<s>
Multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
is	O
a	O
technology	O
for	O
traditional	O
text-based	O
sentiment	O
analysis	O
,	O
which	O
includes	O
modalities	B-General_Concept
such	O
as	O
audio	O
and	O
visual	O
data	O
.	O
</s>
<s>
It	O
can	O
be	O
bimodal	O
,	O
which	O
includes	O
different	O
combinations	O
of	O
two	O
modalities	B-General_Concept
,	O
or	O
trimodal	O
,	O
which	O
incorporates	O
three	O
modalities	B-General_Concept
.	O
</s>
<s>
With	O
the	O
extensive	O
amount	O
of	O
social	O
media	O
data	O
available	O
online	O
in	O
different	O
forms	O
such	O
as	O
videos	O
and	O
images	O
,	O
the	O
conventional	O
text-based	O
sentiment	O
analysis	O
has	O
evolved	O
into	O
more	O
complex	O
models	O
of	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
,	O
which	O
can	O
be	O
applied	O
in	O
the	O
development	O
of	O
virtual	B-Protocol
assistants	I-Protocol
,	O
analysis	O
of	O
YouTube	O
movie	O
reviews	O
,	O
analysis	O
of	O
news	O
videos	O
,	O
and	O
emotion	O
recognition	O
(	O
sometimes	O
known	O
as	O
emotion	O
detection	O
)	O
such	O
as	O
depression	O
monitoring	O
,	O
among	O
others	O
.	O
</s>
<s>
Similar	O
to	O
the	O
traditional	O
sentiment	O
analysis	O
,	O
one	O
of	O
the	O
most	O
basic	O
task	O
in	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
is	O
sentiment	O
classification	O
,	O
which	O
classifies	O
different	O
sentiments	O
into	O
categories	O
such	O
as	O
positive	O
,	O
negative	O
,	O
or	O
neutral	O
.	O
</s>
<s>
Feature	B-General_Concept
engineering	I-General_Concept
,	O
which	O
involves	O
the	O
selection	O
of	O
features	O
that	O
are	O
fed	O
into	O
machine	O
learning	O
algorithms	O
,	O
plays	O
a	O
key	O
role	O
in	O
the	O
sentiment	O
classification	O
performance	O
.	O
</s>
<s>
In	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
,	O
a	O
combination	O
of	O
different	O
textual	O
,	O
audio	O
,	O
and	O
visual	O
features	O
are	O
employed	O
.	O
</s>
<s>
Similar	O
to	O
the	O
conventional	O
text-based	O
sentiment	O
analysis	O
,	O
some	O
of	O
the	O
most	O
commonly	O
used	O
textual	O
features	O
in	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
are	O
unigrams	B-Language
and	O
n-grams	B-Language
,	O
which	O
are	O
basically	O
a	O
sequence	O
of	O
words	O
in	O
a	O
given	O
textual	O
document	O
.	O
</s>
<s>
These	O
features	O
are	O
applied	O
using	O
bag-of-words	B-General_Concept
or	O
bag-of-concepts	O
feature	O
representations	O
,	O
in	O
which	O
words	O
or	O
concepts	O
are	O
represented	O
as	O
vectors	O
in	O
a	O
suitable	O
space	O
.	O
</s>
<s>
Some	O
of	O
the	O
most	O
important	O
audio	O
features	O
employed	O
in	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
are	O
mel-frequency	B-Algorithm
cepstrum	I-Algorithm
(	O
MFCC	O
)	O
,	O
spectral	B-Algorithm
centroid	I-Algorithm
,	O
spectral	B-Algorithm
flux	I-Algorithm
,	O
beat	O
histogram	O
,	O
beat	O
sum	O
,	O
strongest	O
beat	O
,	O
pause	O
duration	O
,	O
and	O
pitch	O
.	O
</s>
<s>
OpenSMILE	B-Application
and	O
Praat	B-Language
are	O
popular	O
open-source	O
toolkits	O
for	O
extracting	O
such	O
audio	O
features	O
.	O
</s>
<s>
Specifically	O
,	O
smile	O
,	O
is	O
considered	O
to	O
be	O
one	O
of	O
the	O
most	O
predictive	O
visual	O
cues	O
in	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
.	O
</s>
<s>
Unlike	O
the	O
traditional	O
text-based	O
sentiment	O
analysis	O
,	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
undergo	O
a	O
fusion	O
process	O
in	O
which	O
data	O
from	O
