I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
,
N
o.
2
,
A
pr
il
20
25
, pp.
1597
~
1604
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
2
.pp
1597
-
1604
1597
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
A
c
om
p
ar
at
i
ve
st
u
d
y of
n
at
u
r
al
l
an
g
u
age
i
n
f
e
r
e
n
c
e
i
n
S
w
ah
i
l
i
u
si
n
g
m
on
ol
i
n
gu
a
l
an
d
m
u
l
t
i
l
i
n
gu
al
m
od
e
l
s
H
aj
r
a F
ak
i
A
li
, A
d
il
a A
lf
a K
r
is
n
ad
h
i
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
, F
a
c
ul
t
y of
C
om
put
e
r
S
c
i
e
nc
e
, U
ni
ve
r
s
i
t
a
s
I
n
done
s
i
a
, D
e
pok,
I
ndone
s
i
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
F
e
b 15, 2024
R
e
vi
s
e
d
O
c
t
29, 2024
A
c
c
e
pt
e
d
N
ov 14, 2024
Recent
advancements
in
large
language
models
(LLMs)
have
led
to
opportunities
for
improving
applications
across
various
domains.
Ho
wever,
existin
g
LLMs
fine
-
tuned
for
Swahili
or
other
African
languages
often
rely
on
pre
-
trained
multilingual
models,
resulting
in
a
relatively
small
por
tion
of
training
data
dedicate
d
to
Swahili.
In
this
study,
we
compa
re
the
performance
of
monolingual
and
multilingual
models
in
Swahili
natural
language
inferenc
e
tasks
using
the
cross
-
lingual
natural
language
inferenc
e
(XNLI)
dataset.
Our
research
demonstrates
the
superior
effectiveness
of
dedicated
Swahili
monolingual
models,
achieving
an
accuracy
rate
of
69%.
These
monolingual
models
exhibit
significantly
enhanced
precision,
recall,
and
F1
scores,
particularly
in
predicting
contradiction
and
neu
trality.
Overall,
the
findings
in
this
a
rticle
emphasize
the
critical
importa
nce
of
using
monolingual
models
in
low
-
resource
language
processing
contexts,
providing
valuable
insights
for
developing
more
efficient
and
t
ailored
natural
language
processing
systems
that
benefit
languages
facing
similar
resource constrai
nts.
K
e
y
w
o
r
d
s
:
M
onol
in
gua
l
m
ode
l
M
ul
ti
li
ngua
l
m
ode
l
N
a
tu
r
a
l
la
ngua
ge
i
nf
e
r
e
nc
e
S
w
a
hi
li
X
N
L
I
da
ta
s
e
t
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
A
di
la
A
lf
a
K
r
is
na
dhi
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
S
c
ie
nc
e
, F
a
c
ul
ty
of
C
om
put
e
r
S
c
ie
nc
e
, U
ni
ve
r
s
it
a
s
I
ndone
s
i
a
U
I
C
a
m
pus
D
e
pok, W
e
s
t
J
a
va
16424, I
ndone
s
ia
E
m
a
il
:
a
di
la
@
c
s
.ui
.a
c
.i
d
1.
I
N
T
R
O
D
U
C
T
I
O
N
S
w
a
hi
li
is
a
r
ic
h
la
ngua
ge
w
id
e
ly
us
e
d
by
ove
r
100
m
il
li
on
pe
opl
e
in
c
ount
r
ie
s
a
c
r
os
s
E
a
s
t
a
nd
C
e
nt
r
a
l
A
f
r
ic
a
,
s
uc
h
a
s
T
a
n
z
a
ni
a
,
K
e
ny
a
,
a
nd
U
ga
nda
[
1]
,
[
2]
.
I
ts
pr
om
in
e
nc
e
le
a
ds
to
it
s
pr
e
s
e
nc
e
in
in
te
r
na
ti
ona
l
m
e
di
a
,
s
u
c
h
a
s
th
e
B
r
it
is
h
B
r
oa
d
c
a
s
ti
ng
C
or
por
a
ti
on
(
B
B
C
)
[
3]
a
nd
E
ur
ope
M
e
di
a
M
oni
to
r
[
4]
.
T
he
s
ig
ni
f
ic
a
nt
pr
e
s
e
nc
e
of
S
w
a
hi
li
m
ot
iv
a
te
s
th
e
r
e
c
e
nt
a
d
opt
io
n
of
na
tu
r
a
l
la
ngua
ge
pr
oc
e
s
s
in
g
(
N
L
P
)
te
c
hnol
ogi
e
s
to
pr
ovi
de
m
or
e
a
s
s
is
ta
nc
e
to
S
w
a
hi
li
-
s
pe
a
ki
ng
us
e
r
s
.
S
om
e
e
xa
m
pl
e
s
of
s
uc
h
te
c
hnol
ogi
e
s
in
c
lu
de
t
he
S
w
a
hi
li
c
ha
tb
ot
s
of
voda
c
a
r
e
a
nd a
da
he
a
lt
h
[
5]
,
[
6]
. H
ow
e
ve
r
, t
he
s
e
c
ha
tb
ot
s
a
r
e
pr
e
tt
y l
im
it
e
d i
n
th
e
ir
a
bi
li
ty
t
o r
e
s
pond to us
e
r
s
'
que
r
ie
s
a
nd, t
hus
,
c
a
nnot
i
nt
e
r
a
c
t
m
uc
h w
it
h us
e
r
s
.
T
he
r
e
c
e
nt
a
dva
n
c
e
m
e
nt
s
in
la
r
ge
la
ngua
ge
m
ode
l
s
(
L
L
M
s
)
pr
ovi
de
a
n
oppor
tu
ni
ty
to
im
p
r
ove
s
uc
h
a
ppl
ic
a
ti
ons
i
n va
r
io
us
doma
in
s
. N
on
e
th
e
le
s
s
, e
xi
s
ti
ng L
L
M
s
f
i
ne
-
tu
ne
d f
or
S
w
a
hi
li
or
a
c
ol
le
c
ti
on of
A
f
r
ic
a
n
la
ngua
ge
s
,
s
uc
h
a
s
r
obus
tl
y
opt
im
iz
e
d
bi
di
r
e
c
ti
ona
l
e
nc
ode
r
r
e
pr
e
s
e
nt
a
ti
ons
f
r
om
tr
a
ns
f
or
m
e
r
s
(
R
oB
E
R
T
a
)
-
ba
s
e
-
w
e
c
h
s
e
l
-
S
w
a
hi
li
[
7]
a
nd
A
f
r
ic
a
n
bi
di
r
e
c
ti
ona
l
e
nc
ode
r
r
e
pr
e
s
e
nt
a
ti
ons
f
r
om
tr
a
ns
f
or
m
e
r
s
(
A
f
r
iB
E
R
T
a
)
[
8]
, of
te
n r
e
ly
on
pr
e
-
t
r
a
in
e
d m
ul
ti
l
in
gua
l
m
ode
ls
l
ik
e
m
ul
ti
li
ng
ua
l
bi
di
r
e
c
ti
ona
l
e
nc
ode
r
r
e
pr
e
s
e
nt
a
ti
ons
f
r
o
m
tr
a
ns
f
or
m
e
r
s
(
m
B
E
R
T
)
[
9]
o
r
c
r
os
s
-
li
ngua
l
la
ngua
ge
m
ode
l
w
it
h
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
ks
(
X
L
M
-
R
)
[
10]
,
w
he
r
e
th
e
pr
e
-
tr
a
in
in
g
da
ta
in
c
lu
de
s
m
a
ny
la
ngua
ge
s
a
r
ound
th
e
w
or
ld
.
C
ons
e
que
nt
ly
,
th
e
S
w
a
hi
li
po
r
ti
on
o
f
th
e
tr
a
in
in
g
da
ta
is
of
te
n
ti
ny
c
om
pa
r
e
d
to
ot
he
r
la
ngua
ge
s
in
th
e
s
e
m
ul
ti
li
ngua
l
m
ode
ls
[
11]
.
F
or
e
xa
m
pl
e
,
th
e
por
ti
on of
S
w
a
hi
li
pr
e
-
t
r
a
in
in
g da
ta
f
or
m
B
E
R
T
m
a
ke
s
up l
e
s
s
t
ha
n 1%
of
i
ts
voc
a
bul
a
r
y
[
12]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 2, A
pr
il
2025
:
1597
-
1604
1598
O
n
th
e
ot
he
r
ha
nd,
S
w
a
hi
li
bi
di
r
e
c
ti
ona
l
e
nc
od
e
r
r
e
pr
e
s
e
nt
a
ti
ons
f
r
om
tr
a
ns
f
or
m
e
r
s
(
S
w
a
hB
E
R
T
)
[
13]
is
a
m
onol
in
gua
l
m
ode
l
s
pe
c
if
ic
a
ll
y
tr
a
in
e
d
f
or
S
w
a
hi
li
.
I
t
w
a
s
tr
a
in
e
d
f
or
th
e
f
ol
lo
w
in
g
N
L
P
ta
s
ks
:
na
m
e
d
e
nt
it
y
r
e
c
ogni
ti
on
(
N
E
R
)
,
a
c
hi
e
vi
ng
88.50%
a
c
c
ur
a
c
y;
ne
w
s
c
la
s
s
if
ic
a
ti
on
w
it
h
90.90%
a
c
c
ur
a
c
y;
e
m
ot
io
n
c
la
s
s
if
ic
a
ti
on
w
it
h
64.46%
a
c
c
ur
a
c
y;
a
nd
s
e
nt
im
e
nt
a
na
ly
s
is
w
it
h
70.94%
a
c
c
ur
a
c
y.
O
ne
im
por
ta
nt
N
L
P
t
a
s
k m
is
s
in
g f
r
om
t
he
s
e
r
e
s
ul
ts
i
s
n
a
tu
r
a
l
la
ngua
ge
i
nf
e
r
e
n
c
e
(
N
L
I
)
.
N
L
I
(
a
ls
o
known
a
s
te
xt
ua
l
e
nt
a
il
m
e
nt
)
is
th
e
ta
s
k
of
de
te
r
m
in
in
g
th
e
e
nt
a
il
m
e
nt
r
e
la
ti
ons
hi
p
be
twe
e
n
a
"
pr
e
m
is
e
"
a
nd
a
"
hypothe
s
is
"
-
w
he
th
e
r
th
e
hypothe
s
i
s
is
tr
ue
(
e
nt
a
il
m
e
nt
)
,
f
a
ls
e
(
c
ont
r
a
di
c
ti
on)
,
or
unde
te
r
m
in
e
d
(
ne
ut
r
a
l)
gi
ve
n
th
e
pr
e
m
is
e
[
14]
,
[
15]
.
E
nt
a
il
m
e
nt
in
N
L
I
di
f
f
e
r
s
f
r
om
s
ta
nda
r
d
lo
gi
c
a
l
e
nt
a
il
m
e
nt
in
th
a
t
it
doe
s
not
r
e
qui
r
e
s
tr
ic
t
lo
gi
c
a
l
s
e
m
a
nt
ic
s
of
th
e
s
e
nt
e
nc
e
s
to
obt
a
in
a
nd
de
r
iv
e
th
e
e
nt
a
il
m
e
nt
.
