one
restaurant.
The
sushi
has
changed!
They
restaurant.
On
one
was
so
incredibly
fresh.
used
to
be
mediocre,
hand,
their
sushi
is
was
so
incredibly
fresh.
used
to
be
mediocre,
hand,
their
sushi
is
W e
h a d
a
b a d
but
now
they
never
fail
unquesPonably
the
W e
h a d
a
b a d
but
now
they
never
fail
unquesPonably
the
experience
at
the
bar,
to
amaze.
We
started
best
in
the
city.
On
the
experience
at
the
bar,
to
amaze.
We
started
best
in
the
city.
On
the
though.
My
chocolate
off
at
the
bar
with
other,
the
atmosphere
though.
My
chocolate
off
at
the
bar
with
other,
the
atmosphere
marPni
was
absolutely
awesome
sake
bombs.
isn’t
that
great.
Plus,
marPni
was
absolutely
awesome
sake
bombs.
isn’t
that
great.
Plus,
terrible.
We
will
be
When
we
got
to
our
t h e i r
d r i n k s
a r e
terrible.
We
will
be
When
we
got
to
our
t h e i r
d r i n k s
a r e
back,
but
we’ll
skip
the
table,
the
sushi
was
completely
watered
back,
but
we’ll
skip
the
table,
the
sushi
was
completely
watered
drinks.
fantasPc.
down.
drinks.
fantasPc.
down.
-‐ Sushi
-‐ Cookies
-‐ Sake
-‐ Cakes
-‐ Dessert
-‐ Pies
→
Aspects
specific
to
each
product
40
41. Corpus-‐driven
Aspect
Defini(on
Allows
comparison
across
mul(ple
reviews
Bakery
I
buy
all
of
my
baked
I
picked
up
a
birthday
This
place
is
nice
for
g o o d s
f r o m
t h i s
cake
for
my
son
here
some
baked
goods,
bakery.
Their
bread
is
yesterday.
It
was
the
but
some
things
are
so
delicious!
It’s
also
most
amazing
cake
really
nasty.
The
loaf
good
for
all
kinds
of
I’ve
ever
seen!
The
of
bread
I
bought
was
baked
goods.
They
d e c o r a P o n s
w e r e
stale!
They
were
also
have
some
truly
outstanding,
and
all
happy
to
take
it
back
beauPful
cakes
on
the
kids
loved
the
and
give
me
another,
display.
Even
their
chocolate
icing.
I’ll
but
I’ll
be
watching
cookies
are
great!
definitely
come
back!
next
Pme.
…truly
beauPful
cakes
on
display.
…most
amazing
cake
I’ve
ever
seen!
– Consensus
(both
posi(ve
and
nega(ve)
What’s
the
best/worst
aspect
of
this
product?
41
42. Corpus-‐driven
Aspect
Defini(on
Allows
comparison
across
mul(ple
reviews
Bakery
I
buy
all
of
my
baked
I
picked
up
a
birthday
This
place
is
nice
for
g o o d s
f r o m
t h i s
cake
for
my
son
here
some
baked
goods,
bakery.
Their
bread
is
yesterday.
It
was
the
but
some
things
are
so
delicious!
It’s
also
most
amazing
cake
really
nasty.
The
loaf
good
for
all
kinds
of
I’ve
ever
seen!
The
of
bread
I
bought
was
baked
goods.
They
d e c o r a P o n s
w e r e
stale!
They
were
also
have
some
truly
outstanding,
and
all
happy
to
take
it
back
beauPful
cakes
on
the
kids
loved
the
and
give
me
another,
display.
Even
their
chocolate
icing.
I’ll
but
I’ll
be
watching
cookies
are
great!
definitely
come
back!
next
Pme.
Their
bread
is
so
delicious!
The
loaf
of
bread
I
bought
was
stale!
– Consensus
(both
posi(ve
and
nega(ve)
What’s
the
best/worst
aspect
of
this
product?
– Conflicts
of
opinion
What
aspects
do
people
disagree
about?
42
43. Task:
Input
Input:
– Food-‐related
snippets
from
restaurant
reviews
• Concise
descrip(on
of
a
user’s
opinion
– Automa(cally
extracted
from
full
review
text
(Sauper
et
al.
2010)
We
went
to
the
restaurant,
and
the
sushi
was
incredibly
fresh.
