Inter
national
J
our
nal
of
Electrical
and
Computer
Engineering
(IJECE)
V
ol.
14,
No.
6,
December
2024,
pp.
7137
∼
7144
ISSN:
2088-8708,
DOI:
10.11591/ijece.v14i6.pp7137-7144
❒
7137
A
semantic
similarity
sear
ch
engine
f
or
mo
vies
Ahmad
Mustafa,
Hammam
Mheidat,
Adam
Shatnawi
Department
of
Computer
Information
Systems,
Jordan
Uni
v
ersity
of
Science
and
T
echnology
,
Irbid,
Jordan
Article
Inf
o
Article
history:
Recei
v
ed
Feb
27,
2024
Re
vised
Aug
6,
2024
Accepted
Aug
14,
2024
K
eyw
ords:
Deep
learning
Machine
learning
Natural
language
processing
Recommender
system
Search
engine
Semantic
similarity
ABSTRA
CT
Semantic
similarity
has
been
g
aining
traction
in
the
eld
of
natural
language
processing.
It
is
a
measure
of
ho
w
similar
tw
o
pieces
of
te
xt
are
in
terms
of
their
meani
ng.
It
can
be
used
to
impro
v
e
search
engine
results.
W
e
propose
a
deep
learning-based
approach
to
b
uild
a
semantic
similarity
search
engine
for
mo
vies
based
on
a
mo
vie
summary
.
Filmmak
ers
can
g
ain
insight
into
audi-
ence
preferences
and
trends,
allo
wing
them
to
create
more
eng
aging
and
suc-
cessful
lms.
The
dataset
used
in
this
study
w
as
g
athered
from
internet
mo
vie
database
(IMDb),
it
includes
mo
vie
summaries
along
with
their
corresponding
name
mo
vies.
The
test
dataset
w
as
generated
using
ChatGPT
to
be
v
ery
close
to
human
input.
The
uni
v
ersal
sentence
encoder
(USE)
model
presented
promising
results
in
semantic
similarity
,
the
model
results
sho
w
that
for
the
top
5
simi-
lar
mo
vies,
the
model
returned
176
out
of
300
mo
vies
(58.6%).
F
or
the
top
10
similar
mo
vies,
the
model
returned
211
out
of
300
mo
vies
(70.3%).
Addition-
ally
,
for
the
top
15
similar
mo
vies,
the
model
returned
229
out
of
300
mo
vies
(76.3%).
And,
for
the
top
20
similar
mo
vies,
the
model
returned
249
of
300
mo
vies
(83%).
This
method
c
an
be
applied
to
mo
vie
recommendation
systems
or
to
or
g
anize
lms
in
a
collection
automatically
.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
Ahmad
Mustaf
a
Department
of
Computer
Information
Systems,
Jordan
Uni
v
ersity
of
Science
and
T
echnology
Irbid,
Jordan
Email:
ammustaf
a@just.edu.jo
1.
INTR
ODUCTION
Recommendation
systems
play
an
essential
role
in
recommending
items,
particularly
in
stream
ing
services.
Streaming
mo
vie
services
such
as
Netix
rely
on
recommendation
systems
to
assist
consumers
in
selecting
ne
w
mo
vies
to
w
atch.
These
systems
tak
e
into
account
se
v
eral
aspects
including
re
vie
ws,
cast,
nar
-
rati
v
e,
cre
w
,
genre,
and
popularity
to
ensure
optimal
outcomes.
It
tak
es
a
user
preference
into
consideration
and
hence
gi
v
es
results
depending
on
the
user’
s
particular
taste
and
decision.
Users
demand
lightning-f
ast,
ef
cient,
and
accurate
results.
As
a
result,
a
search
engine
system
for
mo
vies
has
become
an
important
service
for
the
audience
in
retrie
ving
a
mo
vie
name
without
remembering
its
name.
Currently
,
there
are
v
arious
types
of
recommender
systems
a
v
ailable
[1],
[2].
Natural
language
utterances
are
typically
quite
ambiguous
due
to
polysemous
w
ords,
which
ha
v
e
multiple
possible
meanings,
or
malapropism,
which
is
the
confusion
of
an
intended
w
ord
with
a
w
ord
that
sounds
or
is
spelled
similarly
b
ut
has
a
v
ery
dif
ferent
and
malapropos
meaning.
Generally
,
the
only
w
ay
to
determine
the
interpretation
of
an
utterance
is
to
consider
the
conte
xt
in
which
it
occurred.
Algorithms
and
programs,
ho
we
v
er
,
lack
the
adv
antage
of
human
linguistic
e
xpertise
[3].
An
interaction
model
between
humans
and
computers
is
essential
for
e
xtracting
the
title
of
a
lm
from
a
user’
s
query
.
This
article
introduces
a
search
engine
that
can
accurately
determine
the
title
of
a
mo
vie
J
ournal
homepage:
http://ijece
.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
7138
❒
ISSN:
2088-8708
using
only
a
brief
and
succinct
description.
T
o
handle
this
problem,
we
prepare
a
model
that
understands
and
interprets
the
human
language
for
the
computer
,
using
uni
v
ers
al
sentence
encoder
(USE)
[4],
which
is
a
pre-
trained
deep
learning
model
de
v
eloped
by
Google.
It
is
designed
to
con
v
ert
natural
language
te
xt
into
a
high-
dimensional
continuous
v
ector
representation,
also
kno
wn
as
an
embedding
[5].
These
embeddings
can
then
be
used
as
input
for
a
v
ariety
of
natural
language
processing
tasks,
such
as
te
xt
classication,
semantic
similarity
,
and
machine
translation.