I
ns
te
a
d,
in
N
L
I
,
th
e
pr
e
m
is
e
e
nt
a
il
s
th
e
hypothe
s
is
of
w
h
e
th
e
r
ty
pi
c
a
l
hum
a
n
r
e
a
di
ng
w
oul
d
ju
s
ti
f
y t
ha
t
th
e
e
nt
a
il
m
e
nt
hol
ds
. F
r
om
a
n a
ppl
ic
a
ti
on pe
r
s
pe
c
ti
v
e
, a
s
ol
ut
io
n t
o t
he
N
L
I
t
a
s
k c
a
n of
te
n be
us
e
d
in
ot
he
r
,
m
or
e
c
om
pl
e
x
N
L
P
ta
s
ks
,
s
uc
h
a
s
que
s
ti
on
-
a
n
s
w
e
r
in
g
[
16]
,
doc
um
e
nt
s
um
m
a
r
iz
a
ti
on
[
17]
,
a
n
d
in
f
or
m
a
ti
on
e
xt
r
a
c
ti
on
[
18]
.
I
n
th
e
c
ont
e
xt
of
th
is
pa
pe
r
,
s
e
ve
r
a
l
S
w
a
hi
li
c
ha
tb
ot
a
ppl
ic
a
ti
ons
c
oul
d
be
m
uc
h
im
pr
ove
d i
f
t
he
y a
r
e
e
qui
ppe
d w
it
h N
L
I
m
ode
ls
be
c
a
us
e
c
ur
r
e
n
tl
y, t
hos
e
c
ha
tb
ot
s
a
r
e
pr
e
tt
y s
im
pl
is
ti
c
i
n t
he
ir
a
ppr
oa
c
h,
na
m
e
ly
us
in
g
s
tr
in
g
-
ba
s
e
d
p
a
tt
e
r
n
m
a
tc
hi
ng
a
nd
h
a
ndc
r
a
f
te
d
r
ul
e
s
to
e
na
bl
e
th
e
a
ppl
ic
a
ti
ons
to
r
e
s
pond to us
e
r
'
s
que
r
ie
s
.
I
n
a
ddi
ti
on
to
m
ot
iv
a
ti
on
f
r
om
a
ppl
ic
a
ti
on
a
r
e
a
s
,
th
e
g
a
p
r
e
m
a
in
s
r
e
ga
r
di
ng
w
h
e
th
e
r
a
ny
S
w
a
hi
li
m
onol
in
gua
l
m
ode
l
c
oul
d
pr
ovi
de
a
be
tt
e
r
s
ol
ut
io
n
to
th
e
N
L
I
ta
s
k
th
a
n
m
ul
ti
li
ngua
l
m
ode
ls
.
I
t
ha
s
be
e
n
s
how
n
th
a
t
f
or
lo
w
-
r
e
s
our
c
e
la
ngua
ge
s
,
m
onol
in
gua
l
m
ode
ls
of
te
n
out
pe
r
f
or
m
th
e
ir
m
ul
ti
li
ngua
l
c
ount
e
r
pa
r
ts
[
19]
–
[
23]
.
T
hi
s
a
ppl
ie
s
e
ve
n
to
hi
gh
-
r
e
s
our
c
e
la
ngua
g
e
s
w
he
n
s
pe
c
if
ic
N
L
P
ta
s
ks
a
r
e
c
ons
id
e
r
e
d
[
24]
,
[
25]
.
T
hus
,
a
n
s
w
e
r
in
g
th
is
qu
e
s
ti
on
c
a
n
e
nc
our
a
ge
f
ur
th
e
r
r
e
s
e
a
r
c
h
a
nd
de
ve
lo
pm
e
nt
of
S
w
a
hi
li
m
onol
in
gua
l
m
ode
ls
.
T
hi
s
pa
pe
r
a
im
s
to
pe
r
f
or
m
a
c
om
pa
r
a
ti
ve
a
na
ly
s
is
be
twe
e
n
S
w
a
hB
E
R
T
[
13]
a
nd
m
B
E
R
T
[
9]
.
T
he
f
or
m
e
r
is
th
e
onl
y
known
S
w
a
hi
li
m
onol
in
gua
l
m
ode
l,
w
hi
le
t
he
la
tt
e
r
is
a
w
id
e
ly
us
e
d
m
ul
ti
li
ngua
l
m
ode
l
pr
e
-
tr
a
in
e
d
in
104
la
ngua
g
e
s
,
in
c
lu
di
ng
S
w
a
hi
li
[
26]
.
O
ur
w
or
k
lo
oks
a
t
th
e
pe
r
f
or
m
a
nc
e
of
bot
h
m
od
e
ls
on
th
e
S
w
a
hi
li
N
L
I
ta
s
k.
S
pe
c
if
ic
a
ll
y,
w
e
f
in
e
-
tu
ne
bot
h
m
ode
l
s
f
or
th
e
dow
ns
tr
e
a
m
N
L
I
ta
s
k
on
a
S
w
a
hi
li
s
ubs
e
t
of
th
e
c
r
os
s
-
li
ngua
l
na
tu
r
a
l
la
ngu
a
ge
in
f
e
r
e
nc
e
(
X
N
L
I
)
da
ta
s
e
t
[
27]
.
T
he
r
e
s
ul
ts
f
il
l
th
e
S
w
a
hi
li
N
L
I
r
e
s
e
a
r
c
h ga
p a
s
ne
it
he
r
m
ode
l
w
a
s
t
r
a
in
e
d f
or
t
he
S
w
a
hi
li
N
L
I
t
a
s
k.
T
he
r
e
s
t
of
th
e
pa
pe
r
i
s
or
ga
ni
z
e
d,
m
a
ki
ng
up
le
s
s
th
a
n
1%
.
W
e
f
ol
lo
w
th
is
w
it
h
a
n
e
xpl
a
na
ti
on
of
our
m
e
th
od
in
s
e
c
ti
on
2,
w
he
r
e
w
e
e
xpl
a
in
th
e
m
od
e
ls
,
th
e
d
a
t
a
s
e
t,
a
nd
th
e
e
va
lu
a
ti
on
s
c
e
n
a
r
io
s
.
A
f
te
r
th
a
t,
s
e
c
ti
on
3
de
t
a
il
s
th
e
e
va
lu
a
ti
on
r
e
s
ul
t
s
,
r
e
ve
a
li
ng
e
xc
it
in
g
i
ns
ig
ht
s
in
to
th
e
p
e
r
f
or
m
a
nc
e
of
our
m
ode
ls
.
F
in
a
ll
y, s
e
c
ti
on 4 c
onc
lu
de
s
t
h
e
pa
pe
r
, s
um
m
a
r
iz
in
g our
f
in
di
ngs
a
nd s
ugge
s
ti
ng a
ve
nue
s
f
or
f
ut
ur
e
r
e
s
e
a
r
c
h.
2.
M
E
T
H
O
D
I
n
our
r
e
s
e
a
r
c
h,
w
e
pr
e
pa
r
e
d
our
da
ta
s
e
t
a
nd
s
e
le
c
te
d
S
w
a
hB
E
R
T
f
or
th
e
m
onol
in
gua
l
m
ode
l
a
nd
m
B
E
R
T
f
or
th
e
m
ul
ti
li
ngua
l
m
ode
l.
W
e
f
in
e
-
tu
ne
d
bot
h
S
w
a
hB
E
R
T
a
nd
m
B
E
R
T
f
or
th
e
N
L
I
ta
s
k
a
nd
e
va
lu
a
te
d
bot
h
m
ode
ls
.
F
in
a
ll
y,
w
e
c
onduc
te
d
a
c
om
pa
r
a
ti
ve
a
na
ly
s
is
be
twe
e
n
th
e
two
m
ode
ls
.
O
ur
ge
ne
r
a
l
a
ppr
oa
c
h f
ol
lo
w
s
a
s
ta
nda
r
d w
or
kf
lo
w
, a
s
d
e
pi
c
te
d i
n F
ig
ur
e
1
.
2.1.
D
at
a c
ol
le
c
t
io
n
an
d
p
r
e
-
p
r
oc
e
s
s
in
g
T
hi
s
s
tu
dy
ut
il
iz
e
s
th
e
S
w
a
hi
li
s
e
gm
e
nt
of
th
e
X
N
L
I
da
ta
s
e
t,
w
hi
c
h
c
ont
a
in
s
392,702
tr
a
in
in
g
s
e
t
s
,
2,490
va
li
da
ti
ons
s
e
ts
,
a
nd
5,010
te
s
t
s
e
ts
of
th
e
pa
ir
s
o
f
s
e
nt
e
nc
e
s
[
27]
.
E
a
c
h
s
e
nt
e
nc
e
pa
ir
c
ons
is
ts
of
a
pr
e
m
is
e
,
hypothe
s
is
,
a
nd
la
be
l.
T
he
la
b
e
ls
in
di
c
a
te
w
he
th
e
r
t
he
hypothe
s
is
i
s
e
nt
a
il
e
d
by
th
e
pr
e
m
is
e
(
0)
,
c
ont
r
a
di
c
te
d
by
th
e
pr
e
m
is
e
(
1)
,
or
ne
ut
r
a
l
c
onc
e
r
ni
ng
th
e
pr
e
m
is
e
(
2)
.
T
a
bl
e
1
s
how
s
s
om
e
s
a
m
pl
e
s
f
r
om
th
e
X
N
L
I
da
ta
s
e
t.
T
a
bl
e
1
c
ont
a
in
s
E
ngl
is
h
tr
a
ns
la
ti
ons
in
it
a
li
c
s
s
o
th
e
r
e
a
de
r
c
a
n
unde
r
s
ta
nd
th
e
m
.
W
hi
le
th
e
us
e
of
th
e
m
onol
in
gua
l
m
ode
l
ha
s
s
how
n
pr
om
is
in
g
r
e
s
ul
ts
i
n
s
om
e
dow
ns
tr
e
a
m
ta
s
ks
in
S
w
a
hi
li
,
f
u
r
th
e
r
r
e
s
e
a
r
c
h
is
ne
e
de
d
to
f
in
e
-
tu
ne
th
e
e
xi
s
ti
ng
m
ono
li
ngua
l
m
o
de
l
in
S
w
a
hi
li
,
s
uc
h
a
s
S
w
a
hB
E
R
T
f
or
N
L
I
,
to
de
te
r
m
in
e
i
t
is
s
tr
e
ngt
h i
n N
L
I
t
a
s
ks
t
o a
ddr
e
s
s
t
he
uni
que
l
in
gui
s
ti
c
a
nd c
ul
tu
r
a
l
c
ha
ll
e
nge
s
of
S
w
a
hi
li
N
L
I
[
28]
.
T
o
e
ns
ur
e
th
e
qua
li
ty
of
th
e
da
t
a
a
nd
c
om
pa
ti
bi
li
ty
w
it
h
th
e
m
ode
ls
,
th
e
da
t
a
s
e
t
und
e
r
goe
s
th
or
ough
pr
e
-
pr
oc
e
s
s
in
g.