different	O
modalities	B-General_Concept
(	O
text	O
,	O
audio	O
,	O
or	O
visual	O
)	O
are	O
fused	O
and	O
analyzed	O
together	O
.	O
</s>
<s>
The	O
existing	O
approaches	O
in	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
data	B-General_Concept
fusion	I-General_Concept
can	O
be	O
grouped	O
into	O
three	O
main	O
categories	O
:	O
feature-level	O
,	O
decision-level	O
,	O
and	O
hybrid	O
fusion	O
,	O
and	O
the	O
performance	O
of	O
the	O
sentiment	O
classification	O
depends	O
on	O
which	O
type	O
of	O
fusion	O
technique	O
is	O
employed	O
.	O
</s>
<s>
Feature-level	O
fusion	O
(	O
sometimes	O
known	O
as	O
early	O
fusion	O
)	O
gathers	O
all	O
the	O
features	O
from	O
each	O
modality	B-General_Concept
(	O
text	O
,	O
audio	O
,	O
or	O
visual	O
)	O
and	O
joins	O
them	O
together	O
into	O
a	O
single	O
feature	O
vector	O
,	O
which	O
is	O
eventually	O
fed	O
into	O
a	O
classification	O
algorithm	O
.	O
</s>
<s>
Decision-level	O
fusion	O
(	O
sometimes	O
known	O
as	O
late	O
fusion	O
)	O
,	O
feeds	O
data	O
from	O
each	O
modality	B-General_Concept
(	O
text	O
,	O
audio	O
,	O
or	O
visual	O
)	O
independently	O
into	O
its	O
own	O
classification	O
algorithm	O
,	O
and	O
obtains	O
the	O
final	O
sentiment	O
classification	O
results	O
by	O
fusing	O
each	O
result	O
into	O
a	O
single	O
decision	O
vector	O
.	O
</s>
<s>
One	O
of	O
the	O
advantages	O
of	O
this	O
fusion	O
technique	O
is	O
that	O
it	O
eliminates	O
the	O
need	O
to	O
fuse	O
heterogeneous	O
data	O
,	O
and	O
each	O
modality	B-General_Concept
can	O
utilize	O
its	O
most	O
appropriate	O
classification	O
algorithm	O
.	O
</s>
<s>
It	O
usually	O
involves	O
a	O
two-step	O
procedure	O
wherein	O
feature-level	O
fusion	O
is	O
initially	O
performed	O
between	O
two	O
modalities	B-General_Concept
,	O
and	O
decision-level	O
fusion	O
is	O
then	O
applied	O
as	O
a	O
second	O
step	O
,	O
to	O
fuse	O
the	O
initial	O
results	O
from	O
the	O
feature-level	O
fusion	O
,	O
with	O
the	O
remaining	O
modality	B-General_Concept
.	O
</s>
<s>
Similar	O
to	O
text-based	O
sentiment	O
analysis	O
,	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
can	O
be	O
applied	O
in	O
the	O
development	O
of	O
different	O
forms	O
of	O
recommender	B-Application
systems	I-Application
such	O
as	O
in	O
the	O
analysis	O
of	O
user-generated	O
videos	O
of	O
movie	O
reviews	O
and	O
general	O
product	O
reviews	O
,	O
to	O
predict	O
the	O
sentiments	O
of	O
customers	O
,	O
and	O
subsequently	O
create	O
product	O
or	O
service	O
recommendations	O
.	O
</s>
<s>
Multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
also	O
plays	O
an	O
important	O
role	O
in	O
the	O
advancement	O
of	O
virtual	B-Protocol
assistants	I-Protocol
through	O
the	O
application	O
of	O
natural	B-Language
language	I-Language
processing	I-Language
(	O
NLP	B-Language
)	O
and	O
machine	O
learning	O
techniques	O
.	O
</s>
<s>
In	O
the	O
healthcare	O
domain	O
,	O
multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
can	O
be	O
utilized	O
to	O
detect	O
certain	O
medical	O
conditions	O
such	O
as	O
stress	O
,	O
anxiety	O
,	O
or	O
depression	O
.	O
</s>
<s>
Multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
can	O
also	O
be	O
applied	O
in	O
understanding	O
the	O
sentiments	O
contained	O
in	O
video	O
news	O
programs	O
,	O
which	O
is	O
considered	O
as	O
a	O
complicated	O
and	O
challenging	O
domain	O
,	O
as	O
sentiments	O
expressed	O
by	O
reporters	O
tend	O
to	O
be	O
less	O
obvious	O
or	O
neutral	O
.	O
</s>
