I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
,
p
p
.
1
2
7
2
~
1
2
8
9
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v15
i
3
.
pp
1
2
7
2
-
1
2
8
9
1272
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
Bridg
ing
t
he
li
ng
uistic divid
e
:
re
ce
nt
dev
elo
pmen
ts i
n ma
chine
trans
la
tion f
o
r In
dia
n
l
a
ng
ua
g
es
J
a
y
a
na
nd
A.
K
a
m
ble,
Sh
iv
a
j
ira
o
M
.
J
a
dh
a
v
,
Vino
d J
.
K
a
da
m
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
m
a
t
i
o
n
Te
c
h
n
o
l
o
g
y
,
D
r
.
B
a
b
a
s
a
h
e
b
A
m
b
e
d
k
a
r
Te
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
si
t
y
,
L
o
n
e
r
e
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
an
2
1
,
2
0
2
5
R
ev
is
ed
Ma
y
1
9
,
2
0
2
6
Acc
ep
ted
J
u
l 5
,
2
0
2
6
S
ig
n
ifi
c
a
n
t
a
d
v
a
n
c
e
s
h
a
v
e
b
e
e
n
a
c
h
iev
e
d
i
n
m
a
c
h
in
e
tran
sla
ti
o
n
(M
T)
in
re
c
e
n
t
ti
m
e
s,
p
a
rti
c
u
larly
sta
te
-
of
-
th
e
-
a
rt
(S
OTA)
m
o
d
e
ls
f
o
r
lan
g
u
a
g
e
s
li
k
e
En
g
li
s
h
a
n
d
In
d
ian
h
a
v
in
g
d
isti
n
c
t
g
ra
m
m
a
ti
c
a
l
stru
c
tu
re
s
a
n
d
li
m
it
e
d
m
o
n
o
li
n
g
u
a
l
trai
n
in
g
d
a
ta.
T
h
is
p
a
p
e
r
a
n
a
ly
se
s
v
a
r
io
u
s
re
c
e
n
t
sta
te
-
of
-
th
e
-
a
rt
v
a
rian
ts
o
f
larg
e
la
n
g
u
a
g
e
m
o
d
e
ls
(LL
M
s)
a
n
d
n
e
u
ra
l
m
a
c
h
in
e
t
ra
n
sla
ti
o
n
(NMT
)
f
o
r
I
n
d
ian
lan
g
u
a
g
e
s
i
n
c
o
m
p
a
riso
n
t
o
sta
ti
st
ica
l
m
a
c
h
in
e
t
ra
n
sla
ti
o
n
(
S
M
T
)
.
I
t
t
a
c
k
l
e
s
k
e
y
q
u
e
s
t
i
o
n
s
,
s
u
c
h
a
s
i
d
i
o
m
a
t
i
c
e
x
p
r
e
s
s
i
o
n
s
,
m
o
r
p
h
o
l
o
g
i
c
a
l
l
y
c
o
m
p
lex
g
ra
m
m
a
r
o
r
t
h
e
sc
a
rc
e
n
e
ss
o
f
p
a
ra
ll
e
l
c
o
rp
o
ra
.
F
u
rt
h
e
rm
o
re
,
it
stu
d
ies
b
y
tew
ise
BP
E,
c
o
m
p
a
re
s
t
ra
n
sla
ti
o
n
m
o
d
e
ls
in
term
s
o
f
BL
EU
sc
o
re
s
u
sin
g
se
p
a
ra
te
a
n
d
sh
a
re
d
-
v
o
c
a
b
u
lary
re
p
re
se
n
tati
o
n
wit
h
c
o
p
y
a
c
ti
o
n
s
b
e
twe
e
n
th
e
BP
E
tran
sla
ti
o
n
s,
a
n
d
a
n
a
ly
s
e
s
h
o
w
m
u
l
ti
tas
k
lea
rn
in
g
(Ca
ru
a
n
a
(1
9
9
7
))
a
n
d
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
s
c
a
n
c
o
n
tri
b
u
te
t
o
t
h
e
q
u
a
li
t
y
o
f
tran
sla
ti
o
n
.
In
s
u
m
m
a
ry
,
it
p
ro
v
id
e
s
d
irec
ti
o
n
s
f
o
r
fu
t
u
re
wo
r
k
b
y
s
u
g
g
e
sti
n
g
n
e
w
a
v
e
n
u
e
s
o
f
re
se
a
rc
h
,
i
n
c
lu
d
in
g
b
e
tt
e
r
c
u
ra
te
d
d
a
tas
e
ts,
m
o
r
e
e
fficie
n
t
a
p
p
ro
a
c
h
e
s fo
r
lo
w
-
re
so
u
rc
e
lan
g
u
a
g
e
s
,
a
n
d
c
u
lt
u
ra
ll
y
a
wa
re
tran
sl
a
ti
o
n
s.
K
ey
w
o
r
d
s
:
B
y
te
p
air
en
co
d
i
n
g
I
n
d
ian
lan
g
u
ag
es
L
ar
g
e
lan
g
u
ag
e
m
o
d
els
Ma
ch
in
e
tr
an
s
latio
n
Neu
r
al
m
ac
h
in
e
tr
a
n
s
latio
n
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
J
ay
an
an
d
A.
Kam
b
le
Dep
ar
tm
en
t o
f
I
n
f
o
r
m
atio
n
T
e
ch
n
o
lo
g
y
,
Dr
.
B
ab
asah
eb
Am
b
ed
k
ar
T
ec
h
n
o
lo
g
ical
Un
iv
er
s
ity
L
o
n
er
e,
I
n
d
ia
E
m
ail:
jk
am
b
le@
m
g
m
u
.
ac
.
in
1.
I
NT
RO
D
UCT
I
O
N
Fo
r
a
lin
g
u
is
tically
p
l
u
r
alis
tic
n
atio
n
s
u
ch
as
I
n
d
ia,
m
ac
h
i
n
e
tr
an
s
latio
n
(
MT
)
is
ess
en
tial
to
k
ee
p
v
ar
io
u
s
lan
g
u
ag
es
i
n
in
ter
m
e
d
iar
y
co
m
m
u
n
icatio
n
.
T
r
a
n
s
latio
n
b
etwe
en
I
n
d
ian
lan
g
u
a
g
es
is
ad
v
an
cin
g
m
o
r
e
q
u
ick
ly
,
n
o
t
ju
s
t
u
s
in
g
Go
o
g
le
T
r
an
s
late
b
u
t
also
with
o
th
er
p
o
p
u
lar
MT
s
er
v
ices.
T
h
is
task
h
as
b
ec
o
m
e
m
o
r
e
im
p
o
r
tan
t in
t
h
e
ad
v
e
n
t
o
f
d
ig
i
tal
co
m
m
u
n
icatio
n
[
1
]
.
E
n
g
lis
h
is
a
W
est
Ger
m
an
ic
l
an
g
u
ag
e
wh
ich
was
f
ir
s
t
s
p
o
k
en
in
ea
r
ly
m
ed
iev
al
E
n
g
lan
d
an
d
is
n
o
w
a
lead
in
g
lan
g
u
ag
e
u
s
ed
ar
o
u
n
d
th
e
wo
r
ld
[
2
]
.
I
t
h
as
b
ec
o
m
e
a
lin
g
u
a
f
r
a
n
ca
i
n
p
o
p
u
lar
v
er
n
ac
u
lar
,
s
cien
ce
,
an
d
co
m
m
er
ce
,
as
well
a
s
am
o
n
g
s
ee
r
s
o
r
s
ag
es
[
3
]
.