The
USE
is
trained
on
a
di
v
erse
range
of
t
e
xt
data
and
can
generate
embeddings
for
a
wide
v
ariety
of
sentence
types
and
styles.
It
can
also
handle
input
te
xt
in
multiple
languages.
USE
can
be
used
to
encode
multiple
sentences
at
once
and
return
a
matrix,
which
can
be
used
in
tasks
such
as
te
xt
clustering,
retrie
v
al
and
so
on.
USE
and
sentence
bidirectional
encoder
representations
from
transformers
(SBER
T)
[6]
can
be
used
to
compute
the
semantic
similarity
of
sentences.
Ho
we
v
er
,
based
on
the
results
of
our
test
dataset,
it
is
e
vident
that
USE
is
more
suitable
for
this
particular
task.
USE
sentence
embedding
design,
which
is
pre-trained
on
a
v
ast
and
v
aried
te
xt
dataset,
may
enhance
its
capacity
to
capture
the
entire
te
xtual
meaning.
Ho
we
v
er
,
sen-
tence
BER
T
(S
BER
T)
is
a
specialized
model
designed
specically
for
generating
sentence
embeddings
[7].
It
e
xhibits
strong
generalization
abilities
across
a
wide
range
of
natural
language
processing
(NLP)
applications.
It
surpasses
USE
in
numerous
w
ays.
While
both
models
possess
their
o
wn
merits,
it
is
essential
to
conduct
a
comparison
in
order
to
determine
the
most
suitable
model
for
a
specic
NLP
task.
The
sections
of
the
paper
are
or
g
anized
as
follo
ws.
Sections
2,
3,
and
4.
discuss
the
related
w
ork,
methodology
,
e
xperiments
and
results,
respecti
v
ely
.
The
results
of
the
proposed
methods
are
discussed
in
section
5.
Section
6
concludes
the
w
ork.
2.
RELA
TED
W
ORK
W
e
start
by
looking
at
the
f
ascination
of
the
scientic
literature
with
mo
vie
classication
or
recom-
mendation
system.
The
approaches
can
be
di
vided
into
classical
machine
learning
[8]–[11]
and
deep
learn-
ing
[12]–[14].
Recent
approaches
suggested
multimodal
data[15].
Man
y
ha
v
e
used
collaborati
v
e
ltering
[16],
[17].
Azaria
et
al.
[18]
proposed
that
the
recommendation
system
designer
should
consider
b
usiness
re
v
enue.
Chinnapottu
et
al.
[19]
ha
v
e
applied
v
arious
machine
l
earning
models
(such
as
logistic
re
gression,
random
forest,
nai
v
e
Bayes,
and
support
v
ector
machine)
to
predict
the
name
of
a
mo
vie
using
input
te
xt
that
has
a
semantic
similarity
to
the
mo
vie’
s
script.
The
y
used
a
dataset
of
11
mo
vie
scripts
scraped
from
the
internet
mo
vie
script
database
(IMSDb)
website
and
preprocessed
using
v
arious
natural
language
processing
techniques,
including
stop
w
ord
remo
v
al,
punctuation
remo
v
al,
tok
enization,
stemming,
and
v
ectorization.
The
results
of
the
study
sho
wed
that
the
nai
v
e
Bayes
classier
and
support
v
ector
machine
models
performed
better
than
the
logi
stic
re
gression
and
random
forest
models
in
predicting
the
name
of
the
mo
vie.
It
is
w
orth
noting
that
the
small
size
of
the
training
data
(11
mo
vies’
scripts)
and
testing
data
(166
paraphrased
sentences
from
the
mo
vies’
scripts)
may
ha
v
e
limited
the
generalizability
of
the
results.
Y
oo
et
al.
[20]
ha
v
e
de
v
eloped
a
model
that
can
accurately
identify
a
mo
vie
when
searching
with
a
sequence
of
w
ords.
The
y
proposed
a
model
that
consists
of
three
parts:
one
for
e
xtracting
w
ord
embeddings
using
a
skip-grams
based
approach,
one
for
part-of-speech
tagging
using
hidden
mark
o
v
models
and
V
iterbi
algorithms,
and
one
for
named
entity
recognition
using
a
bidirectional
long
short-term
memory
(Bi-LSTM)
approach.
F
or
model
e
v
aluation,
50
participants
were
randomly
selected
and
ask
ed
to
choose
a
title
from
a
list
of
30
mo
vies.
The
results
sho
wed
that
the
proposed
model
outperformed
e
xisting
search
systems
such
as
Na
v
er
and
Google.
It
w
as
able
to
accurately
identify
the
correct
mo
vie
when
searching
with
a
sequence
of
w
ords.
Bhattacharya
and
Lundia
[21]
ha
v
e
crafted
a
web
application
that
assists
users
in
disco
v
ering
mo
vies
that
align
with
their
preferences.
The
application
generates
a
list
of
highly
similar
mo
vies
based
on
inputs
from
the
user
,
based
on
some
characteristics
of
the
mo
vie
such
as
the
summary
plot,
actors,
and
director
.
T
o
accomplish
this,
the
team
utilized
IMDb
data
and
created
a
representation
for
each
mo
vie
in
the
form
of
a
“bag
of
w
ords”
that
encapsulated
the
plot
summary
,
actors,
and
directors.
The
k
e
y
to
the
application’
s
ability
to
suggest
similar
mo
vies
w
as
the
use
of
cosine
similarity
.