– Segmented
by
restaurant,
but
no
addi(onal
annota(on
Japanese
Restaurant
Bakery
the
sushi
was
so
incredibly
fresh
I’d
recommend
the
apple
pie
best
chicken
katsu
in
town
the
bread
was
disappoinPngly
stale
drinks
are
fun,
fresh,
and
delicious
chocolate
torte
is
the
stuff
of
dreams
43
44. Task:
Output
Output:
– Relevant
aspects
for
each
restaurant
– Aspect
label
for
each
snippet
– Sen(ment
label
for
each
snippet
Mexican
Restaurant
Burrito
Salsa
+ they
had
a
decent
burrito
+ the
salsa
is
incredible
− the
burrito
was
mediocre
at
best
+ the
mango
salsa
is
perfectly
diced
− the
burrito
was
heavily
cilantroed
+ hola
free
chips
&
salsa
44
45. Possible
Solu(on
Use
clustering
based
on
lexical
similarity
the
marPnis
were
very
good
the
sushi
was
the
best
I’d
ever
had
the
marPnis
were
tasty
best
paella
I’d
ever
had
the
fillet
was
the
best
steak
we’d
ever
had
the
wine
list
was
pricey
it’s
the
best
soup
I’ve
ever
had
their
wine
selec(on
is
horrible
ParPal
output
of
state-‐of-‐the-‐art
clustering
system
Problem:
Clusters
and
aspects
are
not
aligned!
45
46. Our
Solu(on
• Jointly
model
aspect
and
sen(ment
• Leverage
data
to
dis(nguish
sen(ment
and
aspect
Bakery
Japanese
Review
1
pies
delicious
salmon
fantas(c
cookies
fresh
sake
smooth
Review
2
cakes
fantas(c
maki
beau(ful
pies
amazing
salmon
fresh
Review
3
cakes
beau(ful
maki
delicious
bread
stale
miso
bland
46
47. Model:
Overview
• Each
snippet
has
an
aspect
and
a
sen(ment
• Each
word
is
drawn
from
a
topic
distribu(on:
– Aspects
are
specific
to
a
single
product
pizza
dessert
pad
thai
– Sen(ment
is
global
across
all
products
great
horrible
amazing
– Background
distribu(on
is
global
was
our
food
• Transi(on
distribu(on
encodes
word
topic
transi(ons
They
had
wonderful
appePzers.
47
48. Model:
Genera(ve
Story
1. Global
distribu(ons
2. Restaurant-‐level
distribu(ons
3. Snippet-‐level
latent
structure
4. Words
48
49. Model:
Genera(ve
Story
Globally,
a. Background
distribu(on
word
distribu(on
for
stop
words
and
in-‐domain
white
noise
b. Sen(ment
distribu(ons
,
word
distribu(ons
over
posi(ve
and
nega(ve
sen(ment
words
small
bias
for
seed
words
c. Transi(on
distribu(on
first-‐order
Markov
distribu(on
of
word
topic
transi(ons
Background
Sen(ment
Transi(on
distribu(on
distribu(ons
distribu(on
B
+
-‐
Λ
49
50. Model:
Genera(ve
Story
For
each
restaurant
,
a. Aspect
distribu(ons
word
distribu(on
for
each
aspect
b. Aspect-‐sen(ment
binomials
probability
of
posi(ve
vs.
nega(ve
sen(ment
for
each
aspect
c. Aspect
mul(nomial
probability
of
each
aspect
Aspect
Aspect
distribu(ons
Aspect-‐sen(ment
binomials
mul(nomial
1
2
… K
φ1
φ2
… φK
ψ
50
51. Model:
Genera(ve
Story
For
each
snippet
from
restaurant
,
Aspect
a. Aspect
chosen
from
aspect
mul(nomial
ψ 2
Sen(ment
b. Sen(ment
chosen
from
aspect-‐sen(ment
binomial
φ2
+
c. Sequence
of
word
topics
Background,
Aspect,
or
Sen(ment
selected
from
transi(on
distribu(on
Word
topic
sequence
Λ
B
A
B
S
S
51
52. Model:
Genera(ve
Story
For
each
word
,
Aspect
a. Word
chosen
from
topic-‐specific
distribu(on
2
based
on
word
topic
sequence
Sen(ment
Word
topic
sequence
+
B
A
B
S
S
Background
B
The
pizza
was
really
great
52
54. Standard
Varia(onal
Inference
• Desired
posterior:
• Op(mizing
directly
is
intractable
• Instead,
op(mize
varia(onal
objec(ve
with
mean-‐field
factoriza(on:
s.t.
factorizes
54
55. Data
Set
Food-‐related
snippets
from
Yelp
restaurant
reviews
(Sauper
et
al.