T
h
er
e
f
o
r
e,
th
e
in
ter
n
atio
n
al
o
r
g
an
is
atio
n
s
,
s
u
ch
as
th
e
E
U,
NAT
O
an
d
th
e
UN
all
u
s
e
E
n
g
lis
h
[
4
]
.
E
n
g
lis
h
is
th
e
w
o
r
ld
’
s
m
o
s
t
-
lear
n
t
f
o
r
eig
n
lan
g
u
ag
e
,
with
s
o
m
e
one
b
illi
o
n
p
eo
p
le
lea
r
n
in
g
it a
t a
n
y
tim
e.
T
h
e
p
leth
o
r
a
o
f
lan
g
u
ag
es
an
d
d
ialec
ts
s
p
o
k
e
n
in
I
n
d
ia
(
w
ith
2
2
lan
g
u
a
g
es
r
ec
o
g
n
ized
as
o
f
f
icial)
also
co
n
s
titu
tes
a
v
er
y
s
tr
o
n
g
m
o
tiv
atio
n
f
o
r
th
e
d
e
v
elo
p
m
e
n
t
o
f
E
n
g
lis
h
to
I
n
d
ian
la
n
g
u
a
g
e
MT
s
y
s
tem
s
[
5
]
.
Alth
o
u
g
h
E
n
g
lis
h
is
n
o
w
u
s
ed
as
a
g
lo
b
al
lin
g
u
a
f
r
an
ca
[
6
]
,
an
d
is
th
e
p
r
im
ar
y
m
o
d
e
o
f
ed
u
ca
ti
on
[
7
]
,
b
u
s
in
ess
[
8
]
an
d
g
o
v
e
r
n
m
e
n
t
ad
m
in
is
tr
atio
n
[
9
]
in
I
n
d
ia
,
m
an
y
p
eo
p
le
s
till
p
r
ef
er
to
s
p
ea
k
th
eir
lo
ca
l
lan
g
u
ag
e.
Stro
n
g
MT
s
y
s
tem
s
ca
n
h
elp
f
ill
th
e
ex
is
tin
g
co
m
m
u
n
icatio
n
b
r
ea
ch
a
n
d
allo
w
ac
ce
s
s
to
o
p
p
o
r
tu
n
ities
,
s
er
v
ices,
an
d
in
f
o
r
m
atio
n
in
lo
ca
l la
n
g
u
ag
es.
I
n
f
ield
s
s
u
c
h
as
ed
u
ca
tio
n
[
10
]
,
g
o
v
er
n
m
en
t
[
11
]
,
h
ea
lth
ca
r
e
[
12
]
,
an
d
tech
n
o
lo
g
y
[
13
]
an
d
s
o
f
o
r
th
,
good
MT
h
as
b
ee
n
in
a
g
r
o
win
g
d
em
a
n
d
.
I
n
t
h
e
teac
h
in
g
d
o
m
ain
,
MT
ca
n
b
e
u
s
ed
to
ass
is
t
s
tu
d
en
ts
th
at
d
o
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
B
r
id
g
in
g
th
e
lin
g
u
is
tic
d
ivid
e:
r
ec
en
t d
ev
elo
p
men
ts
in
ma
ch
i
n
e
tr
a
n
s
la
tio
n
fo
r
… (
Ja
y
a
n
a
n
d
A
.
K
a
mb
le
)
1273
n
o
t
h
av
e
E
n
g
lis
h
as
f
ir
s
t
lan
g
u
ag
e
[
14
]
b
y
p
r
o
v
id
i
n
g
b
ett
er
ac
ce
s
s
to
lear
n
in
g
m
ater
ia
ls
with
co
n
s
eq
u
en
t
in
cr
ea
s
e
in
u
n
d
er
s
tan
d
in
g
an
d
p
ar
ticip
atio
n
.
I
n
p
o
liti
cs,
MT
ca
n
b
r
ea
k
b
ar
r
ier
s
th
r
o
u
g
h
in
clu
d
in
g
th
o
s
e
wh
o
d
o
n
o
t
s
p
ea
k
E
n
g
lis
h
as
a
f
ir
s
t
lan
g
u
ag
e
[
15
]
.
MT
is
co
n
s
id
er
ed
to
b
e
b
en
e
f
icial
in
h
ea
lth
ca
r
e
wh
er
e
cr
itical
m
ed
ical
in
f
o
r
m
atio
n
n
ee
d
s
to
b
e
d
eliv
er
ed
i
n
a
lo
ca
l
lan
g
u
ag
e
wh
ich
th
e
p
atien
t
u
n
d
er
s
t
an
d
s
[
16
]
.
Als
o
,
in
th
is
ag
e
o
f
d
ig
ital
m
ed
ia
tr
an
s
ce
n
d
in
g
th
e
lin
g
u
is
tic
b
ar
r
ier
s
(
b
o
u
n
d
ar
ies),
MT
ass
u
m
es
im
p
o
r
tan
ce
in
d
ev
elo
p
in
g
u
s
er
f
r
ien
d
ly
in
ter
f
ac
es a
n
d
co
n
ten
t
f
o
r
s
p
ea
k
er
s
o
f
o
th
e
r
lan
g
u
ag
es
[
17
]
.
MT
h
as
ev
o
lv
ed
in
wav
es
r
elate
d
to
lin
g
u
is
tic
an
d
co
m
p
u
tatio
n
al
tech
n
o
lo
g
y
s
h
if
ts
.
I
n
th
e
ea
r
ly
d
ay
s
,
r
esear
ch
in
th
is
f
ield
wa
s
m
ain
ly
r
u
le
-
b
ased
MT
(
R
B
MT
)
s
y
s
tem
s
wh
ich
u
tili
s
ed
m
an
u
ally
cr
ea
te
d
r
u
les
an
d
d
ee
p
lin
g
u
is
tic
k
n
o
wled
g
e
[
18
]
.
W
h
ile
s
u
ch
s
y
s
tem
s
s
o
m
e
o
f
wh
ic
h
m
ay
w
o
r
k
b
etter
in
s
o
m
e
s
p
ec
if
ic
d
o
m
ain
s
f
r
eq
u
e
n
tly
b
r
o
k
e
d
o
wn
wh
e
n
it
ca
m
e
t
o
id
i
o
m
s
o
r
t
h
e
c
h
ar
ac
ter
is
in
g
s
y
m
p
to
m
s
o
f
lin
g
u
is
tic
v
ar
iatio
n
.
T
h
is
was
a
tu
r
n
in
g
p
o
in
t
as
r
esear
ch
s
h
if
ted
t
o
s
t
atis
tical
m
ac
h
in
e
tr
an
s
latio
n
(
SMT
)
[
1
9
]
,
w
h
ich
em
p
lo
y
ed
s
tatis
tical
m
o
d
els
p
eg
g
ed
o
n
m
ass
iv
e
p
ar
allel
co
r
p
o
r
a
an
d
g
en
er
ated
tr
a
n
s
latio
n
s
th
at
o
cc
u
r
r
ed
with
ce
r
tain
p
r
o
b
a
b
ilit
ies.
SMT
en
ab
led
u
s
to
p
r
o
ce
s
s
m
u
ch
m
o
r
e
d
ata,
an
d
tak
e
in
to
ac
co
u
n
t
d
if
f
er
en
t
ty
p
es
o
f
lan
g
u
ag
e
p
atter
n
s
b
u
t
b
o
th
f
lu
en
cy
an
d
co
n
te
x
t w
er
e
s
till
an
is
s
u
e.