T
hi
s
in
vol
ve
s
s
e
ve
r
a
l
s
te
ps
,
in
c
lu
di
ng
te
xt
c
le
a
ni
ng
to
r
e
m
ove
s
pe
c
ia
l
c
ha
r
a
c
te
r
s
a
nd
punc
tu
a
ti
on. Additi
ona
ll
y, B
e
r
tT
oke
ni
z
e
r
i
s
a
ppl
ie
d t
o t
oke
ni
z
e
t
he
i
nput
t
e
xt
in
to
s
ubw
or
ds
s
ui
ta
bl
e
f
or
i
npu
t
in
to
a
B
E
R
T
-
ba
s
e
d m
od
e
l.
2.2.
T
h
e
m
od
e
ls
:
S
w
ah
B
E
R
T
an
d
m
B
E
R
T
T
he
m
e
th
odol
ogy
e
m
pl
oye
d
in
th
is
s
tu
dy
ut
il
iz
e
s
two
tr
a
ns
f
or
m
e
r
-
ba
s
e
d
ne
ur
a
l
ne
twor
k
a
r
c
hi
te
c
tu
r
e
s
, na
m
e
ly
S
w
a
hB
E
R
T
a
nd mB
E
R
T
. S
w
a
h
B
E
R
T
i
s
a
m
ode
l
s
pe
c
if
ic
a
ll
y f
in
e
-
tu
ne
d f
or
t
he
S
w
a
hi
li
N
L
I
ta
s
k,
le
ve
r
a
gi
ng
th
e
c
a
p
a
bi
li
ti
e
s
of
th
e
B
E
R
T
m
ode
l.
O
n
t
he
ot
he
r
ha
nd,
m
B
E
R
T
is
a
pr
e
-
tr
a
in
e
d
m
ode
l
de
s
ig
ne
d t
o ha
ndl
e
m
ul
ti
pl
e
l
a
ngua
ge
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
A
c
om
par
at
iv
e
s
tu
dy
of
nat
ur
al
l
anguage
i
nf
e
r
e
nc
e
i
n Sw
ahi
li
us
in
g m
onol
in
gual
…
(
H
aj
r
a F
ak
i
A
li
)
1599
F
ig
ur
e
1
. E
xpe
r
im
e
nt
a
l
f
lo
w
c
ha
r
t
T
a
bl
e
1. S
a
m
pl
e
da
ta
f
r
om
t
he
us
e
d d
a
ta
s
e
t
P
r
e
m
i
s
e
H
ypot
he
s
i
s
L
a
be
l
N
i
l
i
pi
t
i
a s
e
t
i
y
a m
i
l
ango y
a k
abat
i
, na ni
k
aanguk
a c
hi
ni
.
I
w
e
nt
t
hr
ough a
s
e
t
of
c
upboa
r
d door
s
,
a
nd I
f
e
l
l
.
N
i
l
i
pas
ua m
i
l
ango na k
uanguk
a c
hi
ni
.
I
br
oke
t
he
door
s
a
nd f
e
l
l
.
0
K
at
i
k
at
i
y
a
m
uungani
k
o
huu
w
a
aj
abu
w
a
t
am
aduni
ni
s
hauk
u y
a k
ue
nde
l
e
a.
A
pa
s
s
i
on
f
or
pr
ogr
e
s
s
i
s
a
t
t
he
he
a
r
t
of
t
hi
s
r
e
m
a
r
ka
bl
e
f
us
i
on of
c
ul
t
ur
e
s
.
Shauk
u
y
a
k
ue
nde
l
e
a
s
i
o
m
uhi
m
u
z
ai
di
y
a
t
am
aduni
hi
z
i
.
T
hi
s
pa
s
s
i
on
f
or
pr
ogr
e
s
s
i
s
not
t
he
m
os
t
i
m
por
t
a
nt
t
hi
ng a
bout
t
he
s
e
c
ul
t
ur
e
s
1
B
ur
udani
k
w
a w
at
u w
az
i
m
a na W
at
ot
o
E
nt
e
r
t
a
i
nm
e
nt
f
or
a
dul
t
s
a
nd c
hi
l
dr
e
n.
F
ur
aha k
w
a
w
at
ot
o.
F
un f
or
ki
ds
onl
y.
2
M
as
ual
a k
at
i
k
a
u
s
ani
f
u w
a
d
at
a
I
s
s
ue
s
i
n
da
t
a
a
r
c
hi
t
e
c
t
ur
e
.
M
at
at
i
z
o k
at
i
k
a us
ani
s
i
w
a dat
a
P
r
obl
e
m
s
i
n da
t
a
s
ynt
he
s
i
s
.
0
O
na hi
y
o s
i
o m
bay
a s
ana
k
w
a m
as
aa
k
adhaa
S
e
e
, t
ha
t
'
s
not
t
oo ba
d f
or
a
f
e
w
hour
s
Si
pe
ndi
k
us
ubi
r
i
, hat
a k
w
a da
k
i
k
a 10.
I
do not
l
i
ke
w
a
i
t
i
ng, not
e
ve
n f
or
10 m
i
nut
e
s
1
S
w
a
hB
E
R
T
is
a
va
r
ia
nt
of
th
e
B
E
R
T
m
ode
l
s
pe
c
if
ic
a
ll
y
tr
a
i
ne
d
on
a
S
w
a
hi
li
da
ta
s
e
t
c
om
pr
is
in
g
105
M
B
w
it
h
16
M
w
or
ds
s
our
c
e
d
f
r
om
ne
w
s
w
e
bs
it
e
s
,
f
or
um
s
,
a
nd
W
ik
ip
e
di
a
.
I
t
is
de
s
ig
ne
d
f
or
f
our
ke
y
ta
s
ks
:
ne
w
s
c
la
s
s
if
ic
a
ti
on,
na
m
e
d
e
nt
it
y
r
e
c
ogni
ti
on,
e
m
ot
io
n
c
la
s
s
if
ic
a
ti
on,
a
nd
s
e
nt
im
e
nt
c
la
s
s
if
ic
a
ti
on.
T
he
a
r
c
hi
te
c
tu
r
e
of
S
w
a
hB
E
R
T
in
c
lu
de
s
12
e
n
c
ode
r
bl
oc
k
s
w
i
th
768
hi
dde
n
uni
ts
,
a
nd
it
u
s
e
s
s
e
lf
-
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
a
nd
f
e
e
df
or
w
a
r
d
ne
ur
a
l
ne
twor
ks
to
c
a
pt
ur
e
c
om
pl
e
x
to
ke
n
r
e
la
ti
ons
hi
ps
.
D
ur
in
g
pr
e
-
t
r
a
in
in
g,
S
w
a
hB
E
R
T
ut
il
iz
e
s
m
a
s
ke
d
la
ngua
ge
m
ode
li
ng
(
M
L
M
)
a
n
d
ne
xt
s
e
nt
e
nc
e
pr
e
di
c
ti
on
(
N
S
P
)
to
pr
e
di
c
t
m
a
s
ke
d
to
ke
n
s
a
nd
a
s
s
e
s
s
s
e
nt
e
nc
e
c
ohe
r
e
nc
e
,
r
e
s
p
e
c
ti
ve
ly
.
F
o
r
ta
s
k
-
s
pe
c
if
ic
N
L
I
a
da
pt
a
ti
on,
f
ur
th
e
r
tr
a
in
in
g
on
la
be
ll
e
d
da
ta
i
s
e
s
s
e
nt
ia
l,
in
vol
vi
ng
pa
r
a
m
e
t
e
r
upda
te
s
of
t
he
f
in
a
l
c
la
s
s
if
ic
a
ti
on
la
ye
r
to
a
li
gn
w
it
h
th
e
ta
r
ge
t
ta
s
k'
s
r
e
qui
r
e
m
e
nt
s
.
T
hi
s
f
in
e
-
tu
ni
ng
pr
oc
e
s
s
,
c
om
m
o
nl
y
m
in
im
iz
in
g
c
r
os
s
-
e
nt
r
opy
lo
s
s
,
opt
im
iz
e
s
pa
r
a
m
e
te
r
s
it
e
r
a
ti
ve
ly
,
in
c
lu
di
ng
w
e
ig
ht
s
a
nd
bi
a
s
e
s
a
s
s
o
c
ia
te
d
w
it
h
e
a
c
h
c
la
s
s
la
be
l,
ty
pi
c
a
ll
y
in
s
ta
nt
ia
t
e
d
w
it
hi
n t
he
B
e
r
tF
or
S
e
que
nc
e
C
la
s
s
if
ic
a
ti
on mode
l
.
m
B
E
R
T
,
a
v
e
r
s
a
ti
le
pr
e
-
tr
a
in
e
d
m
ode
l,
e
nc
om
pa
s
s
e
s
104
la
ngua
ge
s
,
e
m
pl
oyi
ng
M
L
M
a
nd
N
S
P
te
c
hni
que
s
.
I
t
is
bui
lt
upon
th
e
B
E
R
T
a
r
c
hi
te
c
tu
r
e
w
it
h
m
ul
ti
pl
e
tr
a
ns
f
or
m
e
r
e
nc
ode
r
la
ye
r
s
.
m
B
E
R
T
e
xt
r
a
c
t
s
c
ont
e
xt
ua
l
in
f
or
m
a
ti
on
f
r
om
bot
h
le
f
t
a
nd
r
ig
ht
to
k
e
n
c
ont
e
xt
s
.
T
he
m
B
E
R
T
-
ba
s
e
c
a
s
e
d
m
ode
l,
c
om
pr
is
in
g
12
-
la
ye
r
T
r
a
ns
f
or
m
e
r
bl
oc
ks
w
it
h
110
M
p
a
r
a
m
e
te
r
s
a
nd
a
hi
dde
n
s
iz
e
768,
s
e
r
ve
s
a
s
th
e
f
ounda
ti
on.
A
lt
hough
in
he
r
e
nt
ly
m
ul
ti
li
ngua
l,
it
unde
r
goe
s
f
ur
th
e
r
f
in
e
-
tu
ni
ng
f
or
S
w
a
hi
li
,
e
nha
nc
in
g
it
s
s
ui
ta
bi
li
ty
f
or
S
w
a
hi
li
N
L
I
ta
s
ks
.
F
in
e
-
tu
ni
ng
in
vol
ve
s
tr
a
in
in
g
th
e
m
od
e
l
s
pe
c
if
ic
a
ll
y
f
or
S
w
a
hi
li
N
L
I
,
e
na
bl
in
g
it
to
unde
r
s
ta
nd S
w
a
hi
li
nua
nc
e
s
b
e
tt
e
r
a
nd i
m
pr
ove
t
a
s
k pe
r
f
or
m
a
nc
e
.
2.3.
T
r
ai
n
in
g s
e
t
u
p
T
he
tr
a
in
in
g
pr
oc
e
s
s
in
vol
ve
s
two
tr
a
n
s
f
or
m
e
r
-
ba
s
e
d
ne
ur
a
l
n
e
twor
k
a
r
c
hi
te
c
tu
r
e
s
:
S
w
a
h
B
E
R
T
a
nd
m
B
E
R
T
.