This
mathematical
technique
w
as
applie
d
to
measure
the
similarity
between
the
user’
s
input
and
the
w
ords
that
represent
mo
vies,
thereby
enabling
the
identication
of
the
mo
vies
that
were
most
similar
the
user’
s
input.
Ro
y
et
al.
[22]
ha
v
e
emplo
yed
a
semi-supervised
learning
method
to
de
v
elop
a
mo
vie
recommendation
system.
The
authors
proposed
three
approaches;
popularity-based
Recommender
system,
simple
recommender
Int
J
Elec
&
Comp
Eng,
V
ol.
14,
No.
6,
December
2024:
7137-7144
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
❒
7139
system,
and
content-based
ltering
approach.
Besides
highlighting
the
adv
antages
and
disadv
antages
of
each
approach,
the
study
al
so
demonstrated
ho
w
one
approach
w
as
m
itig
ated
by
another
.
The
project
w
as
b
uilt
us-
ing
Django,
a
popular
web
application
de
v
elopment
technology
,
which
enabled
the
creation
of
a
user
-friendly
interf
ace
for
the
mo
vie
recommendation
system.
The
approaches
were
tested
and
e
v
aluated
on
the
well-kno
wn
mo
vie
lens
dataset,
which
contains
a
lar
ge
collection
of
mo
vie
ratings
and
re
vie
ws
from
users.
The
results
sho
wed
that
the
combination
of
the
three
approaches
resulted
in
more
accurate
and
personalized
recommenda-
tions
for
users.
W
alek
and
F
ojtik
[23]
proposed
a
h
ybrid
system
called
predictory
that
combines
a
collaborati
v
e
l
ter
-
ing
system,
a
content-based
system,
and
a
fuzzy
e
xpert
system
to
recommend
mo
vies
to
users.
The
collaborati
v
e
ltering
system
uses
singular
v
alue
decomposition
(SVD)
algorithm
to
predict
the
e
v
aluation
of
mo
vies.
The
content-based
system
uses
the
properties
of
the
mo
vies,
such
as
the
genre
and
plot,
to
mak
e
recommendations.
The
fuzzy
e
xpert
system
e
v
aluates
the
le
v
el
of
importance
of
mo
vies
for
the
nal
list
of
recommended
mo
vies
based
on
v
arious
parameters,
such
as
the
a
v
erage
mo
vie
rating,
the
number
of
mo
vie
ratings,
and
the
le
v
el
of
similarity
with
already
rated
mo
vies.
The
system
w
as
v
alidated
on
a
group
of
users
using
the
Mo
vie
Lens
dataset
and
sho
wed
promising
result
s.
The
results
were
e
v
aluated
using
standard
metrics
such
as
precision,
recall,
and
F1-measure,
with
v
alues
of
81%,
83%,
and
82%
respecti
v
ely
.
The
system
w
as
also
compared
with
other
typical
approaches,
such
as
a
pure
content-based
system,
a
pure
collaborati
v
e
lter
ing
system,
and
a
weighted
h
ybrid
system.
The
results
sho
wed
that
the
proposed
h
ybrid
system
using
an
e
xpert
system
achie
v
ed
the
highest
ratio
of
rele
v
ant
mo
vies
mark
ed
by
the
users
during
testing.
This
means
the
system
w
as
able
to
recommend
the
most
rele
v
ant
mo
vies
to
the
users.
The
mo
vie
recommender
system
proposed
by
Ahuja
et
al.
[9]
is
designed
to
mak
e
personalized
mo
vie
recommendations
to
users
by
clustering
similar
mo
vies
together
and
using
K-nearest
neighbor
algorithm
to
recommend
mo
vies
that
are
lik
e
the
ones
a
user
has
pre
viously
lik
ed.
The
system
utilizes
k-means
clustering
algorithm
to
group
similar
mo
vies
together
based
on
their
attrib
utes
such
as
genre,
director
,
and
cast.
This
clustering
is
done
to
mak
e
the
mo
vie
recommendation
more
accurate
by
recommending
mo
vies
that
are
lik
e
the
mo
vies
a
user
has
pre
viously
lik
ed.
The
data
for
this
system
is
tak
en
from
the
mo
vie
lens
dataset,
which
is
a
widely
used
dataset
in
mo
vie
recommendation
systems
research.
This
dataset
contains
information
on
mo
vies
and
the
ratings
gi
v
en
to
them
by
users.
The
system
is
implemented
in
Python
programming
language,
which
is
a
popular
choice
for
machine
learning
projects
due
to
its
ease
of
use
and
man
y
libraries
a
v
ailable
for
machine
learning
tasks.
The
system
demonstrates
impro
v
ed
root
mean
squared
error
(RMSE)
compared
to
e
xisting
techniques.
RMSE
is
a
measure
of
ho
w
well
the
recommendations
of
a
system
match
the
actual
ratings
gi
v
en
by
users.
Y
e
et
al.
[24]
introduced
a
solution
to
Xbox
One’
s
mo
vie
search
job
that
deri
v
es
granular
rele
v
ance
of
training
samples
from
session
dat
a.
Se
v
eral
user
-eng
agement
measures
ha
v
e
sho
wn
that
the
strate
gy
produces
impro
v
ed
test
results.
When
the
score
is
computed,
it
also
gi
v
es
a
lar
ge
performance
boost.
Search
engines
rely
signicantly
on
their
ability
to
interpret
the
intent
and
conte
xt
of
a
user’
s
query
.
It
is
challenging
to
create
a
search
engine
that
recognizes
these
phrases
and
pages,
especially
if
the
query
is
brief
and
v
ague.