T
h
e
ad
v
en
t
o
f
n
e
u
r
al
m
ac
h
in
e
tr
an
s
latio
n
(
NM
T
)
[
2
0
]
in
th
e
last
f
ew
y
ea
r
s
h
as
b
ee
n
a
r
ev
o
lu
tio
n
f
o
r
d
ec
ad
es
-
o
ld
tech
n
iq
u
es
b
ase
d
o
n
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es.
NM
T
m
o
d
els,
p
ar
ticu
la
r
ly
th
o
s
e
with
th
e
tr
an
s
f
o
r
m
er
ar
c
h
itectu
r
e,
lev
er
ag
e
co
m
p
lex
al
g
o
r
ith
m
s
t
h
at
an
aly
ze
th
e
f
u
ll
co
n
te
x
t
o
f
a
s
en
ten
ce
s
im
u
ltan
eo
u
s
ly
r
esu
ltin
g
in
m
o
r
e
co
h
er
en
t
an
d
f
lu
e
n
t
tr
an
s
latio
n
s
.
T
h
is
p
r
o
g
r
ess
io
n
h
as
r
esu
lted
in
th
e
s
u
b
s
tan
tial
im
p
r
o
v
em
en
t
in
d
ea
lin
g
with
s
y
n
tactic
v
ar
ian
ts
alo
n
g
with
cu
ltu
r
ally
d
ep
en
d
en
t
s
em
an
tics
,
esp
ec
ially
f
o
r
tr
an
s
latin
g
E
n
g
l
is
h
an
d
I
n
d
ian
lan
g
u
ag
es
[
2
1
]
,
b
u
t
ev
en
th
ese
h
a
v
e
f
u
r
th
e
r
ch
allen
g
es
s
u
ch
as
lo
w
-
r
eso
u
r
ce
d
lan
g
u
ag
es
t
r
a
n
s
latio
n
s
u
p
p
o
r
t
an
d
p
r
ec
is
e
id
en
tific
atio
n
o
f
I
n
d
ian
s
p
ec
if
ic
id
io
m
atic
ex
p
r
ess
io
n
s
.
T
h
e
p
ap
e
r
s
u
r
v
ey
s
th
e
r
ec
en
t
tr
en
d
s
,
tech
n
iq
u
es,
an
d
is
s
u
es
in
co
n
tin
u
o
u
s
ly
-
g
r
o
win
g
ar
ea
o
f
MT
s
y
s
tem
s
f
o
r
E
n
g
lis
h
to
I
n
d
ian
lan
g
u
ag
e
t
r
an
s
latio
n
.
W
e
also
h
o
p
e
to
d
ir
ec
t
atten
tio
n
to
wa
r
d
cu
r
r
en
t
s
tatu
s
o
f
MT
,
an
d
p
o
s
s
ib
le
way
s
o
f
im
p
r
o
v
i
n
g
MT
f
r
o
m
p
r
e
v
io
u
s
r
esear
ch
f
in
d
in
g
s
th
r
o
u
g
h
c
o
m
p
ar
in
g
d
iv
er
s
if
ied
ap
p
r
o
ac
h
es.
W
e
s
ee
k
to
p
r
o
m
o
te
b
etter
co
m
m
u
n
icatio
n
an
d
u
n
d
er
s
tan
d
i
n
g
in
I
n
d
ia’
s
m
u
lti
lin
g
u
al
s
o
ciety
.
T
ab
le
1
s
h
o
ws
th
e
esti
m
ated
n
u
m
b
er
o
f
s
p
ea
k
er
s
,
g
eo
g
r
ap
h
i
c
co
v
er
ag
e
a
n
d
th
e
o
n
lin
e
av
a
ilab
le
d
ata
co
r
p
o
r
a
f
o
r
E
n
g
lis
h
an
d
s
o
m
e
o
f
th
e
lead
in
g
I
n
d
ian
la
n
g
u
a
g
es.
E
n
g
lis
h
is
a
s
p
ec
ial
ca
s
e,
f
o
r
it
h
as
1
.
5
b
illi
o
n
s
p
ea
k
er
s
in
1
4
6
co
u
n
tr
ies
an
d
an
alr
ea
d
y
m
ass
iv
e
o
n
lin
e
co
r
p
u
s
o
f
d
ata.
On
th
e
o
th
er
h
a
n
d
,
Hin
d
i,
B
en
g
ali
an
d
T
am
il h
av
e
lar
g
e
o
r
m
o
d
e
r
ate
s
ized
o
n
lin
e
co
r
p
o
r
a
b
u
t in
d
if
f
er
e
n
t la
n
g
u
a
g
es with
th
eir
s
p
ea
k
er
b
ases
an
d
co
u
n
tr
y
d
is
tr
ib
u
tio
n
s
.
T
h
er
e
a
r
e
s
ev
er
al
o
t
h
er
I
n
d
ian
lan
g
u
a
g
es
(
e.
g
.
,
T
elu
g
u
,
Ma
r
ath
i,
K
an
n
ad
a)
with
lar
g
e
n
u
m
b
er
s
o
f
s
p
ea
k
er
s
wh
o
s
e
o
n
lin
e
d
ata
co
r
p
o
r
a
ar
e
r
elati
v
ely
lim
ited
,
in
d
icatin
g
a
d
e
m
an
d
f
o
r
e
n
h
an
ce
d
d
ig
ital r
eso
u
r
ce
s
in
th
ese
lan
g
u
ag
es.
T
ab
le
1
.
E
n
g
lis
h
an
d
t
h
e
to
p
I
n
d
ian
lan
g
u
ag
es with
n
u
m
b
er
o
f
s
p
ea
k
er
s
an
d
av
ailab
le
o
n
lin
e
d
ata
co
r
p
u
s
[
22
],
[
23
]
La
n
g
u
a
g
e
N
u
mb
e
r
o
f
sp
e
a
k
e
r
s
(
a
p
p
r
o
x
.
)
N
u
mb
e
r
o
f
c
o
u
n
t
r
i
e
s
sp
o
k
e
n
O
n
l
i
n
e
c
o
r
p
u
s
/
d
a
t
a
a
v
a
i
l
a
b
l
e
En
g
l
i
sh
1
.
5
b
i
l
l
i
o
n
1
4
6
Ex
t
e
n
si
v
e
H
i
n
d
i
6
0
0
m
i
l
l
i
o
n
20
La
r
g
e
B
e
n
g
a
l
i
3
0
0
m
i
l
l
i
o
n
4
M
o
d
e
r
a
t
e
Te
l
u
g
u
9
6
mi
l
l
i
o
n
3
Li
mi
t
e
d
M
a
r
a
t
h
i
9
5
mi
l
l
i
o
n
3
Li
mi
t
e
d
Ta
mi
l
7
5
mi
l
l
i
o
n
6
M
o
d
e
r
a
t
e
U
r
d
u
7
0
mi
l
l
i
o
n
26
M
o
d
e
r
a
t
e
G
u
j
a
r
a
t
i
6
0
mi
l
l
i
o
n
6
Li
mi
t
e
d
K
a
n
n
a
d
a
5
6
mi
l
l
i
o
n
3
Li
mi
t
e
d
O
d
i
a
4
0
mi
l
l
i
o
n
2
Li
mi
t
e
d
Te
l
u
g
u
9
6
mi
l
l
i
o
n
3
Li
mi
t
e
d
2.
MET
H
O
D
2
.
1
.