T
he
s
e
m
ode
ls
a
r
e
f
in
e
-
tu
ne
d
f
or
th
e
S
w
a
hi
li
N
L
I
ta
s
k
us
in
g
th
e
pr
e
-
pr
oc
e
s
s
e
d
da
ta
.
T
he
m
ode
ls
a
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 2, A
pr
il
2025
:
1597
-
1604
1600
tr
a
in
e
d
f
or
100
e
poc
hs
w
it
h
e
a
r
ly
s
to
ppi
ng
us
in
g
a
ba
tc
h
s
iz
e
of
32
a
nd
th
e
A
da
m
opt
im
iz
e
r
w
i
th
a
le
a
r
ni
n
g
r
a
te
1e
-
5.
W
e
u
s
e
th
e
D
G
X
-
A
100
s
e
r
ve
r
e
qui
ppe
d
w
it
h
T
e
s
la
A
100
S
X
M
4
40
G
B
.
G
P
U
s
,
e
a
c
h
w
it
h
40
G
B
.
of
G
P
U
m
e
m
o
r
y. T
hi
s
s
e
tu
p i
s
i
de
a
l
f
or
e
xe
c
ut
in
g our
m
ode
ls
, w
hi
c
h r
e
qui
r
e
a
ppr
oxi
m
a
te
ly
15 hour
s
t
o r
un.
2.4.
M
od
e
l
e
val
u
at
io
n
T
he
pe
r
f
or
m
a
nc
e
of
th
e
m
ode
ls
is
e
v
a
lu
a
te
d
us
in
g
s
e
pa
r
a
te
t
e
s
t
da
ta
s
e
ts
,
e
m
pl
oyi
ng
m
e
tr
ic
s
s
u
c
h
a
s
a
c
c
ur
a
c
y
(
c
om
put
e
d
in
1)
,
pr
e
c
is
io
n
(
c
om
put
e
d
in
2)
,
r
e
c
a
ll
(
c
om
put
e
d
in
3)
,
a
nd
F
1
-
s
c
or
e
(
c
om
put
e
d
in
4
)
a
c
r
os
s
va
r
io
us
c
l
a
s
s
e
s
in
S
w
a
hi
li
N
L
I
.
T
he
s
e
m
e
tr
ic
s
,
w
hi
c
h
ut
il
iz
e
tr
ue
pos
it
iv
e
(
T
P
)
,
tr
ue
ne
ga
ti
v
e
(
T
N
)
,
f
a
ls
e
pos
it
iv
e
(
F
P
)
,
a
nd
f
a
ls
e
ne
ga
ti
ve
(
F
N
)
,
pr
ovi
de
a
c
om
pr
e
he
ns
iv
e
ove
r
vi
e
w
of
th
e
m
ode
l'
s
s
tr
e
ngt
hs
a
nd
w
e
a
kne
s
s
e
s
a
c
r
os
s
di
f
f
e
r
e
nt
N
L
I
c
a
te
gor
ie
s
.
F
ur
th
e
r
m
or
e
,
m
a
c
r
o
a
nd
w
e
ig
ht
e
d
a
ve
r
a
ge
s
a
r
e
de
r
iv
e
d
to
s
um
m
a
r
iz
e
th
e
ove
r
a
ll
pe
r
f
or
m
a
nc
e
,
of
f
e
r
in
g
va
lu
a
bl
e
in
s
ig
ht
s
in
to
how
e
f
f
e
c
ti
ve
ly
th
e
m
ode
ls
in
te
r
pr
e
t
th
e
r
e
la
ti
ons
hi
ps
be
twe
e
n
s
e
nt
e
nc
e
s
in
S
w
a
hi
li
N
L
I
ta
s
ks
.
R
e
f
e
r
r
in
g
to
s
ta
te
-
of
-
th
e
-
a
r
t
le
a
r
ni
ng
a
lg
or
it
hm
s
s
uc
h
a
s
[
29]
,
[
30]
,
us
in
g
a
c
onf
u
s
io
n
m
a
tr
ix
e
nha
n
c
e
s
th
e
e
va
lu
a
ti
on
pr
oc
e
s
s
by
pr
ovi
di
ng
d
e
e
pe
r
in
s
ig
ht
s
in
to
th
e
m
ode
ls
'
pe
r
f
or
m
a
nc
e
a
c
r
os
s
di
f
f
e
r
e
nt
c
la
s
s
e
s
,
th
e
r
e
by
e
nr
i
c
hi
ng
th
e
de
ve
lo
pm
e
nt
of
N
L
I
m
ode
ls
f
or
lo
w
-
r
e
s
our
c
e
l
a
ngua
ge
s
.
=
(
)
+
(
)
+
(
)
(
)
(
1)
=
+
(
2)
=
+
(
3)
1
−
=
2
∗
∗
+
(
4)
3.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
T
he
s
tu
dy
in
ve
s
ti
ga
te
d
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
a
de
ve
lo
pe
d
m
onol
in
gua
l
m
ode
l
f
or
N
L
I
ta
s
ks
in
S
w
a
hi
li
.
W
hi
le
pr
e
vi
ous
r
e
s
e
a
r
c
h
pr
im
a
r
il
y
f
oc
us
e
d
on
N
L
I
m
ode
ls
w
it
hi
n
m
ul
ti
li
ngua
l
f
r
a
m
e
w
or
ks
,
th
e
y
di
d
not
e
xpl
ic
it
ly
a
ddr
e
s
s
th
e
pe
r
f
or
m
a
nc
e
of
m
onol
in
gua
l
m
ode
ls
in
S
w
a
hi
li
N
L
I
ta
s
ks
.
T
he
ga
p
is
a
ddr
e
s
s
e
d
by
e
va
lu
a
ti
ng
th
e
im
pa
c
t
a
nd
e
f
f
e
c
ti
ve
ne
s
s
of
m
onol
in
gua
l
m
ode
ls
in
S
w
a
hi
li
N
L
I
ta
s
ks
c
om
pa
r
e
d
to
m
ul
ti
li
ngua
l
m
ode
ls
.
T
he
s
tu
dy
r
e
s
ul
t
s
w
il
l
s
he
d
li
ght
on
m
onol
in
gua
l
m
ode
ls
'
pot
e
nt
ia
l
a
dva
nt
a
ge
s
a
nd
li
m
it
a
ti
ons
c
om
pa
r
e
d t
o t
he
ir
m
ul
ti
li
ngua
l
c
ount
e
r
pa
r
ts
f
or
S
w
a
hi
li
N
L
I
t
a
s
ks
.
3.1.
K
e
y f
in
d
in
gs
O
ur
s
tu
dy
r
e
ve
a
le
d
th
a
t
th
e
S
w
a
hi
li
N
L
I
m
onol
in
gua
l
m
ode
l
a
c
hi
e
ve
d
a
n
im
pr
e
s
s
iv
e
a
c
c
ur
a
c
y
of
69%
,
out
pe
r
f
or
m
in
g
th
e
m
ul
ti
li
ngua
l
m
ode
l'
s
a
c
c
ur
a
c
y
of
66
%
,
a
s
di
s
pl
a
ye
d
in
T
a
bl
e
2
.
T
hi
s
in
di
c
a
te
s
a
c
ons
is
te
nt
a
li
gnm
e
nt
be
twe
e
n
th
e
m
ode
l'
s
pr
e
di
c
ti
ons
a
nd
th
e
a
c
tu
a
l
la
be
ls
in
m
os
t
te
s
t
s
a
m
pl
e
s
.
N
ot
a
bl
y,
th
e
m
onol
in
gua
l
m
ode
l
e
xhi
bi
te
d
e
nha
nc
e
d pr
e
c
is
io
n, r
e
c
a
ll
, a
nd F
1 s
c
or
e
s
, e
s
pe
c
ia
ll
y i
n pr
e
di
c
ti
ng c
ont
r
a
di
c
ti
on
a
nd
ne
ut
r
a
li
ty
,
hi
ghl
ig
ht
in
g
it
s
pr
of
ic
ie
nc
y
in
a
c
c
ur
a
te
opt
im
is
ti
c
pr
e
di
c
ti
ons
w
hi
le
m
in
im
iz
in
g
e
r
r
or
s
.
T
he
m
ode
l'
s
c
ons
i
s
te
nt
pe
r
f
or
m
a
nc
e
a
c
r
os
s
m
ic
r
o
a
nd
w
e
ig
ht
e
d
a
ve
r
a
ge
s
s
ugge
s
ts
unbi
a
s
e
d
pr
e
di
c
ti
on
s
una
f
f
e
c
te
d
by
im
ba
la
nc
e
d
da
ta
di
s
tr
ib
ut
io
n.
I
n
c
ont
r
a
s
t,
th
e
m
ul
ti
li
ngua
l
m
ode
l
s
how
e
d
une
xpe
c
te
d
s
tr
e
ngt
h
in
pr
e
di
c
ti
ng
e
nt
a
il
m
e
nt
,
unde
r
s
c
or
in
g
th
e
im
por
ta
nc
e
of
le
ve
r
a
gi
ng
th
e
uni
que
c
a
pa
bi
li
ti
e
s
of
e
a
c
h
m
ode
l.
A
ddi
ti
ona
ll
y,
th
e
c
onf
us
io
n
m
a
tr
ix
(
F
ig
ur
e
2
)
s
how
s
a
hi
ghe
r
num
be
r
of
c
or
r
e
c
tl
y
p
r
e
di
c
te
d
in
s
ta
nc
e
s
in
a
ll
c
la
s
s
e
s
f
or
th
e
m
onol
in
gua
l
m
ode
l
c
om
pa
r
e
d
to
th
e
m
ul
ti
li
ngua
l
m
ode
l,
a
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
2
(
a
)
f
or
th
e
m
onol
in
gua
l
(
S
w
a
hB
E
R
T
)
m
ode
l
a
nd
F
ig
ur
e
2
(
b)
f
or
th
e
m
u
lt
il
in
gua
l
(
m
B
E
R
T
)
m
ode
l.
T
he
s
e
unde
r
s
c
or
e
th
e
im
por
ta
nc
e
of
ta
il
or
e
d
a
ppr
oa
c
h
e
s
f
or
e
nha
n
c
in
g
lo
w
-
r
e
s
our
c
e
la
ngua
ge
pr
oc
e
s
s
in
g,
of
f
e
r
in
g
va
lu
a
bl
e
in
s
ig
ht
s
f
or
de
ve
lo
pi
ng e
f
f
e
c
ti
ve
N
L
P
s
ys
te
m
s
f
or
unde
r
r
e
pr
e
s
e
nt
e
d l
a
ngua
ge
s
l
ik
e
S
w
a
hi
li
.