T
o
solv
e
this
issue,
se
v
eral
models
ha
v
e
been
presented;
belo
w
are
some
of
the
greatest
e
xamples
from
the
lit
erature.
T
opic
inde
xing
nonlinear
models
can
be
used
to
the
demands
of
semantic
matching
for
search
engines.
Examples
of
nonlinear
models
include
probabilistic
latent
analysis
(PLSI)
and
latent
dirichlet
allocation
(LD
A).
One
important
application
of
these
models
is
to
analyze
t
he
similarity
between
query
documents
at
the
topic
and
semantic
le
v
el.
Another
set
of
latent
semantic
models
are
neural
netw
ork
models,
which
can
be
used
to
calculate
the
query-document
similarity
.
3.
METHOD
3.1.
T
raining
dataset
T
o
obtain
a
comprehensi
v
e
dataset
of
mo
vies,
we
scra
p
e
the
dataset
from
IMDb
website
using
Beau-
tifulSoup4
.
T
o
ensure
that
the
dataset
is
as
specic
and
rele
v
ant
as
possible,
we
applied
a
series
of
lters
to
the
website’
s
adv
anced
search.
These
lters
include
a
release
date
range
of
between
January
1
st
1980
and
January
1
st
2023,
a
minimum
rating
count
of
10,000,
and
a
requirement
that
the
m
o
vi
es
be
in
English.
Additionally
,
we
set
the
“including
adults
titl
e”
lter
to
“true”
to
ensure
that
the
dataset
w
ould
include
a
di
v
erse
range
of
mo
vie
titles.
As
a
result
of
these
lters,
we
collected
7,653
mo
vies
that
met
our
criteria.
This
dataset
includes
9
features
such
as
the
mo
vie’
s
title,
director
,
release
year
,
summary
,
duration,
rating,
v
ote
count,
gross
re
v
enue
A
semantic
similarity
sear
c
h
engine
for
mo
vies
(Ahmad
Mustafa)
Evaluation Warning : The document was created with Spire.PDF for Python.
7140
❒
ISSN:
2088-8708
in
millions
of
dollars,
and
genre.
F
or
a
visual
representation
of
the
distrib
ution
of
mo
vies
in
this
dataset
o
v
er
the
years,
please
refer
to
Figure
1.
Figure
1.
The
distrib
ution
of
mo
vies
o
v
er
years
3.2.
T
esting
dataset
Gi
v
en
the
limited
a
v
ailability
of
datasets
in
the
eld
of
natural
language
processing,
we
used
Chat-
GPT
to
generate
a
dataset
for
our
research.
ChatGPT
is
a
lar
ge
language
model
de
v
eloped
by
OpenAI,
which
is
based
on
the
generati
v
e
pre-training
transformer
(GPT)
architecture.
It
is
trained
using
a
v
ast
dataset
of
te
xt,
allo
wing
it
to
produce
human-lik
e
responses
to
te
xt-based
prompts.
This
v
ersatile
model
can
be
ne-tuned
for
v
arious
natural
language
processing
tasks
such
as
language
trans
lation,
te
xt
summarization,
te
xt
completion,
and
question
answering,
as
well
as
more
creati
v
e
endea
v
ors
l
ik
e
writing
poetry
and
other
forms
of
te
xt.
Chat-
GPT
is
designed
to
be
user
-friendly
,
with
a
simple
API
that
can
be
inte
grated
into
v
arious
applications
such
as
chatbots,
v
oice
assistants,
and
language
understanding
systems.
T
o
simulate
human
input
(query),
we
used
a
prompt
that
ask
ed
ChatGPT
to
“generate
a
description
about
a
mo
vie
without
mentioning
the
name
of
the
mo
vie
or
the
names
of
actors
and
characters
in
one
sentence.
”
This
generated
400
descriptions.
These
descrip-
tions
represents
the
user
queries.
The
dataset
contains
tw
o
columns
the
mo
vie
description
(query)
and
the
tar
get
mo
vie
name.
Examples
of
mo
vies
are
sho
wn
in
T
able
1.
T
able
1.
Examples
of
ChatGPT
generated
descriptions
for
the
mo
vies:
baby
dri
v
er
,
American
History
X,
and
The
Lion
King
The
description
generated
by
ChatGPT
Mo
vie
name
A
young,
talented
geta
w
ay
dri
v
er
tri
es
to
lea
v
e
his
criminal
past
behind
and
start
a
ne
w
life
b
ut
is
pulled
back
into
the
w
orld
of
crime
by
a
group
of
dangerous
bank
robbers.
Baby
Dri
v
er
A
young
man
who
becomes
in
v
olv
ed
in
white
supremacist
acti
vity
as
a
result
of
a
tragic
f
amily
e
v
ent
is
sent
to
prison,
where
he
under
goes
a
transformation
and
w
orks
to
pre
v
ent
his
younger
brother
from
follo
wing
in
his
footsteps.
American
History
X
A
young
lion
prince
is
e
xiled
from
his
kingdom
after
the
death
of
his
f
ather
,
b
ut
e
v
entually
returns
to
reclaim
his
place
as
rightful
king
and
end
the
reign
of
his
po
wer
-hungry
uncle.
The
Lion
King
3.3.
Pr
epr
ocessing
F
or
the
training
dataset
(which
is
collected
from
IMDB),
we
drop
unnecessary
columns
(i.e.,
“direc-
tor
,
”
“release
year
,
”
“duration,
”
“rating,
”
“v
ote
count,
”
“gross,
”
and
“genre”)
because
we
aim
to
train
the
model
with
the
mo
vie
title
and
the
summary
.