Co
m
pa
ra
t
iv
e
a
na
ly
s
is
o
f
MT
m
et
ho
ds
MT
h
as
ev
o
lv
e
d
s
ig
n
if
ican
tly
o
v
er
th
e
y
ea
r
s
,
a
n
d
m
u
ltip
le
m
eth
o
d
s
h
a
v
e
b
ee
n
d
ev
el
o
p
ed
to
ad
d
r
ess
th
e
ch
allen
g
es
o
f
tr
a
n
s
latin
g
b
etwe
en
I
n
d
ian
la
n
g
u
a
g
es
an
d
E
n
g
lis
h
.
I
n
t
h
is
ar
ea
o
f
wo
r
k
,
SMT
,
NM
T
an
d
L
L
Ms
ar
e
ex
ten
s
iv
ely
u
s
ed
.
E
ac
h
h
as
its
u
n
iq
u
e
s
tr
en
g
th
s
a
n
d
wea
k
n
ess
es
th
at
m
ak
e
it
m
o
r
e
o
r
less
ef
f
ec
tiv
e
in
d
if
f
er
e
n
t tr
an
s
latio
n
s
itu
atio
n
s
.
SMT
is
wid
ely
u
s
ed
f
o
r
tr
a
n
s
latin
g
co
n
te
n
t
in
s
p
ec
if
ic
f
ield
s
,
r
ea
d
in
g
th
r
o
u
g
h
th
e
m
u
ltil
in
g
u
al
co
n
ten
t
an
d
m
ak
in
g
p
r
ed
ictio
n
s
o
f
wh
o
m
will
b
e
tr
an
s
lated
b
ased
o
n
s
tatis
tical
m
o
d
els
[
2
4
]
,
[
2
5
]
.
On
e
o
f
th
e
b
en
ef
its
in
SMT
is
co
n
tr
o
llin
g
s
en
ten
ce
f
r
a
g
m
en
ts
an
d
s
tr
u
ctu
r
ed
p
h
r
ases
,
u
s
in
g
p
h
r
ase
-
b
ased
m
o
d
els
co
n
s
id
er
in
g
wo
r
d
f
r
eq
u
e
n
cy
a
n
d
co
-
o
cc
u
r
r
en
ce
p
atter
n
s
.
I
ts
tr
ain
in
g
d
ata
[
2
6
]
,
o
f
ten
y
ield
s
d
ec
en
t
tr
an
s
latio
n
s
f
o
r
a
f
ew
lan
g
u
ag
e
p
air
s
.
De
s
p
ite
th
ese
b
en
ef
its
,
SMT
als
o
h
as
its
d
r
awb
ac
k
s
.
B
ec
au
s
e
it
d
ep
en
d
s
alm
o
s
t
s
o
lely
o
n
la
r
g
e
a
m
o
u
n
ts
o
f
p
ar
allel
d
ata
(
f
o
r
wh
ich
d
ata
f
o
r
m
o
s
t
lan
g
u
ag
es
a
r
e
n
o
n
ex
is
ten
t)
,
id
io
m
atic
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
1
2
7
2
-
1
2
8
9
1274
ex
p
r
ess
io
n
s
an
d
h
ig
h
er
-
lev
el
lan
g
u
ag
e
u
s
ag
e
f
ar
e
p
o
o
r
ly
[
2
7
]
.
Ad
d
itio
n
ally
,
SMT
is
o
f
ten
in
ef
f
ec
tiv
e
f
o
r
in
tr
icate
p
h
r
ase
s
tr
u
ctu
r
es.
I
t
lack
s
a
s
ec
o
n
d
ar
y
lay
er
o
f
u
n
d
er
s
tan
d
in
g
o
n
h
o
w
to
m
ain
tain
co
h
er
en
ce
a
n
d
f
lu
id
ity
in
tr
an
s
latio
n
s
,
esp
ec
i
ally
in
lar
g
e
an
d
c
o
m
p
lex
s
en
ten
ce
s
wh
er
e
th
e
o
v
e
r
ar
ch
in
g
co
n
tex
t
is
r
ea
lly
im
p
o
r
tan
t.
Dee
p
lear
n
in
g
h
as
o
p
tim
ized
th
e
tr
an
s
latio
n
m
o
d
el
f
o
r
NM
T
th
r
o
u
g
h
o
u
t
th
e
p
r
o
ce
s
s
.
R
em
in
d
th
at
m
o
s
t
o
f
th
e
NM
T
is
b
ased
o
n
en
co
d
er
-
d
ec
o
d
er
a
n
d
a
d
d
in
g
a
tten
tio
n
m
ec
h
a
n
is
m
s
.
T
h
e
m
o
r
e
co
n
te
x
t
in
ter
m
s
o
f
s
ettin
g
(
wh
ich
wo
r
d
s
ap
p
ea
r
n
ea
r
est
to
o
n
e
an
o
t
h
er
)
s
u
r
r
o
u
n
d
in
g
a
s
tatem
en
t,
th
e
h
i
g
h
e
r
th
e
ac
cu
r
ac
y
an
d
f
lu
id
ity
o
f
tr
a
n
s
latio
n
s
.
An
ad
v
an
tag
e
o
f
NM
T
is
th
at
it
ca
n
m
an
ag
e
lo
n
g
-
d
is
tan
ce
tr
an
s
la
tio
n
al
d
ep
en
d
en
cies
b
etwe
en
wo
r
d
s
in
a
s
en
ten
ce
an
d
m
ai
n
tain
s
n
atu
r
al
n
ess
u
p
to
co
m
p
lex
wo
r
d
o
r
d
e
r
.
T
h
e
cr
u
cial
r
o
le
h
er
e
is
p
lay
ed
b
y
th
e
atten
tio
n
th
at
h
elp
s
o
u
r
m
o
d
el
to
f
o
cu
s
o
n
r
elev
an
t
p
ar
ts
o
f
in
p
u
ts
wh
en
tr
an
s
latin
g
,
th
u
s
p
r
o
v
id
i
n
g
lin
g
u
is
tically
an
d
co
n
tex
tu
ally
r
ich
t
r
an
s
latio
n
s
.
NM
T
m
y
s
ter
io
u
s
ly
h
as
n
o
p
r
o
b
lem
talk
i
n
g
s
p
ec
if
ics
,
b
u
t
o
n
ly
b
ec
a
u
s
e
th
e
tr
ain
in
g
d
ata
an
d
m
o
d
el
p
ar
am
eter
s
ar
e
lar
g
em
.
I
n
o
r
d
er
to
p
er
f
o
r
m
co
r
r
ec
tly
,
it
n
ee
d
s
a
lo
t
o
f
tr
ain
in
g
d
ata,
an
d
t
h
is
is
th
e
is
s
u
e
w
ith
lo
w
-
r
eso
u
r
ce
la
n
g
u
a
g
es:
th
er
e
ju
s
t
is
n
ète
en
o
u
g
h
d
ata
a
v
ailab
le.
Ad
d
itio
n
ally
,
m
o
d
els
lik
e
th
ese
r
eq
u
ir
e
a
lo
t
o
f
co
m
p
u
tin
g
p
o
wer
to
tr
ain
,
m
ea
n
in
g
th
at
team
s
th
at
d
o
n
o
t
h
av
e
w
o
r
ld
-
class
r
eso
u
r
ce
s
m
ay
n
o
t
ev
e
n
b
e
ab
le
to
d
o
s
o
.