2010)
– 13,879
total
snippets
– 328
restaurants
– 42.1
snippets
per
restaurant
(high
variance)
– 7.8
words
per
snippet
Seed
words
for
sen(ment
distribu(ons
– 42
posi(ve,
33
nega(ve
– Relevant
to
domain
(e.g.,
“delicious”)
55
56. Experiments:
Aspect
Clustering
• Gold
standard
– Clusters
over
3,250
snippets
– Collected
via
Mechanical
Turk
• Baseline
– CLUTO
clustering
weighted
by
TF*IDF
• MUC
cluster
evalua(on
metric
– Based
on
number
of
cluster
merges
and
splits
required
to
achieve
gold
data
• Both
systems
allowed
10
clusters
per
restaurant
56
57. Experiments:
Aspect
Clustering
MUC
F1
80
75,5
69,3
70
60
Baseline
Our
model
Our
model
Our
model
the
marPnis
are
very
good
the
carrot
cake
was
delicious
the
marPni
selec(on
looked
delicious
the
best
carrot
cake
I’ve
ever
eaten
the
s’mores
marPni
sounded
excellent
carrot
cake
was
deliciously
moist
the
marPnis
are
very
good
the
carrot
cake
was
delicious
the
mozzarella
was
very
fresh
it
was
rich,
creamy,
and
delicious
the
fish
and
various
meets
were
well
made
the
pasta
bolognese
was
rich
and
robust
Baseline
Baseline
57
58. Error
Analysis
Number
of
sen(ment
and
aspect
errors
approximately
equal
Aspect
errors
Sen(ment
errors
− Similar
aspect
words
in
different
− Rare
sen(ment
words
contexts
belgian
frites
are
very
crave-‐able
the
blackened
chicken
was
meh
chicken
enchiladas
are
yummy
− Nega(on,
some(mes
the
cream
cheese
wasn’t
bad
the
cream
cheese
was
n’t
bad
ice
cream
was
just
delicious
58
59. Paper
&
Code
• Paper
hlp://groups.csail.mit.edu/rbg/code/content_a†tude/sauper-‐acl-‐11.pdf
• Code
hlp://groups.csail.mit.edu/rbg/code/content_a†tude/code.tar.gz
59
60. The
Task
• Goal:
Automa(c
construc(on
of
even
records
from
Twiler
• Input:
Stream
of
Twiler
messages
Seated
at
@carnegiehall
wai'ng
for
@CraigyFerg’s
show
@DJPaulyD
absolutely
killed
it
at
Terminal
5
last
night.
Craig,
nice
seeing
you
#noelnight
this
weekend
@becksdavis!
• Output:
Table
of
event
records
Ar#st
Venue
Craig
Ferguson
Carnegie
Hall
DJ
Pauly
D
Terminal
5
60
61. Example
Output
Artist Venue
Bardavon Opera
Amos Lee
House
Jim Gaffigan Best Buy Theater
Jeff Tweedy Bowery Ballroom
Hall & Oates Beacon Theater
J. Cole Highline Ballroom
Sunday Gospel B.B. King Blues
Brunch Club
61
62. IE
for
Social
Media:
Challenges
• Messages
are
short
⇒ Individual
message
may
not
contain
all
event
fields.
• Message
are
expressed
in
colloquial
language
⇒ Mapping
between
messages
and
event
record
is
not
obvious
Seated
at
@carnegiehall
wai(ng
for
@CraigyFerg’s
show
Ar(st:
Craig
Ferguson
RT
@leerader
:
ge†ng
REALLY
Venue:
Carnegie
Hall
stoked
for
#CraigyAtCarnegie
sat
night.
62
63. IE
for
Social
Media:
Opportunity
Significant
redundancy
in
Twiler
stream:
Seated
at
@carnegiehall
wai'ng
for
@CraigyFerg’s
show
@DJPaulyD
absolutely
killed
it
at
Terminal
5
last
night.
Craig,
nice
seeing
you
#noelnight
this
weekend
@becksdavis!
Approach:
Drive
event
extrac(on
by
modeling
agreement
in
message
stream.
63
64. Model
Func(onality
• Message
level
analysis:
Tag
words
in
message
with
event-‐field
labels.
Label
(y)
ar'st
none
venue
venue
@YonderMountain
rocking
Mercury
Lounge
Message
(x)
64
65. Model
Func(onality
• Message
level
analysis:
Tag
words
in
message
with
event-‐field
labels.
• Message
clustering:
Group
messages
based
on
events.
• Event
records:
Induce
canonical
value
for
each
field.
Record
(R) Alignment
(A)
#CraigAtCarnie
is
star'ng
now!
#iamsoexcited
Ar#st
Venue
Craig
Ferguson
Carnegie
Hall
Going
to
see
Radiohead
at
the
Coliseum
tonight!
Craig
Ferguson,
what
a
riot!
Carnegie
is
in
s'tches
ArPst
Venue
Radiohead
Coliseum
Pumped
for
R
A
D
I
O
H
E
A
D
!!!
65
66. Model
Overview
Source
of
supervision:
Example
event
records
-‐
Alignment
between
records
and
messages
not
observed.
-‐
Message
level
field
annota(ons
not
observed.
July
16,
5:30pm
at
American
Folk
Art
Museum
Jun
17,
8:00
PM
at
Izod
Center
Jun
17,
8:00
PM
at
Tarrytown
Music
Hall
66