Preprocessing
is
not
typically
necessary
when
using
the
USE
because
it
is
a
pre-trained
model
that
is
able
to
handle
a
wide
v
ariety
of
input
te
xt,
including
te
xt
with
dif
ferent
capitaliza-
tion,
punctuation,
and
grammatical
st
ructures.
Additionally
,
the
USE
is
designed
to
handle
input
te
xt
that
may
contain
typos,
misspellings,
or
other
errors.
Therefore,
it
can
produce
useful
e
mbeddings
e
v
en
when
applied
to
ra
w
,
unprocessed
te
xt.
F
or
the
testing
dataset
after
we
look
for
the
data,
we
found
descriptions
of
mo
vies
such
that
their
mo
vie
name
is
in
the
descriptions.
So,
after
remo
ving
them
the
dataset
becomes
300
mo
vies.
Figure
2
sho
ws
the
methodology
applied
in
our
research.
Int
J
Elec
&
Comp
Eng,
V
ol.
14,
No.
6,
December
2024:
7137-7144
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
❒
7141
Figure
2.
The
methodology
3.4.
Ev
aluation
It
w
as
not
possible
to
use
mean
reciprocal
rank
(MRR)
[25]
because
all
the
results
retrie
v
ed
for
mo
vie
titles
were
rele
v
ant
to
the
query
.
The
same
applies
to
mean
a
v
erage
precision
(MAP)
[26].
Based
on
the
abo
v
e,
we
were
only
looking
at
whether
the
requested
mo
vie
title
w
as
on
the
retrie
v
ed
list,
re
g
ardless
of
its
order
among
the
remaining
mo
vies.
W
e
di
vided
the
number
of
queries
returning
the
intended
mo
vie
title
by
the
total
number
of
queries
from
the
list
of
titles
retrie
v
ed.
Therefore,
our
v
alue
dif
fered
according
to
the
number
of
lms
in
the
retrie
v
ed
list.
Accur
acy
=
Q
T
r
ue
N
(1)
where
N
is
the
total
number
of
queries
and
Q
T
r
ue
is
the
number
of
queries
that
returned
the
correct
mo
vie
among
the
top
k
.
4.
RESUL
TS
The
results
from
using
the
USE
model
for
nding
the
most
similar
mo
vies
are
as
follo
ws:
for
the
top
5
similar
mo
vies,
the
model
returned
176
out
of
300
mo
vies
(58.6%).
F
or
the
top
10
similar
mo
vies,
the
model
returned
211
out
of
300
mo
vies
(70.3%).
Additionally
,
for
the
top
15
similar
mo
vies,
the
model
returned
229
out
of
300
mo
vies
(76.3%).
And,
f
or
the
top
20
similar
mo
vies,
the
model
returned
249
out
of
300
mo
vies
(83%).
F
or
further
details,
please
refer
to
T
able
2.
In
T
able
3
we
sho
w
the
accurac
y
reported
by
tw
o
selected
related
w
ork
[19],
[23].
Both
approaches
were
applied
to
mo
vie
semantic
search.
Ho
we
v
er
,
the
y
search
using
mo
vie
summa
ry
and
scenario.
In
[19].
The
original
mo
vie
name
is
predicted
by
using
random
paraphrased
sentences
from
mo
vie
scripts.
On
the
other
hand,
in
[23]
mo
vies
are
recommended
by
e
v
alua
ting
their
importance.
So,
the
testing
dataset
is
dif
ferent
than
ours.
T
able
2.
The
results
we
obtained
from
using
USE
T
op
5
T
op
10
T
op
15
T
op
20
Number
of
mo
vies
returned
176
211
229
249
Accurac
y
58.6%
70.3%
76.3%
83%
T
able
3.
The
accurac
y
obtained
by
the
related
w
ork
Approach
Accurac
y
[19]
59.63%
[23]
81%
Our
W
ork
83%
A
semantic
similarity
sear
c
h
engine
for
mo
vies
(Ahmad
Mustafa)
Evaluation Warning : The document was created with Spire.PDF for Python.
7142
❒
ISSN:
2088-8708
W
e
also
used
the
test
dataset
to
e
v
aluate
ChatGPT’
s
performance
as
a
mo
vie
search
engine
for
the
T
op
K
mo
vies.
T
o
compare
with
our
presented
method
using
the
USE
model,
we
also
used
dif
ferent
models
from
SBER
T
[27]
(namely
,
all-mpnet-base-v2,
msmarco-distilbert-base-v2,
multi-qa-mpnet-base-cos-v1,
and
paraphrase-distilroberta-base-v1).
Although
the
SBER
T
model
outperforms
the
USE
m
odel
in
v
arious
te
xtual
similarity
tasks,
USE
is
not
our
contrib
ution,
and
the
results
sho
wn
in
T
able
4
illustrate
that
USE
outperforms
SBER
T
in
our
case.
T
able
4.
The
accurac
y
obtained
by
our
model
vs.
other
models
on
the
same
testing
data
Our
Model
Chat
GPT
multi-qa-
mpnet-base-
cos-v1
msmarco-
distilbert-
base-v2
all-mpnet-
base-v2
paraphrase-
distilroberta-
base-v1
T
op5
Retrie
v
ed
mo
vies
176
157
185
155
162
164
Accurac
y
(%)
58.6
52.3
61.7
51.7
54
54.7
T
op10
Retrie
v
ed
mo
vies
211
173
203
178
184
186
Accurac
y
(%)
70.3
57.6
67.67
59.33
61.33
62
T
op15
Retrie
v
ed
mo
vies
229
59.7
211
185
195
197
Accurac
y
(%)
76.3
179
70.3
61.7
65
65.7
T
op20
Retrie
v
ed
mo
vies
249
197
219
190
215
209
Accurac
y
(%)
83
65.7
73
63.3
71.7
69.7
Figure
3
illustrates
an
e
xample
of
a
mo
vie
search
using
(ChatGPT)
description
of
“F
ast
and
Furious:
T
ok
yo
Drift.