Su
ch
ch
alle
n
g
es
ca
n
r
ed
u
ce
th
e
ef
f
ec
tiv
en
ess
o
r
av
ailab
ilit
y
o
f
NM
T
in
s
o
m
e
ca
s
es,
p
ar
ticu
lar
ly
f
o
r
lo
wer
-
r
eso
u
r
ce
la
n
g
u
a
g
es a
n
d
co
m
p
u
tin
g
s
y
s
tem
s
with
less
a
d
v
an
ce
d
tech
n
o
l
o
g
y
[
2
8
]
.
So
m
e
r
ec
en
t
lar
g
e
-
s
ca
le
lan
g
u
ag
e
m
o
d
els
(
L
L
Ms)
s
u
ch
as
GPT
an
d
B
E
R
T
[
2
9
]
h
av
e
also
m
ad
e
g
r
ea
t
im
p
ac
ts
b
ec
au
s
e
th
e
y
c
an
tr
an
s
late
task
s
u
s
in
g
m
u
lti
p
le
d
atasets
an
d
d
ee
p
n
e
u
r
al
n
etwo
r
k
s
.
T
h
ey
a
r
e
ca
p
ab
le
o
f
tr
an
s
latin
g
m
u
ltip
l
e
lan
g
u
a
g
es
b
ec
au
s
e
th
e
y
ca
n
f
u
n
ctio
n
ac
r
o
s
s
d
if
f
er
en
t
f
o
r
m
ats.
On
e
o
f
th
eir
p
r
im
e
s
tr
en
g
th
s
is
in
f
ew
-
s
h
o
t
an
d
ev
en
ze
r
o
-
s
h
o
t
lear
n
in
g
.
T
h
is
co
m
es
o
u
t
to
b
e
esp
ec
i
ally
u
s
ef
u
l
in
lo
w
-
r
eso
u
r
ce
lan
g
u
ag
es,
wh
er
e
a
l
ar
g
e
co
r
p
u
s
o
f
an
n
o
tated
d
ata
d
o
es
n
o
t
ex
is
t,
s
in
ce
n
o
w
th
e
y
g
et
th
e
ab
ilit
y
to
g
en
er
alize
to
th
eir
tr
an
s
latio
n
co
u
n
ter
p
a
r
t w
ith
litt
le
s
u
p
er
v
is
io
n
.
Ho
wev
er
,
L
L
Ms
co
m
e
with
a
s
et
o
f
ch
allen
g
es
o
f
th
eir
o
wn
.
T
h
ey
r
e
q
u
ir
e
a
lo
t
o
f
m
em
o
r
y
an
d
p
o
wer
wh
ic
h
m
an
y
c
o
m
p
an
i
es
ju
s
t
d
o
n
’
t
h
av
e
th
e
i
n
f
r
astru
ctu
r
e
f
o
r
.
Mo
r
e
o
v
er
,
d
esp
it
e
b
ein
g
ab
le
t
o
b
e
f
in
e
-
tu
n
e
d
f
o
r
a
p
a
r
ticu
lar
tr
a
n
s
latio
n
task
,
it
r
eq
u
ir
es
l
o
ts
o
f
ex
p
er
ts
an
d
h
u
g
e
d
ataset
a
cc
ess
,
wh
ich
is
n
o
t
r
ea
d
ily
av
ailab
le
o
r
ca
n
c
o
s
t
a
lo
t.
Du
e
to
t
h
is
,
ad
ap
tiv
ely
tu
n
in
g
L
L
Ms
f
o
r
d
o
m
ain
-
s
p
ec
if
ic
tr
an
s
latio
n
co
u
ld
b
e
n
o
n
-
tr
iv
ial
esp
ec
ially
in
a
lo
w
-
r
eso
u
r
ce
co
n
tex
t.
Similar
to
w
h
at
was
d
o
n
e
in
[
1
]
,
we
co
m
p
ar
e
th
e
b
eh
av
io
r
s
o
f
ea
ch
p
r
o
p
o
s
ed
tr
a
n
s
latio
n
tech
n
iq
u
e,
as p
r
esen
t
in
T
ab
le
2.
T
ab
le
2
.
C
o
m
p
a
r
ativ
e
s
u
m
m
ar
y
o
f
s
tr
en
g
th
s
,
lim
itatio
n
s
,
an
d
B
L
E
U
s
co
r
es f
o
r
SMT
,
NM
T
,
an
d
L
L
Ms
M
e
t
h
o
d
S
t
r
e
n
g
t
h
s
Li
mi
t
a
t
i
o
n
s
B
LEU
S
c
o
r
e
(
A
v
e
r
a
g
e
)
S
t
a
t
i
st
i
c
a
l
mac
h
i
n
e
t
r
a
n
s
l
a
t
i
o
n
(
S
M
T)
G
o
o
d
f
o
r
st
r
u
c
t
u
r
e
d
p
h
r
a
s
e
s a
n
d
sh
o
r
t
se
n
t
e
n
c
e
s
S
t
r
u
g
g
l
e
s w
i
t
h
i
d
i
o
ma
t
i
c
e
x
p
r
e
ssi
o
n
s;
r
e
q
u
i
r
e
s
l
a
r
g
e
p
a
r
a
l
l
e
l
d
a
t
a
~
2
5
-
3
0
(
v
a
r
i
e
s
b
y
p
a
i
r
)
N
e
u
r
a
l
ma
c
h
i
n
e
t
r
a
n
sl
a
t
i
o
n
(
N
M
T)
H
a
n
d
l
e
s l
o
n
g
-
r
a
n
g
e
d
e
p
e
n
d
e
n
c
i
e
s;
b
e
t
t
e
r
f
l
u
e
n
c
y
R
e
q
u
i
r
e
s
l
a
r
g
e
d
a
t
a
s
e
t
s
;
c
o
m
p
u
t
a
t
i
o
n
a
l
l
y
e
x
p
e
n
si
v
e
~
3
0
-
35
La
r
g
e
l
a
n
g
u
a
g
e
mo
d
e
l
s
(
LLM
s)
M
u
l
t
i
l
i
n
g
u
a
l
s
u
p
p
o
r
t
:
f
e
w
-
sh
o
t
a
n
d
z
e
r
o
-
s
h
o
t
l
e
a
r
n
i
n
g
H
i
g
h
r
e
s
o
u
r
c
e
r
e
q
u
i
r
e
me
n
t
s
;
n
e
e
d
s
f
i
n
e
t
u
n
i
n
g
~
3
5
-
40
W
e
also
p
r
o
v
id
e
co
m
p
ar
is
o
n
s
with
o
th
er
m
ajo
r
ap
p
r
o
ac
h
es
to
MT
in
clu
d
in
g
SMT
a
n
d
NM
T
,
as
well
a
s
L
L
Ms,
to
i
llu
s
tr
ate
th
eir
r
elativ
e
s
tr
en
g
th
s
an
d
w
ea
k
n
ess
es.
SMT
m
ak
es
h
u
g
e
u
s
e
o
f
p
ar
allel
d
ata
(
th
e
k
in
d
with
tr
an
s
latio
n
s
)
,
i
s
ex
tr
em
ely
ac
c
u
r
ate
f
o
r
s
h
o
r
t
s
en
ten
ce
s
n
icely
f
r
am
e
d
b
y
p
h
r
ases
,
an
d
s
im
p
ly
b
r
ea
k
s
d
o
wn
wh
e
n
u
s
ed
o
n
v
e
r
n
ac
u
lar
.
B
leu
s
co
r
e
is
ex
p
ec
t
ed
to
b
e
s
o
m
ewh
er
e
b
etwe
en
2
5
-
3
0
d
e
p
en
d
i
n
g
o
n
lan
g
u
ag
e
p
air
.