T
a
bl
e
2
. C
om
pa
r
is
on of
m
onol
in
gua
l
a
nd mul
ti
li
ngua
l
m
ode
ls
A
c
c
ur
a
c
y (
%
)
M
onol
i
ngua
l
m
ode
l
(
S
w
a
hB
E
R
T
)
M
ul
t
i
l
i
ngua
l
m
ode
l
(
m
B
E
R
T
)
69
66
P
r
e
c
i
s
i
on (
%
)
R
e
c
a
l
l
(
%
)
F1
-
S
c
or
e
(
%
)
P
r
e
c
i
s
i
on (
%
)
R
e
c
a
l
l
(
%
)
F1
-
s
c
or
e
(
%
)
E
nt
a
i
l
m
e
nt
71
70
71
81
67
74
C
ont
r
a
di
c
t
i
on
74
70
72
63
71
67
N
e
ut
r
a
l
62
66
64
57
60
58
M
a
c
r
o a
ve
r
a
ge
69
69
69
67
66
66
W
e
i
ght
e
d a
ve
r
a
ge
69
69
69
67
66
66
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
A
c
om
par
at
iv
e
s
tu
dy
of
nat
ur
al
l
anguage
i
nf
e
r
e
nc
e
i
n Sw
ahi
li
us
in
g m
onol
in
gual
…
(
H
aj
r
a F
ak
i
A
li
)
1601
3.2.
C
om
p
ar
at
iv
e
an
al
ys
is
an
d
i
n
t
e
r
p
r
e
t
at
io
n
of
r
e
s
u
lt
s
O
ur
s
tu
dy
c
onduc
te
d
a
de
ta
il
e
d
c
om
pa
r
a
ti
ve
a
na
ly
s
is
be
twe
e
n m
onol
in
gua
l
a
nd
m
ul
ti
li
ngua
l
m
ode
ls
in
S
w
a
hi
li
N
L
I
ta
s
ks
,
r
e
ve
a
li
ng
th
e
ir
uni
que
s
tr
e
ngt
hs
a
nd
w
e
a
kne
s
s
e
s
.
T
he
m
onol
in
gua
l
m
ode
l,
ut
il
iz
in
g
th
e
e
nt
ir
e
S
w
a
hi
li
voc
a
bul
a
r
y,
de
m
ons
tr
a
te
d
a
c
om
pr
e
he
ns
iv
e
un
de
r
s
ta
ndi
ng
of
th
e
la
ngua
ge
,
c
ont
r
a
s
ti
ng
w
it
h
m
ul
ti
li
ngua
l
m
ode
ls
th
a
t
a
ll
oc
a
te
a
s
m
a
ll
e
r
por
ti
on
of
th
e
ir
voc
a
bul
a
r
y
to
S
w
a
hi
li
[
12]
.
T
hi
s
a
ppr
oa
c
h
pot
e
nt
ia
ll
y
boos
ts
pe
r
f
or
m
a
nc
e
in
S
w
a
hi
li
N
L
I
ta
s
ks
,
a
s
e
vi
de
nc
e
d
by
th
e
hi
ghe
r
num
be
r
of
a
c
c
ur
a
te
la
be
ls
id
e
nt
if
ie
d
by
th
e
m
onol
in
gua
l
m
ode
l
c
om
pa
r
e
d
to
th
e
m
ul
ti
li
ngua
l
m
ode
l.
T
he
pe
r
f
or
m
a
nc
e
of
bot
h
m
ode
ls
is
vi
s
ua
ll
y
de
pi
c
te
d
in
th
e
c
onf
us
io
n
m
a
tr
ic
e
s
pr
e
s
e
nt
e
d
in
F
ig
ur
e
2(
a
)
f
or
th
e
m
onol
in
gua
l
(
S
w
a
hB
E
R
T
)
m
ode
l
a
nd F
ig
ur
e
2(
b)
f
or
th
e
m
ul
ti
li
ngua
l
(
m
B
E
R
T
)
m
ode
l.
F
ur
th
e
r
m
or
e
,
a
s
out
li
ne
d
in
T
a
bl
e
2
,
our
de
ta
il
e
d
c
om
pa
r
is
on
unde
r
s
c
or
e
d
th
e
m
onol
in
gua
l
m
ode
l'
s
s
upe
r
io
r
pe
r
f
or
m
a
nc
e
, a
c
hi
e
vi
ng a
n a
c
c
ur
a
c
y s
c
or
e
, m
a
c
r
o a
ve
r
a
ge
, a
nd w
e
ig
ht
e
d a
ve
r
a
ge
of
69%
, m
a
r
gi
na
ll
y
out
pe
r
f
or
m
in
g
th
e
m
ul
ti
li
ngua
l
m
ode
l'
s
66%
.
C
la
s
s
-
w
is
e
c
o
m
pa
r
is
ons
r
e
ve
a
le
d
th
e
m
onol
in
gua
l
m
ode
l'
s
pr
of
ic
ie
nc
y
in
pr
e
di
c
ti
ng
c
ont
r
a
di
c
ti
on
a
nd
ne
ut
r
a
l
la
be
ls
,
a
c
hi
e
vi
ng
th
e
hi
ghe
s
t
pr
e
c
is
io
n,
r
e
c
a
ll
,
a
nd
F1
-
s
c
or
e
s
in
th
e
s
e
c
a
t
e
gor
ie
s
.
D
e
s
pi
te
it
s
e
xc
e
ll
e
n
c
e
,
th
e
m
onol
in
gua
l
m
ode
l
out
pe
r
f
or
m
e
d
th
e
m
ul
ti
li
ngua
l
m
ode
l
in
pr
e
di
c
ti
ng
e
nt
a
il
m
e
nt
,
s
how
c
a
s
in
g
it
s
a
bi
li
ty
to
di
s
c
e
r
n
lo
gi
c
a
l
r
e
la
ti
on
s
hi
ps
a
c
r
os
s
di
ve
r
s
e
la
ngua
ge
s
.
T
h
e
s
e
f
in
di
ngs
e
m
pha
s
i
z
e
th
e
im
por
ta
nc
e
of
a
ba
la
nc
e
d
a
ppr
oa
c
h
th
a
t
le
ve
r
a
ge
s
th
e
s
tr
e
ngt
hs
of
bot
h m
ode
ls
t
o opti
m
iz
e
ove
r
a
ll
pe
r
f
or
m
a
nc
e
w
it
hout
c
om
pr
om
is
in
g a
c
c
ur
a
c
y i
n s
p
e
c
if
ic
t
a
s
k doma
in
s
.
(
a
)
(
b)
F
ig
ur
e
2
. C
onf
us
io
n m
a
tr
ix
f
or
(
a
)
m
onol
in
gua
l
(
S
w
a
hB
E
R
T
)
a
nd (
b)
m
ul
ti
li
ngua
l
(
m
B
E
R
T
)
m
ode
ls
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 2, A
pr
il
2025
:
1597
-
1604
1602
3.3.
A
d
d
r
e
s
s
in
g l
im
it
at
io
n
s
C
om
pa
r
in
g
th
e
m
onol
in
gua
l
a
nd
m
ul
ti
li
ngua
l
m
ode
l
pr
e
di
c
ti
on
s
pr
ovi
de
s
va
lu
a
bl
e
in
s
ig
ht
s
in
to
th
e
ir
m
is
c
la
s
s
if
ic
a
ti
on
s
.
B
ot
h
m
ode
ls
e
nc
ount
e
r
c
ha
ll
e
nge
s
due
to
le
xi
c
a
l
va
r
ia
ti
on,
s
e
m
a
nt
ic
e
qui
va
le
nc
e
,
s
ubt
le
di
f
f
e
r
e
nc
e
s
,
a
nd
que
s
ti
on
f
or
m
a
ti
on,
a
s
il
lu
s
tr
a
te
d
in
T
a
bl
e
3
.
I
n
in
s
ta
nc
e
s
w
h
e
r
e
th
e
pr
e
m
is
e
a
nd
hypothe
s
i
s
di
s
c
us
s
te
c
hni
que
s
ut
il
iz
e
d
by
or
ga
ni
z
a
ti
ons
,
bot
h
m
ode
ls
m
is
ta
ke
nl
y
pr
e
di
c
t
la
be
l
1
(
e
nt
a
il
m
e
nt
)
in
s
te
a
d
of
la
be
l
0
(
c
ont
r
a
di
c
ti
on)
.
T
hi
s
m
is
c
la
s
s
if
ic
a
ti
on
unde
r
s
c
or
e
s
th
e
ir
f
a
il
ur
e
to
di
s
c
e
r
n
th
e
nu
a
nc
e
d
di
f
f
e
r
e
nc
e
be
twe
e
n
"
k
w
a
k
aw
ai
da
"
(
us
u
a
ll
y)
a
nd
"
z
in
az
ot
um
i
w
a
s
ana
"
(
he
a
vi
ly
us
e
d)
,
in
di
c
a
ti
ng
a
li
m
it
a
ti
on
in
c
om
pr
e
he
ndi
ng
c
ont
e
xt
ua
l
nu
a
nc
e
s
.
S
im
il
a
r
ly
,
th
e
m
onol
in
gua
l
m
ode
l
a
de
pt
ly
id
e
nt
if
ie
s
s
e
m
a
nt
ic
e
qui
va
le
nc
e
s
,
hi
ghl
ig
ht
in
g
it
s
pr
of
ic
ie
nc
y
in
c
a
pt
ur
in
g
s
ubt
le
n
ua
nc
e
s
,
w
hi
le
th
e
m
ul
ti
li
ngua
l
m
ode
l
s
tr
uggl
e
s
due
to
li
ngui
s
ti
c
va
r
ia
ti
on
a
c
r
os
s
la
ngua
ge
s
.
M
or
e
ov
e
r
,
th
e
m
ul
ti
li
ngua
l
m
ode
l
e
nc
ount
e
r
s
c
ha
ll
e
nge
s
w
it
h
s
e
nt
e
nc
e
pa
ir
s
in
vol
vi
ng
que
s
ti
on
s
,
w
he
r
e
m
is
c
la
s
s
if
ic
a
ti
on
oc
c
ur
s
due
to
di
f
f
ic
ul
ti
e
s
in
na
vi
ga
ti
ng
li
ngui
s
ti
c
di
f
f
e
r
e
nc
e
s
.
D
e
s
pi
te
th
e
s
e
li
m
it
a
ti
ons
,
th
e
m
onol
in
gua
l
m
o
de
l
ge
ne
r
a
ll
y
out
pe
r
f
or
m
s
,
e
m
pha
s
iz
in
g
th
e
s
ig
ni
f
ic
a
nc
e
of
l
a
ngua
ge
-
s
pe
c
if
ic
c
om
pr
e
he
ns
io
n i
n a
c
c
ur
a
te
pr
e
di
c
ti
ons
a
m
id
s
t
li
ngui
s
ti
c
c
om
pl
e
xi
ti
e
s
.
T
a
bl
e
3
.
I
ll
us
tr
a
ti
ve
e
xa
m
pl
e
s
f
or
m
onol
in
gua
l
a
nd mul
ti
li
ngua
l
m
ode
l
pr
e
di
c
ti
ons
SN
P
r
e
m
i
s
e
H
ypot
he
s
i
s
T
r
ue
l
a
be
l
P
r
e
di
c
t
e
d
m
onol
i
ngua
l
P
r
e
di
c
t
e
d
m
ul
t
i
l
i
ngua
l
1
T
ul
i
j
ar
i
bu
k
ut
am
bua
m
bi
nu
am
baz
o
z
i
l
i
t
e
k
e
l
e
z
w
a
k
w
a
k
aw
ai
da
na
m
as
hi
r
i
k
a
k
at
i
k
a m
uda w
a m
i
ak
a 5
i
l
i
y
opi
t
a
.