”
T
able
4
sho
ws
the
top
10
similar
mo
vies
to
the
query
“
A
person
who
mo
v
ed
to
another
country
because
of
his
obsession
with
racing
cars.
”
The
actual
mo
vie
summary
is
“
A
teenager
becomes
a
major
com-
petitor
in
the
w
orld
of
drift
racing
after
mo
ving
in
with
his
f
ather
in
T
ok
yo
to
a
v
oid
a
jail
sentence
in
America.
”
The
results
indicate
that
the
mo
vie
“F
ast
and
Furious:
T
ok
yo
Drift”
is
rank
ed
number
one
with
a
similarity
score
of
0.5318.
Figure
3.
Example
for
retrie
v
ed
mo
vies
when
using
the
description
of
“F
ast
and
Furious:
T
ok
yo
Drift”
5.
DISCUSSION
AND
LIMIT
A
TIONS
Although
we
produced
encouraging
results
utilizing
USE
on
our
test
data
compared
to
other
models
on
the
same
mo
vie
data,
it
is
e
vident
that
there
is
much
potential
for
accurac
y
impro
v
ement
as
we
e
xplore
the
realm
of
search
engines.
One
plausible
a
p
pr
o
a
ch
to
addressing
this
issue
could
in
v
olv
e
the
inte
gration
of
an
e
xpanded
dataset
or
the
uti
lization
of
storylines
and
scenarios
from
the
mo
vies
as
a
method
for
training
the
model.
USE
may
struggle
to
comprehend
the
signicance
of
idiomatic
or
gurati
v
e
language
due
to
its
improper
usage,
especially
when
it
represents
a
general
summary
of
a
specic
mo
vie.
A
potential
consequence
of
this
phenomenon
is
the
decrease
in
performance
when
eng
aging
in
tasks
that
strongly
depend
on
comprehending
the
subtleties
of
language,
such
as
sarcasm
or
metaphor
.
Furthermore,
the
model
may
encounter
challenges
in
comprehending
the
sentence’
s
conte
xt
or
the
interconnections
among
v
arious
concepts
in
the
te
xt.
This
phenomenon
may
result
in
a
decline
in
performance
when
e
v
aluating
te
xts
that
include
identical
semantic
content
b
ut
e
xhibit
distinct
surf
ace-le
v
el
grammar
.
Another
thing
that
mak
es
it
hard
to
use
is
that
it
relies
on
a
transformer
-based
architecture
that
is
not
really
made
for
semantic
similarity
Int
J
Elec
&
Comp
Eng,
V
ol.
14,
No.
6,
December
2024:
7137-7144
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
❒
7143
b
ut
for
understanding
language
in
general.
This
may
result
in
reduced
ef
cienc
y
or
accurac
y
in
some
specic
cases.
Des
pite
the
current
le
v
el
of
precision,
the
results
indicate
that
model
USE
achie
v
ed
the
highest
le
v
el
of
accurac
y
when
compared
to
models
SBER
T
and
ChatGPT
,
e
v
en
though
we
used
model
ChatGPT
to
generate
the
test
data.
This
is
still
remarkable
and
can
serv
e
as
a
benchmark
for
e
v
aluating
the
ef
fecti
v
eness
of
a
synopsis
search
engine,
especially
for
mo
vie
ent
husiasts
who
frequently
look
for
a
specic
mo
vie
based
on
a
synopsis
the
y
ha
v
e
heard
or
encountered,
e
v
en
without
kno
wing
an
ything
about
the
title
of
the
lm.
6.
CONCLUSION
In
this
paper
,
we
highlight
the
promising
results
of
using
the
USE
for
mo
vie
semantic
search,
out-
performing
other
models
lik
e
SBER
T
and
ChatGPT
in
accurac
y
.
Ho
we
v
er
,
we
identied
potential
for
im-
pro
v
ement,
particularly
in
handling
idiomatic
language
and
understanding
conte
xt
and
connections
in
te
xt.
W
e
suggest
e
xpanding
the
dataset
and
incorporating
mo
vie
storylines
for
better
train-
ing.
Our
ndings
instill
hope
for
both
a
vid
mo
vie
enthusiasts
and
academics
in
the
search
engine
and
natural
language
processing
elds.
REFERENCES
[1]
J.
Lund
and
Y
.-K.
Ng,
“Mo
vie
recommendations
using
the
deep
learning
approach,
”
in
2018
IEEE
International
Confer
ence
on
Information
Reuse
and
Inte
gr
ation
(IRI)
,
Jul.
2018,
pp.
47–54,
doi:
10.1109/IRI.2018.00015.
[2]
K.
Raj,
A.
A.
Das,
A.
Guha,
P
.
Sharma,
and
M.
K.
S,
“Mo
vie
recommendation
system,
”
International
J
ournal
of
Computer
Sciences
and
Engineering
,
v
ol.
7,
no.
4,
pp.
1024–1028,
2019,
doi:
10.26438/ijcse/v7i4.10241028.
[3]
H.