I
n
co
n
tr
ast,
NM
T
ca
n
allo
w
f
lu
e
n
t
d
ec
o
d
in
g
a
n
d
b
e
r
o
b
u
s
t
to
co
m
p
licated
ar
ch
itectu
r
es
s
u
ch
as
lo
n
g
co
n
ten
ts
d
ep
e
n
d
en
cies.
T
h
e
B
L
E
U
s
co
r
es
ar
e
ty
p
ically
s
o
m
ewh
er
e
b
etwe
en
3
0
a
n
d
3
5
,
th
o
u
g
h
b
o
th
o
f
th
o
s
e
m
etr
ics
ar
e
d
ep
en
d
en
t
o
n
a
m
ass
iv
e
am
o
u
n
t
o
f
d
ata
an
d
co
m
p
u
ter
p
o
wer
.
Du
e
to
t
h
e
m
u
ltil
in
g
u
al
tr
an
s
latio
n
ca
p
a
b
ilit
ies
an
d
ze
r
o
-
s
h
o
t,
f
ew
-
s
h
o
t
lear
n
in
g
ab
ilit
y
,
L
L
M
ex
c
els
in
th
is
r
eg
ar
d
,
wh
ich
is
lar
g
ely
b
e
n
ef
icial
f
o
r
l
o
w
-
r
eso
u
r
ce
lan
g
u
ag
es.
L
L
Ms
o
b
tain
t
h
e
s
tate
-
of
-
t
h
e
-
ar
t
p
er
f
o
r
m
a
n
ce
s
(
m
o
r
e
th
a
n
3
5
-
4
0
B
L
E
U)
b
u
t
ar
e
co
m
p
u
tatio
n
ally
d
e
m
an
d
in
g
an
d
r
eq
u
ir
e
h
ea
v
y
f
in
e
-
t
u
n
in
g
.
T
h
e
er
a
o
f
NM
T
a
n
d
L
L
M
h
as
in
d
ee
d
attr
ac
ted
u
s
in
t
o
a
n
u
n
p
r
ec
ed
en
ted
l
y
h
ig
h
-
q
u
ality
a
n
d
ef
f
icien
t
p
r
o
ce
s
s
es
o
f
tr
an
s
latio
n
th
o
u
g
h
SMT
laid
th
e
co
r
n
er
s
to
n
e
o
f
MT
.
Ho
we
v
er
,
t
h
er
e
ar
e
s
o
m
e
p
r
o
b
lem
s
with
ea
ch
o
f
th
e
ap
p
r
o
ac
h
es
wh
en
tr
ea
tin
g
m
o
r
e
c
o
m
p
lex
lin
g
u
is
tic
s
tr
u
ctu
r
es
an
d
r
eso
u
r
ce
d
e
m
an
d
s
.
B
y
k
n
o
win
g
th
ese
ap
p
r
o
ac
h
es,
esp
ec
ially
in
v
iew
o
f
th
e
d
if
f
er
e
n
t
lan
g
u
ag
e
r
eq
u
ir
em
en
ts
p
o
s
ed
b
y
th
e
I
n
d
ian
c
o
n
tex
t,
r
esear
ch
er
s
a
n
d
d
ev
elo
p
e
r
s
ca
n
d
ec
id
e
wh
ic
h
s
tr
ateg
y
to
a
d
o
p
t f
o
r
a
p
ar
ticu
lar
tr
a
n
s
latio
n
t
ask
.
W
h
en
th
e
s
ize
o
f
tr
ain
in
g
d
ata
b
ec
o
m
es
lar
g
er
,
we
co
m
p
ar
e
B
L
E
U
s
co
r
es
o
f
th
r
e
e
d
if
f
er
e
n
t
tr
an
s
latio
n
ap
p
r
o
ac
h
es
in
Fig
u
r
e
1
:
L
L
Ms,
NM
T
,
an
d
SMT
.
T
h
e
B
L
E
U
s
co
r
e
-
a
m
ea
s
u
r
e
o
f
th
e
q
u
ality
o
f
tr
an
s
latio
n
—
g
o
es
u
p
as
y
o
u
u
s
e
m
o
r
e
d
ata.
Ho
wev
e
r
,
ir
r
e
s
p
ec
tiv
e
o
f
t
h
e
s
ca
le
o
f
d
ata,
L
L
Ms
co
n
s
is
ten
tly
s
ig
n
if
ican
tly
im
p
r
o
v
e
o
v
e
r
b
o
th
SMT
an
d
NM
T
h
av
in
g
h
ig
h
er
B
L
E
U
s
co
r
e.
SMT
s
h
r
in
k
s
v
er
y
s
lo
wly
,
with
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
B
r
id
g
in
g
th
e
lin
g
u
is
tic
d
ivid
e:
r
ec
en
t d
ev
elo
p
men
ts
in
ma
ch
i
n
e
tr
a
n
s
la
tio
n
fo
r
… (
Ja
y
a
n
a
n
d
A
.
K
a
mb
le
)
1275
litt
le
f
u
r
th
er
im
p
r
o
v
em
e
n
t
af
t
er
g
ettin
g
a
ce
r
tain
am
o
u
n
t
o
f
tr
ain
in
g
d
ata,
wh
ile
NM
T
co
n
tin
u
es
to
g
et
b
etter
as
it
is
p
r
o
v
id
ed
m
o
r
e
tr
ain
i
n
g
d
ata
ev
en
with
o
u
t
r
ea
c
h
in
g
S
MT
o
r
L
L
M
lev
el.
Su
ch
g
r
ap
h
s
h
o
ws
th
at
L
L
Ms
ar
e
s
u
p
er
io
r
to
SMT
an
d
NM
T
in
h
ar
d
tr
an
s
latio
n
s
,
an
d
th
ey
b
en
ef
it th
e
m
o
s
t f
r
o
m
b
ig
co
r
p
o
r
a.
Fig
u
r
e
1
.
B
L
E
U
s
co
r
e
co
m
p
ar
is
o
n
f
o
r
SMT
,
NM
T
,
a
n
d
L
L
Ms w
ith
in
cr
ea
s
in
g
tr
ain
in
g
d
a
ta
2
.
2
.
Cha
lleng
es in
MT
f
o
r
I
nd
ia
n L
a
ng
ua
g
es
MT
f
o
r
I
n
d
ia
n
lan
g
u
ag
es
[
30
]
h
as
s
ee
n
s
ig
n
i
f
ican
t
g
r
o
wt
h
in
r
ec
e
n
t
tim
es.
T
h
e
r
e
ar
e,
h
o
wev
er
,
s
ev
er
al
im
p
o
r
tan
t
o
b
s
tacle
s
t
o
MT
s
y
s
tem
s
b
ein
g
d
ep
lo
y
e
d
an
d
ev
o
lv
ed
ef
f
ec
tiv
ely
.
T
h
ese
co
n
s
tr
ain
ts
ar
e
d
is
cu
s
s
ed
in
th
is
s
ec
tio
n
,
an
d
a
n
in
s
ig
h
t to
th
e
f
u
tu
r
e
wo
r
k
o
f
r
esear
ch
an
d
d
ev
elo
p
m
en
t is p
r
esen
ted
.
2
.
2
.
1
.
Cha
lleng
es wit
h
da
t
a
s
ca
rc
it
y
On
e
o
f
th
e
p
r
im
ar
y
c
h
allen
g
es
in
cr
ea
tin
g
r
eliab
le
MT
f
o
r
I
n
d
ian
lan
g
u
ag
es
in
p
ar
tic
u
lar
is
th
e
lim
ited
av
ailab
ilit
y
o
f
lar
g
e
p
a
r
allel
co
r
p
o
r
a
[
3
1
]
.