W
e
t
r
i
e
d
t
o
i
de
nt
i
f
y
t
he
m
e
t
hods
t
ha
t
w
e
r
e
c
om
m
onl
y
i
m
pl
e
m
e
nt
e
d
by
or
ga
ni
z
a
t
i
ons
dur
i
ng t
he
l
a
s
t
5 ye
a
r
s
.
T
unat
ak
a
k
ut
am
bua
m
bi
nu
z
i
naz
ot
um
i
w
a
s
ana
na
m
as
hi
r
i
k
a
k
at
i
k
a m
i
ak
a 5 i
l
i
y
opi
t
a.
W
e
w
a
nt
t
o
i
de
nt
i
f
y
t
he
m
e
t
hods
m
os
t
us
e
d
by
or
ga
ni
z
a
t
i
ons
i
n
t
he
l
a
s
t
5 ye
a
r
s
.
0
1
1
2
N
i
l
az
i
m
a
m
at
ai
f
a
y
aony
e
s
he
m
ae
nde
l
e
o
y
a
k
ur
i
dhi
s
ha
k
at
i
k
a
m
i
pango
y
ao
y
a
ut
e
k
e
l
e
z
aj
i
y
a
m
aj
i
m
bo
k
u
e
l
e
k
e
a
l
e
ngo
l
i
l
i
l
oam
r
i
s
hw
a
na
bunge
l
a
k
ur
e
j
e
a
hal
i
y
a
as
i
l
i
k
at
i
k
a
m
buga
z
a
k
i
t
ai
f
a
na
m
ae
ne
o y
a ny
i
k
a.
S
t
a
t
e
s
m
us
t
de
m
ons
t
r
a
t
e
s
a
t
i
s
f
a
c
t
or
y
pr
ogr
e
s
s
i
n
t
he
i
r
s
t
a
t
e
a
c
t
i
on
pl
a
ns
t
ow
a
r
d
t
he
c
ongr
e
s
s
i
ona
l
l
y
m
a
nda
t
e
d
go
a
l
of
r
e
s
t
or
i
ng
na
t
ur
a
l
c
ondi
t
i
ons
i
n
na
t
i
ona
l
pa
r
ks
a
nd w
i
l
de
r
ne
s
s
a
r
e
a
s
.
Si
o
l
az
i
m
a
k
w
a
ubor
e
s
haj
i
w
ow
ot
e
.
I
t
i
s
not
ne
c
e
s
s
a
r
y
f
or
a
ny
i
m
pr
ove
m
e
nt
.
2
2
1
3
J
e
, w
al
i
k
uw
a hum
o ndani
?
W
e
r
e
t
he
y i
ns
i
de
?
J
e
, w
al
i
pas
w
a
k
uw
a hum
o ndani
?
W
e
r
e
t
he
y s
uppos
e
d t
o b
e
i
ns
i
de
?
1
1
0
3.4.
I
m
p
li
c
at
io
n
s
f
o
r
f
u
t
u
r
e
r
e
s
e
ar
c
h
an
d
m
od
e
l
r
e
f
in
e
m
e
n
t
T
he
f
in
di
ngs
of
th
is
s
tu
dy
s
ugge
s
t
th
a
t
la
ngua
ge
-
s
pe
c
if
ic
m
ode
ls
a
r
e
c
r
uc
ia
l
f
or
ta
s
ks
li
ke
N
L
I
,
w
he
r
e
la
ngua
ge
nua
n
c
e
s
gr
e
a
tl
y
a
f
f
e
c
t
pe
r
f
or
m
a
nc
e
.
M
ul
ti
li
ng
ua
l
m
ode
ls
m
a
y
not
e
f
f
e
c
ti
ve
ly
c
a
pt
ur
e
th
e
s
e
nua
nc
e
s
,
p
a
r
ti
c
ul
a
r
ly
f
or
lo
w
-
r
e
s
our
c
e
la
ngua
ge
s
.
F
ut
ur
e
r
e
s
e
a
r
c
h
c
oul
d
f
oc
us
on
e
nha
nc
in
g
m
ul
ti
li
ngua
l
m
ode
l
pe
r
f
or
m
a
nc
e
by
in
c
or
por
a
ti
ng
la
ngua
ge
-
s
pe
c
if
ic
pr
e
-
pr
oc
e
s
s
in
g
t
e
c
hni
que
s
or
e
xpl
or
in
g
tr
a
ns
f
e
r
le
a
r
ni
ng f
r
om
l
a
ngua
ge
-
s
pe
c
if
ic
m
ode
ls
. A
ddi
ti
ona
ll
y, i
nve
s
ti
ga
ti
ng hybr
id
m
ode
ls
t
ha
t
c
om
bi
ne
t
he
s
tr
e
ngt
hs
of
m
onol
in
gua
l
a
nd
m
ul
ti
li
ngua
l
a
ppr
oa
c
he
s
c
oul
d
yi
e
ld
p
r
om
is
in
g
r
e
s
ul
ts
in
im
pr
ovi
ng
ove
r
a
ll
m
ode
l
pe
r
f
or
m
a
nc
e
on NL
I
t
a
s
ks
.
F
ur
th
e
r
m
or
e
,
th
e
r
e
is
a
ne
e
d
to
e
nha
n
c
e
e
nt
a
il
m
e
nt
pr
e
d
ic
ti
on
w
it
hi
n
m
onol
in
gua
l
m
ode
ls
,
pa
r
ti
c
ul
a
r
ly
f
or
la
ngua
ge
s
li
ke
S
w
a
hi
li
.
T
hi
s
c
oul
d
in
vol
ve
in
te
gr
a
ti
ng
e
xt
e
r
na
l
knowle
dge
s
our
c
e
s
or
e
xpl
or
in
g
a
lt
e
r
na
ti
ve
ne
ur
a
l
ne
twor
k
a
r
c
hi
te
c
tu
r
e
s
ta
il
or
e
d
f
or
e
nt
a
il
m
e
nt
ta
s
ks
.
F
ur
th
e
r
va
li
da
ti
on
us
in
g
m
e
ti
c
ul
ous
ly
c
ur
a
te
d
da
ta
s
e
ts
is
e
s
s
e
nt
ia
l
to
a
ddr
e
s
s
li
m
it
a
ti
on
s
a
nd
ga
in
a
c
om
pr
e
he
ns
iv
e
unde
r
s
ta
ndi
ng
of
th
e
m
ode
l'
s
c
a
pa
bi
li
ti
e
s
,
pa
r
ti
c
ul
a
r
ly
it
s
im
pa
c
t
on
S
w
a
hi
li
N
L
I
ta
s
ks
.
R
e
c
e
nt
obs
e
r
va
ti
ons
s
ugg
e
s
t
th
a
t
m
onol
in
gua
l
m
ode
ls
of
te
n
out
pe
r
f
o
r
m
m
ul
t
il
in
gua
l
one
s
,
hi
ghl
i
ght
in
g
th
e
im
por
ta
nc
e
of
th
is
li
ne
of
r
e
s
e
a
r
c
h.
O
ur
s
tu
dy
c
ont
r
ib
ut
e
s
to
th
is
di
s
c
our
s
e
by
e
m
pha
s
iz
in
g
th
e
s
upe
r
io
r
pr
e
di
c
ti
on
c
a
pa
bi
li
ty
of
m
onol
in
gua
l
m
ode
ls
, s
uppor
te
d by higher
a
c
tu
a
l
pr
e
di
c
ti
on c
ount
s
a
nd c
om
p
e
ti
ti
ve
a
ve
r
a
ge
s
c
or
e
s
a
c
r
os
s
di
f
f
e
r
e
nt
c
la
s
s
e
s
.
4.
C
O
N
C
L
U
S
I
O
N
T
hi
s
s
tu
dy
hi
ghl
ig
ht
s
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
S
w
a
hB
E
R
T
in
c
a
pt
ur
in
g
th
e
nua
nc
e
s
of
th
e
S
w
a
hi
li
la
ngua
ge
, out
pe
r
f
or
m
in
g m
B
E
R
T
i
n t
he
N
L
I
t
a
s
k. T
hi
s
unde
r
s
c
or
e
s
t
he
pot
e
nt
ia
l
of
l
a
ngua
ge
-
s
pe
c
if
ic
m
ode
ls
li
ke
S
w
a
hB
E
R
T
to
e
nha
n
c
e
N
L
P
a
ppl
ic
a
ti
ons
ta
il
or
e
d
to
lo
w
-
r
e
s
our
c
e
la
ngua
ge
s
.
H
ow
e
ve
r
,
a
ddr
e
s
s
in
g
e
xi
s
ti
ng
li
m
it
a
ti
ons
a
nd
r
e
f
in
in
g
m
ode
l
pe
r
f
o
r
m
a
nc
e
,
pa
r
ti
c
ul
a
r
ly
in
ta
s
ks
r
e
la
te
d
to
e
nt
a
il
m
e
nt
in
te
r
pr
e
ta
ti
on,
r
e
qui
r
e
s
f
ur
th
e
r
r
e
s
e
a
r
c
h
a
nd
de
v
e
lo
pm
e
nt
.
M
ovi
ng
f
or
w
a
r
d,
th
is
r
e
s
e
a
r
c
h
w
il
l
f
oc
u
s
on
r
e
f
in
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
A
c
om
par
at
iv
e
s
tu
dy
of
nat
ur
al
l
anguage
i
nf
e
r
e
nc
e
i
n Sw
ahi
li
us
in
g m
onol
in
gual
…
(
H
aj
r
a F
ak
i
A
li
)
1603
S
w
a
hB
E
R
T
'
s
p
e
r
f
or
m
a
nc
e
by
e
xpl
or
in
g
s
pe
c
ia
li
z
e
d
da
ta
s
e
t
s
a
nd
m
e
th
odol
ogi
e
s
to
be
tt
e
r
a
ddr
e
s
s
th
e
li
ngui
s
ti
c
c
om
pl
e
xi
ti
e
s
in
he
r
e
nt
in
lo
w
-
r
e
s
our
c
e
la
ngua
ge
s
li
ke
S
w
a
hi
li
.
A
ddi
ti
ona
ll
y,
p
r
om
ot
in
g
in
c
lu
s
iv
it
y
in
la
ngua
ge
pr
oc
e
s
s
in
g
te
c
hnol
ogi
e
s
f
or
lo
w
-
r
e
s
our
c
e
la
n
gua
ge
s
r
e
m
a
in
s
a
pr
io
r
it
y.
U
ti
li
z
in
g
bot
h
m
onol
in
gua
l
a
nd
m
ul
ti
li
ngua
l
a
ppr
oa
c
h
e
s
,
th
is
r
e
s
e
a
r
c
h
a
i
m
s
to
de
v
e
lo
p
m
or
e
r
obus
t
N
L
P
to
ol
s
f
or
S
w
a
hi
li
-
s
pe
a
ki
ng
c
om
m
uni
ti
e
s
a
nd
a
dva
nc
e
lo
w
-
r
e
s
our
c
e
la
ng
ua
ge
pr
oc
e
s
s
in
g
gl
oba
ll
y,
in
c
lu
di
ng
e
xpl
or
in
g
a
dva
nc
e
d t
e
c
hni
que
s
l
ik
e
f
uz
z
y l
ogi
c
.