Ezzik
ouri,
M.
Erritali,
and
M.
Ouk
essou,
“Semantic
similarity/relatedness
for
cross
language
plagiarism
detec-
tion,
”
in
2016
13
th
International
Confer
ence
on
Computer
Gr
aphics,
Ima
ging
and
V
isualization
(CGiV)
,
Mar
.
2016,
v
ol.
1,
no.
2,
pp.
372–374,
doi:
10.1109/CGiV
.2016.78.
[4]
D.
Cer
et
al.
,
“Uni
v
ersal
sentence
encoder
for
Englis
h,
”
in
EMNLP
2018
-
Confer
ence
on
Empirical
Methods
in
Natur
al
Langua
g
e
Pr
ocessing:
System
Demonstr
ations,
Pr
oceedings
,
2018,
pp.
169–174,
doi:
10.18653/v1/d18-
2029.
[5]
S.
Lai,
K.
Liu,
S.
He,
and
J.
Zhao,
“Ho
w
to
generate
a
good
w
ord
embedding?,
”
IEEE
Intellig
ent
Systems
,
pp.
1–1,
2017,
doi:
10.1109/mis.2017.2581325.
[6]
J.
De
vlin,
M.-W
.
Chang,
K.
Lee,
K.
T
.
Google,
and
A.
I.
Language,
“BER
T
:
Pre-training
of
deep
bidirectional
transformers
for
language
understanding,
”
Naacl-Hlt
2019
,
pp.
4171–4186,
2018.
[7]
N.
Reimers
and
I.
Gure
vych,
“Sentence-BER
T
:
Sentence
embeddings
using
siamese
BER
T
-netw
orks,
”
in
Pr
oceed-
ings
of
the
2019
Confer
ence
on
Empirical
Methods
in
Natur
al
Langua
g
e
Pr
ocessing
and
the
9
th
International
J
oint
Confer
ence
on
Natur
al
Langua
g
e
Pr
ocessing
(EMN
LP-IJCNLP)
,
2019,
pp.
3980–3990,
doi:
10.18653/v1/D19-1410.
[8]
S.
Sridhar
,
D.
Dha
nasekaran,
and
G.
C.
P
.
Latha,
“Content-based
mo
vie
recommendation
system
using
MBO
with
DBN,
”
Intellig
ent
A
utomation
and
Soft
Computing
,
v
ol.
35,
no.
3,
pp.
3241–3257,
2023,
doi:
10.32604/i-
asc.2023.030361.
[9]
R.
Ahuja,
A.
Solanki,
and
A.
Nayyar
,
“Mo
vie
recommender
system
using
K-means
clustering
and
K-nearest
neigh-
bor
,
”
in
2019
9
th
International
Confer
ence
on
Cloud
Computing
,
Data
Science
and
Engineering
(Conuence)
,
Jan.
2019,
pp.
263–268,
doi:
10.1109/CONFLUENCE.2019.8776969.
[10]
M.
R.
Hasa
n
and
J.
Ferdous,
“Dominance
of
AI
and
machine
learning
techniques
in
h
ybrid
mo
vie
recommendation
system
applying
te
xt-to-number
con
v
ersion
and
cosine
similarity
approaches,
”
J
ournal
of
Computer
Science
and
T
ec
hnolo
gy
Studies
,
v
ol.
6,
no.
1,
pp.
94–102,
2024,
doi:
10.32996/jcsts.2024.6.1.10.
[11]
H.
Leng
et
al.
,
“Finding
similar
mo
vies:
Dataset,
tools,
and
methods,
”
Computer
Science
Resear
c
h
Notes
,
v
ol.
2802,
pp.
115–124,
2018,
doi:
10.24132/CSRN.2018.2802.15.
[12]
A.
K
utlimurato
v
and
N.
Atadj
ano
v
a
,
“Mo
vie
recommender
system
using
con
v
olutional
neural
netw
orks
algorithm,
”
Science
and
Inno
vation
,
v
ol.
2,
no.
3,
2023.
[13]
K.
K.
Y
ada
v
,
H.
K.
Soni,
G.
Y
ada
v
,
and
M.
Sharma,
“Collaborati
v
e
ltering
based
h
ybrid
recommendation
system
using
neural
netw
ork
and
matrix
f
actorization
techniques,
”
International
J
ournal
of
Intellig
ent
Systems
and
Applica-
tions
in
Engineering
,
v
ol.
12,
no.
8s,
pp.
695–701,
2024.
[14]
S.
Reddy
and
J.
K.
Sarode,
“Hybrid
recommendation
system
using
graph
neural
netw
ork
and
BER
T
embeddings,
”
arXiv:2310.04878
,
2023.
[15]
Y
.
Mu
and
Y
.
W
u,
“Multimodal
mo
vie
recommendation
system
using
deep
learning,
”
Mathematics
,
v
ol.
11,
no.
4,
2023,
doi:
10.3390/math11040895.
[16]
A.
Li,
X.
Liu,
and
B.
Y
ang,
“Item
attrib
ute-a
w
are
graph
collaborati
v
e
ltering,
”
Expert
Systems
with
Applications
,
v
ol.
238,
2024,
doi:
10.1016/j.esw
a.2023.122242.
[17]
V
.
L.
Chetana
and
H.
Seetha,
“Handling
massi
v
e
sparse
data
in
recommendation
systems,
”
J
ournal
of
Information
and
Knowledg
e
Mana
g
ement
,
v
ol.
23,
no.
3,
2024,
doi:
10.1142/S0219649224500217.