Desp
ite
s
o
m
e
p
r
o
jects
lik
e
Sam
an
an
tar
[
3
2
]
b
ein
g
ef
f
ec
tiv
e
in
cr
awlin
g
d
ata
f
o
r
a
f
ew
lan
g
u
ag
e
p
air
s
,
th
er
e
ar
e
v
er
y
f
e
w
o
r
h
ar
d
ly
an
y
co
m
p
r
eh
en
s
i
v
e
d
atasets
av
ailab
le
to
tr
ain
an
d
test
m
o
d
els
th
at
wo
u
ld
wo
r
k
o
n
s
ev
er
al
o
f
th
e
r
e
g
io
n
al
lan
g
u
a
g
es.
T
h
e
d
ata
s
p
ar
s
en
ess
p
r
esen
ts
a
b
asic b
o
ttlen
ec
k
f
o
r
tr
ain
in
g
a
n
d
test
in
g
o
f
tr
a
n
s
latio
n
m
o
d
els.
T
h
ese
co
r
p
o
r
a
s
h
o
u
ld
b
e
en
l
ar
g
ed
in
t
h
e
f
u
tu
r
e,
an
d
d
e
v
elo
p
m
e
n
t
o
f
s
u
ch
r
eso
u
r
ce
s
m
u
s
t
illu
s
tr
ate
n
e
w
s
tep
s
in
p
ar
allel
d
ata
p
r
eser
v
atio
n
p
lan
n
in
g
em
p
h
asizin
g
co
n
tin
u
o
u
s
co
l
lab
o
r
atio
n
b
etwe
en
lan
g
u
ag
e
co
m
m
u
n
ities
,
ed
u
ca
tio
n
al
in
s
titu
tio
n
s
an
d
tech
n
o
lo
g
y
g
en
e
r
atin
g
en
titi
e
s
to
k
ee
p
u
p
h
ig
h
-
q
u
ality
p
a
r
allel
d
ata
ab
le
to
f
ac
e
b
o
t
h
d
ialec
tal
as
well
a
s
s
o
ci
o
-
r
eg
is
ter
s
itu
atio
n
s
.
2
.
2
.
2
.
I
dio
m
a
t
ic
a
nd
cult
ura
l nua
nces
I
n
d
ian
lan
g
u
a
g
es
h
av
e
h
u
g
e
c
o
n
cu
r
r
en
t
id
io
m
s
an
d
cu
ltu
r
e
-
s
p
ec
if
ic
p
h
ases
,
m
o
s
t
o
f
wh
ic
h
ca
n
o
n
ly
b
e
u
n
d
er
s
to
o
d
u
s
in
g
th
e
s
am
e
m
eth
o
d
in
MT
th
ese
d
a
y
s
[
3
3
]
.
T
h
er
e
is
a
v
ast
v
ar
iab
ilit
y
in
to
s
u
ch
id
io
m
atic
s
tr
u
ctu
r
es
b
ased
o
n
s
o
u
r
ce
lan
g
u
ag
e
cu
ltu
r
e,
in
an
y
s
in
g
le
tar
g
et
lan
g
u
ag
e
it
ca
n
v
er
y
d
if
f
e
r
en
t.
Mo
d
els
(
p
ar
ticu
lar
l
y
th
e
o
n
e
s
wh
ich
ar
e
t
r
ain
ed
o
n
s
m
aller
d
atasets
)
s
tr
u
g
g
le
t
o
ca
p
t
u
r
e
s
u
ch
n
u
a
n
ce
s
,
r
esu
ltin
g
in
tech
n
ical
tr
an
s
latio
n
c
o
r
r
ec
t
o
u
tp
u
ts
th
at
m
is
s
t
h
e
m
ar
k
.
T
o
t
h
is
en
d
,
id
io
m
atic
p
h
r
ase
tr
an
s
latio
n
is
r
elativ
ely
u
n
d
e
r
-
ex
p
lo
r
ed
as
a
task
o
n
its
o
wn
a
n
d
i
n
r
elati
o
n
to
th
e
g
o
al
o
f
cr
ea
tin
g
s
y
s
tem
s
th
at
ar
e
ab
le
to
u
n
d
er
s
tan
d
/c
o
r
r
ec
t
c
u
ltu
r
e
-
s
p
e
cif
ic
tr
an
s
latio
n
s
.
T
h
e
New
Y
o
r
k
team
h
as
p
ar
t
n
er
ed
with
l
o
ca
l
an
d
lin
g
u
is
tic
ex
p
er
ts
to
m
ak
e
s
u
r
e
th
at
th
e
y
ar
e
r
ea
d
y
f
o
r
th
is
task
,
to
o
;
as
cu
ltu
r
al
p
o
licies
wil
l
b
ec
o
m
e
k
ey
in
en
s
u
r
in
g
MT
s
y
s
tem
s
m
ee
t th
e
p
er
f
o
r
m
an
ce
ex
p
ec
tatio
n
s
an
d
s
u
itab
ilit
y
o
f
th
e
n
ativ
e
s
p
ea
k
er
.
2
.
2
.
3
.
M
o
del
co
m
plex
it
y
a
nd
co
m
pu
t
a
t
io
na
l c
o
s
t
s
NM
T
an
d
L
L
Ms
im
p
r
o
v
e
f
lu
en
cy
a
n
d
co
r
r
ec
t
n
ess
,
b
u
t
th
e
co
m
p
u
tatio
n
al
co
s
t
in
v
o
lv
in
g
th
e
u
s
e
o
f
NM
T
an
d
L
L
M
is
f
ar
b
ey
o
n
d
th
e
ca
p
ab
ilit
y
o
f
m
o
s
t
s
m
all
r
esear
ch
g
r
o
u
p
s
an
d
m
u
ch
o
f
in
d
u
s
tr
y
.
Dev
elo
p
in
g
an
d
tr
ain
in
g
th
ese
m
o
d
els
ca
n
also
b
e
ex
p
en
s
iv
e
f
o
r
n
o
n
-
p
r
o
f
it
o
r
lo
w
r
eso
u
r
ce
o
r
g
a
n
izatio
n
s
,
wh
o
m
ay
lac
k
th
e
in
f
r
astru
ctu
r
e
to
s
u
p
p
o
r
t
s
u
ch
i
n
itiativ
es
f
in
an
cially
.
T
h
is
is
u
n
d
o
u
b
te
d
ly
a
m
ajo
r
n
eg
ativ
e,
wh
ich
war
r
an
ts
d
esig
n
in
g
al
g
o
r
ith
m
s
an
d
m
o
d
els
th
at
p
r
o
v
id
e
h
ig
h
-
q
u
ality
tr
an
s
latio
n
s
with
o
u
t
b
ein
g
co
m
p
u
tatio
n
ally
p
r
o
h
ib
itiv
ely
ex
p
en
s
iv
e.
Mo
d
el
co
m
p
r
ess
io
n
an
d
o
p
tim
izatio
n
tech
n
iq
u
es
ar
e
also
an
o
p
tio
n
to
im
p
r
o
v
e
ac
ce
s
s
ib
ilit
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
1
2
7
2
-
1
2
8
9
1276
2
.
2
.
4
.