A
C
K
N
O
WL
E
D
G
E
M
E
N
T
S
T
hi
s
r
e
s
e
a
r
c
h
is
s
uppor
te
d
by
in
te
r
na
l
f
undi
ng
f
r
om
th
e
F
a
c
ul
ty
of
C
om
put
e
r
S
c
ie
nc
e
,
U
ni
ve
r
s
it
a
s
I
ndone
s
ia
,
a
nd
th
e
hi
gh
-
pe
r
f
or
m
a
nc
e
c
om
put
in
g
f
a
c
il
it
ie
s
a
t
th
e
T
okope
di
a
-
U
I
A
I
C
e
nt
e
r
,
U
ni
ve
r
s
it
a
s
I
ndone
s
ia
. T
he
a
ut
hor
s
w
oul
d l
ik
e
t
o t
ha
nk a
ll
of
t
he
m
f
or
t
he
ir
s
uppor
t.
R
E
F
E
R
E
N
C
E
S
[
1]
C
.
S
.
S
hi
ka
l
i
a
nd
R
.
M
okhos
i
,
“
E
nha
nc
i
ng
A
f
r
i
c
a
n
l
ow
-
r
e
s
our
c
e
l
a
ngua
ge
s
:
S
w
a
hi
l
i
da
t
a
f
or
l
a
ngua
ge
m
ode
l
l
i
ng,”
D
at
a
i
n
B
r
i
e
f
,
vol
. 31, 2020, doi
:
10.1016/
j
.di
b.2020.105951.
[
2]
M
. J
. R
obi
ns
on,
A
l
anguage
f
or
t
he
w
or
l
d:
t
he
s
t
andar
di
z
at
i
on of
Sw
ahi
l
i
. A
t
he
ns
, O
hi
o:
O
hi
o U
ni
ve
r
s
i
t
y P
r
e
s
s
, 2022.
[
3]
A
.
N
.
N
w
a
m
m
uo
a
nd
A
.
S
a
l
a
w
u,
“
A
r
e
r
a
di
o
pr
og
r
a
m
m
e
s
vi
a
i
ndi
ge
nous
l
a
ngua
ge
s
t
he
s
ol
ut
i
on?
A
s
t
udy
of
I
gbo
s
c
hol
a
r
s
’
a
s
s
e
s
s
m
e
nt
of
t
he
e
f
f
e
c
t
i
ve
ne
s
s
of
t
he
B
r
i
t
i
s
h
B
r
oa
dc
a
s
t
i
ng
C
or
por
a
t
i
on
(
B
B
C
)
i
n
pr
om
ot
i
ng
A
f
r
i
c
a
n
l
a
ngua
ge
s
,”
A
f
r
i
c
an
R
e
nai
s
s
anc
e
, vol
. 16, no. 1, pp. 83
–
99, 2019, doi
:
10.31920/
2516
-
5305/
2019/
V
16n1a
5.
[
4]
P
.
M
pof
u,
I
.
A
.
F
a
di
pe
,
a
nd
T
.
T
s
ha
ba
ngu,
I
ndi
ge
nou
s
A
f
r
i
c
an
L
anguage
M
e
di
a
.
S
i
nga
por
e
:
S
pr
i
nge
r
N
a
t
ur
e
S
i
nga
por
e
,
2023
,
doi
:
10.1007/
978
-
981
-
99
-
0305
-
4.
[
5]
A
.
O
w
oye
m
i
,
J
.
O
w
oye
m
i
,
A
.
O
s
i
ye
m
i
,
a
nd
A
.
B
oyd,
“
A
r
t
i
f
i
c
i
a
l
i
nt
e
l
l
i
ge
n
c
e
f
or
he
a
l
t
hc
a
r
e
i
n
A
f
r
i
c
a
,”
F
r
ont
i
e
r
s
i
n
D
i
gi
t
al
H
e
al
t
h
, vol
. 2, 2020, doi
:
10.3389/
f
dgt
h.2020.00006.
[
6]
C
. H
ol
s
t
e
t
al
.
,
“
D
e
ve
l
opm
e
nt
of
di
gi
t
a
l
he
a
l
t
h
m
e
s
s
a
ge
s
f
or
r
ur
a
l
popul
a
t
i
ons
i
n T
a
nz
a
ni
a
:
m
ul
t
i
-
a
nd i
nt
e
r
di
s
c
i
pl
i
na
r
y a
ppr
oa
c
h,
”
J
M
I
R
m
H
e
al
t
h and uH
e
al
t
h
, vol
. 9, no. 9, 2021, doi
:
10.2196/
25558.
[
7]
B
.
M
i
ni
xhof
e
r
,
F
.
P
a
i
s
c
he
r
,
a
nd
N
.
R
e
ka
bs
a
z
,
“
W
E
C
H
S
E
L
:
E
f
f
e
c
t
i
ve
i
ni
t
i
a
l
i
z
a
t
i
on
of
s
ubw
or
d
e
m
be
ddi
ngs
f
or
c
r
os
s
-
l
i
ngua
l
t
r
a
ns
f
e
r
of
m
onol
i
ngua
l
l
a
ngua
ge
m
ode
l
s
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
2022
C
onf
e
r
e
nc
e
of
t
he
N
or
t
h
A
m
e
r
i
c
an
C
hapt
e
r
of
t
he
A
s
s
oc
i
at
i
on
f
or
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
:
H
um
an
L
anguage
T
e
c
hnol
ogi
e
s
,
2022,
pp.
3992
–
4006
,
doi
:
10.18653/
v1/
2022.na
a
c
l
-
m
a
i
n.293.
[
8]
K
.
O
gue
j
i
,
“
A
f
r
i
B
E
R
T
a
:
T
ow
a
r
ds
vi
a
bl
e
m
ul
t
i
l
i
ngua
l
l
a
ngua
ge
m
ode
l
s
f
or
l
ow
-
r
e
s
our
c
e
l
a
ngua
ge
s
,”
M
.Sc
.
T
he
s
i
s
,
D
e
pa
r
t
m
e
nt
of
M
a
t
he
m
a
t
i
c
s
,
U
ni
ve
r
s
i
t
y of
W
a
t
e
r
l
oo, W
a
t
e
r
l
oo, C
a
na
da
, 2022.
[
9]
G
.
L
.
M
a
r
t
i
n,
M
.
E
.
M
s
w
a
hi
l
i
,
a
nd
Y
.
-
S
.
J
e
ong,
“
S
e
nt
i
m
e
nt
c
l
a
s
s
i
f
i
c
a
t
i
on
i
n
S
w
a
hi
l
i
l
a
ngua
ge
us
i
ng
m
ul
t
i
l
i
ngua
l
B
E
R
T
,”
ar
X
i
v
-
C
om
put
e
r
Sc
i
e
n
c
e
, pp. 1
–
5, 2021.
[
10]
Y
. X
i
a
o
e
t
al
.
, “
A
r
e
B
E
R
T
f
a
m
i
l
y good i
ns
t
r
uc
t
i
on f
ol
l
ow
e
r
s
?
A
s
t
udy on t
he
i
r
pot
e
nt
i
a
l
a
nd l
i
m
i
t
a
t
i
ons
,”
i
n
I
C
L
R
2024
, pp. 1
–
20
,
2024
.
[
11]
K
.
A
l
na
j
j
a
r
a
nd
M
.
H
ä
m
ä
l
ä
i
n
e
n,
“
H
a
r
ne
s
s
i
ng
m
ul
t
i
l
i
ngua
l
r
e
s
our
c
e
s
t
o
qu
e
s
t
i
on
a
ns
w
e
r
i
ng
i
n
A
r
a
bi
c
,”
ar
X
i
v
-
C
om
put
e
r
Sc
i
e
nc
e
,
pp. 1
–
6, 2022.
[
12]
S
.
W
u
a
nd
M
.
D
r
e
dz
e
,
“
A
r
e
a
l
l
l
a
ngua
ge
s
c
r
e
a
t
e
d
e
qua
l
i
n
m
ul
t
i
l
i
ngua
l
B
E
R
T
?
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
5t
h
W
or
k
s
hop
on
R
e
pr
e
s
e
nt
at
i
on L
e
ar
ni
ng f
or
N
L
P
, 2020, pp. 120
–
130
, doi
:
10.18653/
v1/
2020.r
e
pl
4nl
p
-
1.16.
[
13]
G
.
M
a
r
t
i
n,
M
.
E
.
M
s
w
a
hi
l
i
,
Y
.
-
S
.
J
e
ong,
a
nd
J
.
Y
oung
-
S
e
ob,
“
S
w
a
hB
E
R
T
:
l
a
ngua
ge
m
ode
l
of
S
w
a
hi
l
i
,”
i
n
P
r
oc
e
e
di
ng
s
of
t
he
2022
C
onf
e
r
e
nc
e
of
t
he
N
or
t
h
A
m
e
r
i
c
an
C
hapt
e
r
of
t
he
A
s
s
oc
i
at
i
on
f
or
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
:
H
um
an
L
anguage
T
e
c
hnol
ogi
e
s
, 2022, pp. 303
–
313
, doi
:
10.18653/
v1/
2022.na
a
c
l
-
m
a
i
n.23.
[
14]
Z
.
W
a
ng,
L
.
L
i
,
a
nd
D
.
Z
e
ng,
“
K
now
l
e
dge
-
e
nha
nc
e
d
na
t
ur
a
l
l
a
ngua
ge
i
nf
e
r
e
nc
e
ba
s
e
d
on
know
l
e
dge
gr
a
phs
,”
i
n
P
r
oc
e
e
di
ng
s
of
t
he
28t
h I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
, 2020, pp. 649
8
–
6508
, doi
:
10.18653/
v1/
2020.c
ol
i
ng
-
m
a
i
n.571.
[
15]
A
.
D
e
,
M
.
S
.
D
e
s
a
r
ka
r
,
a
nd
A
.
E
kba
l
,
“
T
ow
a
r
ds
i
m
pr
ove
m
e
nt
of
gr
ounde
d
c
r
os
s
-
l
i
ngua
l
na
t
ur
a
l
l
a
ngua
g
e
i
nf
e
r
e
nc
e
w
i
t
h
V
i
s
i
oT
e
xt
ua
l
a
t
t
e
nt
i
on,”
N
at
ur
al
L
anguage
P
r
oc
e
s
s
i
ng J
ou
r
nal
, vol
. 4, 2023, doi
:
10.1016/
j
.nl
p.2023.100023.
[
16]
Q
.
W
u,
P
.
W
a
ng,
X
.
W
a
ng,
X
.
H
e
,
a
nd
W
.
Z
hu,
“
Q
ue
s
t
i
on
a
ns
w
e
r
i
ng
(
Q
A
)
ba
s
i
c
s
,”
i
n
A
dv
anc
e
s
i
n
C
om
put
e
r
V
i
s
i
on
and
P
at
t
e
r
n
R
e
c
ogni
t
i
on
, S
pr
i
nge
r
, S
i
nga
po
r
e
, 2022, pp. 27
–
31
, doi
:
10.1007/
978
-
981
-
19
-
0
964
-
1_3.