A
semantic
similarity
sear
c
h
engine
for
mo
vies
(Ahmad
Mustafa)
Evaluation Warning : The document was created with Spire.PDF for Python.
7144
❒
ISSN:
2088-8708
[18]
A.
Azaria,
A.
Hassidim,
S.
Kraus,
A.
Eshk
ol,
O.
W
eintraub,
and
I.
Netanely
,
“Mo
vie
recommender
system
for
prot
maximization,
”
AAAI
W
orkshop
-
T
ec
hnical
Report
,
v
ol.
WS-13-11,
pp.
2–8,
2013.
[19]
B.
Chinnapottu,
G.
Arikatla,
and
S.
Memeti,
“Mo
vie
prediction
based
on
mo
vie
scripts
using
natural
language
pro-
cessing
and
machine
learning
algorithms,
”
M.S.
thesis,
F
aculty
of
Computing,
Blekinge
Institute
of
T
echnology
,
2021.
[20]
H.
Y
oo,
M.
Kang,
and
K.
Oh,
“
A
semantic
search
model
using
w
ord
embedding,
POS
tagging,
and
named
entity
recognition,
”
in
2018
International
Confer
ence
on
Computational
Science
and
Computational
Intellig
ence
(CSCI)
,
Dec.
2018,
pp.
1204–1209,
doi:
10.1109/CSCI46756.2018.00231.
[21]
S.
Bhattacharya
and
A.
Lundia,
“Mo
vie
recommendation
system
using
bag
of
w
ords
and
Scikit-Learn,
”
International
J
ournal
of
Engineering
Applied
Sciences
and
T
ec
hnolo
gy
,
v
ol
.
04,
no.
05,
pp.
526–528,
2019,
doi:
10.33564/i-
jeast.2019.v04i05.076.
[22]
S.
Ro
y
,
M.
Sharma,
and
S.
K.
Singh,
“Mo
vie
recommendation
system
using
semi-supervised
learning,
”
in
2019
Global
C
onfer
ence
for
Advancement
in
T
ec
hnolo
gy
(GCA
T)
,
Oct.
2019,
pp.
1–5,
doi:
10.1109/G-
CA
T47503.2019.8978353.
[23]
B.
W
alek
and
V
.
F
ojtik,
“
A
h
ybrid
recommender
system
for
recommending
rele
v
ant
mo
vies
using
an
e
xpert
system,
”
Expert
Systems
with
Applications
,
v
ol.
158,
2020,
doi:
10.1016/j.esw
a.2020.113452.
[24]
X.
Y
e,
Z.
Qi,
X.
Song,
X.
He,
and
D.
Masse
y
,
“Generalized
learning
of
neural
netw
ork
based
semantic
similarity
models
and
its
application
in
mo
vie
search,
”
in
2015
IEEE
International
Confer
ence
on
Data
Mining
W
orkshop
(ICDMW)
,
No
v
.
2015,
pp.
86–93,
doi:
10.1109/ICDMW
.2015.34.
[25]
N.
Craswell,
“Mean
reciprocal
rank,
”
Encyclopedia
of
Database
Systems
,
pp.
2217–2217,
2018,
doi:
10.1007/978-1-
4614-8265-9
488.
[26]
E.
Zhang
and
Y
.
Zhang,
“
A
v
erage
precision,
”
Encyclopedia
of
Database
Systems
,
pp.
192–193,
2009,
doi:
10.1007/978-0-387-39940-9
482.
[27]
N.
Reimers
and
I.
Gure
vych,
“Sentence-BER
T
:
Sentence
embeddings
using
siames
e
BER
T
-netw
orks,
”
in
EMNLP-
IJCNLP
2019
-
2019
Confer
ence
on
Empirical
M
ethods
in
Natur
al
Langua
g
e
Pr
ocessing
and
9
th
International
J
oint
Confer
ence
on
Natur
al
Langua
g
e
Pr
ocessing
,
Pr
oceedings
of
the
Confer
ence
,
2019,
pp.
3982–3992,
doi:
10.18653/v1/d19-1410.
BIOGRAPHIES
OF
A
UTHORS
Ahmad
Mustafa
recei
v
ed
the
Ph.D.
de
gree
in
computer
science
from
the
Uni
v
ersit
y
of
T
e
xas
at
Dallas
in
2018.
He
is
currently
an
assistant
professor
at
the
Jordan
Uni
v
ersity
of
Science
and
T
echnology
.
His
current
research
interests
include
continual
learning,
acti
v
e
learning,
and
natural
language
processing.
He
can
be
contacted
at
email:
ammustaf
a@just.edu.jo.
Hammam
Mheidat
is
a
master
student
and
teaching
assistant
in
Data
Science
P
rogram
in
the
Jordan
Uni
v
ersity
of
Science
and
T
echnology
.
His
research
interests
include
machine
learning,
deep
learning,
computer
vision,
and
natural
language
proce
ssing.
He
can
be
contacted
at
email:
hmmheidat206@cit.just.edu.jo.
Adam
Shatnawi
is
a
master
student
in
Data
Science
Program
in
the
Jordan
Uni
v
er
-
sity
of
Science
and
T
echnology
.
His
areas
of
research
enc
ompass
deep
learning,
data
analytics
and
visualization,
as
well
as
natura
l
language
processing.
He
can
be
contacted
at
email:
azshat-
na
wi20@cit.just.edu.jo.
Int
J
Elec
&
Comp
Eng,
V
ol.
14,
No.
6,
December
2024:
7137-7144
Evaluation Warning : The document was created with Spire.PDF for Python.