[
17]
V
.
A
ga
t
e
,
S
.
M
i
r
a
j
ka
r
,
a
nd
G
.
T
or
a
dm
a
l
,
“
B
ook
s
um
m
a
r
i
z
a
t
i
on
us
i
ng
N
L
P
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
I
nnov
at
i
v
e
R
e
s
e
ar
c
h
i
n
E
ngi
ne
e
r
i
ng
, vol
. 11, no. 4, pp. 476
–
480, 2023, doi
:
10.59256/
i
j
i
r
e
.2023040218
.
[
18]
W
.
Z
hou,
“
R
e
s
e
a
r
c
h
on
i
nf
or
m
a
t
i
on
e
xt
r
a
c
t
i
on
t
e
c
hnol
ogy
a
ppl
i
e
d
f
or
k
now
l
e
dge
gr
a
phs
,”
A
ppl
i
e
d
and
C
om
put
at
i
onal
E
ngi
ne
e
r
i
ng
, vol
. 4, no. 1, pp. 26
–
31, 2023, doi
:
10.54254/
2755
-
2721/
4/
202303
40.
[
19]
H
.
T
a
nvi
r
,
C
.
K
i
t
t
a
s
k,
S
.
E
i
c
he
,
a
nd
K
.
S
i
r
t
s
,
“
E
s
t
B
E
R
T
:
A
pr
e
t
r
a
i
ne
d
l
a
ngua
ge
-
s
pe
c
i
f
i
c
B
E
R
T
f
or
E
s
t
oni
a
n,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
23r
d N
or
di
c
C
onf
e
r
e
nc
e
on C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
(
N
o
D
aL
i
D
a)
, 2021, pp. 11
–
19.
[
20]
A
.
V
e
l
a
nka
r
,
H
.
P
a
t
i
l
,
a
nd
R
.
J
os
hi
,
“
M
ono
vs
m
ul
t
i
l
i
ngua
l
B
E
R
T
f
or
ha
t
e
s
pe
e
c
h
de
t
e
c
t
i
on
a
nd
t
e
xt
c
l
a
s
s
i
f
i
c
a
t
i
on:
a
c
a
s
e
s
t
u
dy
i
n
M
a
r
a
t
hi
,”
i
n
A
r
t
i
f
i
c
i
al
N
e
ur
al
N
e
t
w
or
k
s
i
n
P
at
t
e
r
n
R
e
c
ogni
t
i
on
(
A
N
N
P
R
20
22)
,
2023,
pp.
121
–
128
,
doi
:
10.1007/
978
-
3
-
031
-
20650
-
4_10.
[
21]
M
.
S
t
r
a
ka
,
J
.
N
á
pl
a
va
,
J
.
S
t
r
a
ková
,
a
nd
D
.
S
a
m
ue
l
,
“
R
obe
C
z
e
c
h:
C
z
e
c
h
R
oB
E
R
T
a
,
a
m
onol
i
ngua
l
c
ont
e
xt
ua
l
i
z
e
d
l
a
ngua
ge
r
e
pr
e
s
e
nt
a
t
i
on m
ode
l
,”
i
n
T
e
x
t
, Spe
e
c
h, and D
i
al
ogue
(
T
SD
2021)
, 2021, pp. 197
–
209
, doi
:
10.1007/
978
-
3
-
030
-
83527
-
9_17.
[
22]
D
.
Q
.
N
guye
n
a
nd
A
.
T
.
N
guye
n,
“
P
hoB
E
R
T
:
P
r
e
-
t
r
a
i
ne
d
l
a
ngua
ge
m
ode
l
s
f
or
V
i
e
t
na
m
e
s
e
,”
i
n
F
i
ndi
ngs
of
t
he
A
s
s
oc
i
at
i
on
f
or
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
:
E
M
N
L
P
2020
, 2020, pp. 1037
–
1042
, doi
:
10.18653
/
v1/
2020.f
i
ndi
ngs
-
e
m
nl
p.92.
[
23]
K
.
J
a
i
n,
A
.
D
e
s
hpa
nde
,
K
.
S
hr
i
dha
r
,
F
.
L
a
um
a
nn,
a
nd
A
.
D
a
s
h,
“
I
ndi
c
-
t
r
a
ns
f
or
m
e
r
s
:
a
n
a
na
l
ys
i
s
of
t
r
a
ns
f
or
m
e
r
l
a
ngua
ge
m
ode
l
s
f
or
I
ndi
a
n l
a
ngua
ge
s
,”
ar
X
i
v
-
C
om
put
e
r
Sc
i
e
nc
e
, pp. 1
–
14, N
ov. 2020
.
[
24]
R
.
S
c
he
i
bl
e
,
F
.
T
hom
c
z
yk,
P
.
T
i
ppm
a
nn,
V
.
J
a
r
a
vi
ne
,
a
nd
M
.
B
oe
ke
r
,
“
G
ot
t
B
E
R
T
:
a
pur
e
G
e
r
m
a
n
l
a
ngua
ge
m
ode
l
,
”
ar
X
i
v
-
C
om
put
e
r
Sc
i
e
nc
e
, pp. 1
–
6, 2020.
[
25]
A
.
F
.
M
.
D.
P
a
ul
a
a
nd
I
.
B
.
S
c
hl
i
c
ht
,
“
A
I
-
U
P
V
a
t
I
be
r
L
E
F
-
2021
D
E
T
O
X
I
S
t
a
s
k:
t
oxi
c
i
t
y
de
t
e
c
t
i
on
i
n
i
m
m
i
gr
a
t
i
on
-
r
e
l
a
t
e
d
w
e
b
ne
w
s
c
om
m
e
nt
s
u
s
i
ng t
r
a
ns
f
or
m
e
r
s
a
nd s
t
a
t
i
s
t
i
c
a
l
m
ode
l
s
,”
a
r
X
i
v
-
C
om
put
e
r
Sc
i
e
nc
e
, pp. 1
–
20, N
ov. 2021
.
[
26]
S
.
W
u
a
nd
M
.
D
r
e
dz
e
,
“
B
e
t
o,
B
e
nt
z
,
B
e
c
a
s
:
T
he
s
ur
pr
i
s
i
ng
c
r
os
s
-
l
i
ngua
l
e
f
f
e
c
t
i
ve
ne
s
s
of
B
E
R
T
,”
ar
X
i
v
-
C
om
put
e
r
S
c
i
e
nc
e
,
pp. 1
–
12, 2019.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 2, A
pr
il
2025
:
1597
-
1604
1604
[
27]
A
.
K
.
U
pa
dhya
y
a
nd
H
.
K
.
U
pa
dhya
,
“
X
N
L
I
2.0:
I
m
pr
ovi
ng
X
N
L
I
da
t
a
s
e
t
a
nd
pe
r
f
or
m
a
nc
e
on
c
r
os
s
l
i
ngua
l
unde
r
s
t
a
ndi
ng
(
X
L
U
)
,”
i
n
2023
I
E
E
E
8t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
f
or
C
onv
e
r
g
e
nc
e
i
n
T
e
c
hnol
ogy
(
I
2C
T
)
,
2023,
pp.
1
–
6
,
doi
:
10.1109/
I
2C
T
57861.2023.10126332.
[
28]
M
. S
a
da
t
a
nd
C
. C
a
r
a
ge
a
, “
L
e
a
r
ni
ng t
o i
nf
e
r
f
r
om
unl
a
be
l
e
d da
t
a
:
a
s
e
m
i
-
s
up
e
r
vi
s
e
d l
e
a
r
ni
ng a
ppr
oa
c
h f
or
r
obus
t
na
t
ur
a
l
l
a
ngua
g
e
i
nf
e
r
e
nc
e
,”
i
n
F
i
ndi
ngs
of
t
he
A
s
s
oc
i
at
i
on
f
or
C
om
put
at
i
onal
L
i
ngui
s
t
i
c
s
:
E
M
N
L
P
2022
,
2022,
pp.
4763
–
4776
,
doi
:
10.18653/
v1/
2022.f
i
ndi
ngs
-
e
m
nl
p.351.
[
29]
J
.
D
.
L
.
C
.
N
t
i
vugur
uz
w
a
a
nd
T
.
A
hm
a
d,
“
A
c
onvol
ut
i
ona
l
n
e
ur
a
l
ne
t
w
or
k
t
o
de
t
e
c
t
pos
s
i
bl
e
hi
dde
n
da
t
a
i
n
s
pa
t
i
a
l
dom
a
i
n
i
m
a
ge
s
,”
C
y
b
e
r
s
e
c
ur
i
t
y
, vol
. 6, no. 1, 2023, doi
:
10.1186/
s
42400
-
023
-
00156
-
x.
[
30]
N
.
J
.
D
.
L
.
C
r
oi
x
a
nd
T
.
A
hm
a
d,
“
T
ow
a
r
d
s
e
c
r
e
t
da
t
a
l
oc
a
t
i
on
vi
a
f
uz
z
y
l
o
gi
c
a
nd
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k,”
E
gy
pt
i
an
I
nf
or
m
at
i
c
s
J
our
nal
, vol
. 24, no. 3, 2023, doi
:
10.1016/
j
.e
i
j
.2023.05.010.
B
I
O
G
R
A
P
H
I
E
S
O
F
A
U
T
H
O
R
S
Hajra
Faki
Ali
is
advancing
her
computer
science
knowledge
through
her
master'
s
degree
at
the
Universitas
Indonesia.
Her
academic
journey
be
gan
at
the
University
o
f
Dar
es
Salaam,
Tanzania,
where
she
obtained
her
b
achelor
of
science
in
computer
science
in
2017.
Her
scientific
interests
are
deeply
rooted
in
machine
learning,
with
a
particular
focus
on
natural
language
processing.
Her
research
work
includes
an
intri
guing
study
on
natural
language
inferenc
e
in
low
-
resource
languages
such
as
Swahili.
She
ca
n
be
contacted
at
email:
hajra.ali@
judiciaryzanzibar.go.tz.
Adila
Alfa
Krisnadhi
received
a
B.S.
in
computer
science
fro
m
Universitas
Indonesia
in
2002,
an
M.Sc.
in
computat
ional
logic
from
Technisc
he
Universität
Dresden,
Germany
in
2007,
and
a
Ph.D.
in
computer
science
from
Wright
St
ate
University,
USA
in
2015.
He
joined
as
a
permanent
faculty
member
at
the
Faculty
of
Computer
Science
,
Universitas
Indonesia
in
2007.
Since
2019,
he
also
held
a
position
as
a
co
-
director
of
th
e
Tokopedia
-
UI
AI
Center
of
Excellence.
He
has
authored/co
-
a
uthored
more
than
80
publications
in
the
area
of
semantic
technologies,
knowledge
repre
sentation,
and
machine
learning.
His
resear
ch
interests
include
knowledge
graph
constr
uction,
reasoning
and
questio
n
answering
over
knowledg
e
graphs
and
ontolo
gies,
as
well
as
kn
owledge
modelli
ng
and
engineerin
g. He can b
e contacted
at email
: adila@
cs.ui.ac.i
d.
Evaluation Warning : The document was created with Spire.PDF for Python.