M
ultiling
ua
l
m
o
dels
a
nd
t
ra
ns
f
er
lea
rning
T
r
an
s
f
er
lear
n
in
g
h
as
em
er
g
e
d
as
an
ef
f
ec
tiv
e
ap
p
r
o
ac
h
f
o
r
im
p
r
o
v
in
g
MT
s
y
s
tem
s
[
3
4
]
esp
ec
ially
lo
w
r
eso
u
r
ce
I
n
d
ian
lan
g
u
a
g
es.
On
e
ca
n
s
ca
le
lo
w
r
eso
u
r
ce
tr
an
s
latio
n
ab
ilit
y
to
s
ca
le
f
o
r
e
v
en
lo
w
r
eso
u
r
ce
s
(
s
u
ch
as
Od
ia,
Kan
n
ad
a
an
d
Ass
am
ese)
u
s
in
g
d
ata/m
o
d
els
tr
ain
ed
o
n
h
ig
h
r
eso
u
r
ce
lan
g
u
ag
es
s
u
ch
as
Hin
d
i
an
d
Ma
r
ath
i.
T
h
is
s
tr
ateg
y
le
v
er
ag
es
th
e
i
n
f
o
r
m
atio
n
alr
ea
d
y
p
r
esen
t
in
th
e
d
ata
a
n
d
e
n
ab
les
cr
o
s
s
-
lin
g
u
al
tr
an
s
f
er
o
f
k
n
o
wled
g
e
th
at
ca
n
en
h
an
ce
tr
an
s
latio
n
q
u
ality
an
d
e
f
f
icien
cy
.
T
o
s
y
s
tem
atica
lly
tr
y
an
d
v
alid
ate
s
u
ch
an
ap
p
r
o
ac
h
o
v
er
a
m
u
lt
itu
d
e
o
f
I
n
d
ian
lan
g
u
ag
es
ca
ll
s
f
o
r
well
-
d
ef
in
ed
tr
an
s
f
er
lea
r
n
in
g
m
et
h
o
d
s
an
d
f
r
am
ewo
r
k
s
in
t
h
e
f
u
t
u
r
e.
2
.
3
.
K
ey
co
ncept
s
in
MT
I
n
t
h
i
s
s
e
c
t
i
o
n
w
e
p
r
e
s
e
n
t
t
h
e
w
e
l
l
s
u
i
t
e
d
s
o
l
i
d
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
m
e
t
h
o
d
s
a
n
d
w
e
c
l
a
s
s
i
f
y
t
h
e
m
b
e
s
i
d
e
s
t
r
a
n
s
l
a
t
i
o
n
l
o
w
a
n
d
h
i
g
h
p
r
e
c
i
s
i
o
n
.
I
t
d
e
s
c
r
i
b
e
s
s
e
v
e
r
a
l
m
a
c
h
i
n
e
t
r
a
n
s
l
a
t
i
o
n
a
p
p
r
o
a
c
h
e
s
i
m
p
l
e
m
e
n
t
e
d
i
n
d
e
p
t
h
v
a
r
y
i
n
g
f
r
o
m
s
t
a
t
i
s
t
i
c
a
l
u
p
t
o
m
o
r
e
m
o
d
e
r
n
o
n
e
s
.
T
h
e
a
b
i
l
i
t
y
f
o
r
t
h
e
s
e
m
o
d
e
l
s
t
o
p
r
o
v
i
d
e
t
r
a
n
s
l
a
t
i
o
n
a
c
c
u
r
a
c
y
f
r
o
m
s
i
m
p
l
e
s
t
a
t
i
s
t
i
c
a
l
m
e
t
h
o
d
s
t
o
s
o
p
h
i
s
t
i
c
a
t
e
d
d
e
e
p
l
e
a
r
n
i
n
g
a
n
d
m
u
l
t
i
l
i
n
g
u
a
l
m
o
d
e
l
s
i
s
h
i
g
h
l
i
g
h
t
e
d
i
n
t
h
e
s
e
c
t
i
o
n
.
2
.
3
.
1
.
Sta
t
is
t
ica
l
m
a
chine t
ra
n
s
la
t
io
n
SMT
r
elies
o
n
s
tatis
tica
l
m
o
d
els
to
p
r
ed
ict
tr
an
s
latio
n
b
ase
d
o
n
p
r
o
b
ab
ilit
ies
d
er
iv
ed
f
r
o
m
a
lar
g
e
co
llectio
n
o
f
p
ar
allel
tex
ts
.
I
t
was
b
ased
o
n
h
o
w
o
f
ten
wo
r
d
s
an
d
p
h
r
ases
n
atu
r
al
ly
o
cc
u
r
to
g
eth
er
(
s
p
o
n
tan
eo
u
s
ly
co
m
b
in
e)
i
n
p
ar
allel
tex
ts
in
two
lan
g
u
ag
es,
an
d
m
a
d
e
u
s
e
o
f
th
ese
p
r
o
b
a
b
ilit
ies
to
co
n
s
tr
u
ct
tr
an
s
latio
n
s
.
SMT
d
o
es
well
wh
en
th
e
p
h
r
ase
len
g
th
is
s
m
all
an
d
s
en
ten
ce
s
tr
u
ctu
r
e
g
et
s
class
ic
b
u
t
SMT
r
elies
o
n
p
ar
allel
co
r
p
u
s
f
o
r
m
atch
in
g
p
atter
n
s
b
etwe
en
s
o
u
r
ce
lan
g
u
ag
e
an
d
tar
g
et.
On
th
is
co
u
n
t,
p
h
r
ase
-
b
ased
m
o
d
els ex
ce
l,
p
r
o
d
u
cin
g
o
cc
asio
n
ally
q
u
ite
s
atis
f
ac
to
r
y
tr
an
s
latio
n
s
o
f
th
ese
b
asic sen
ten
ce
s
.
Ho
wev
er
,
SMT
h
as
its
lim
itatio
n
s
.
Dee
p
er
lin
g
u
is
tic
s
tr
u
ctu
r
es
an
d
id
io
m
s
,
o
r
f
in
e
lan
g
u
a
g
e
u
s
e
a
r
e
o
u
t
o
f
i
ts
leag
u
e.
An
d
th
e
p
e
r
f
o
r
m
a
n
ce
g
ets
wo
r
s
e
as
th
e
s
en
ten
ce
b
ec
o
m
es
lo
n
g
er
a
n
d
m
o
r
e
co
m
p
licated
.
T
h
is
is
b
ec
au
s
e
SMT
is
h
ea
v
ily
d
ep
en
d
en
t
o
n
m
ass
iv
e
p
ar
allel
co
r
p
o
r
a
an
d
d
o
es
n
o
t
h
av
e
an
y
k
in
d
o
f
co
n
ce
p
tu
al
co
n
s
tr
ain
ts
,
wh
ich
ar
e
n
ec
ess
ar
y
f
o
r
d
ee
p
er
o
r
d
e
r
u
s
e
o
f
th
e
lan
g
u
ag
e
[
35
],
[
36
]
.
T
h
e
b
asis
f
o
r
th
e
SMT
ar
e
p
h
ar
s
e
-
b
ased
m
o
d
el
s
(
th
er
e
is
e.
g
.
,
n
o
s
in
g
le
wo
r
d
tr
an
s
latio
n
m
o
d
el)
,
an
d
a
T
M/L
M
co
m
b
in
atio
n
g
en
er
atin
g
p
r
o
p
e
r
-
s
h
ap
e
s
en
te
n
ce
s
with
f
lu
en
t u
s
e
o
f
lan
g
u
a
g
e,
r
esp
ec
tiv
ely
.
De
co
d
i
n
g
is
a
ce
n
tr
al
co
m
p
o
n
en
t
o
f
SMT
:
th
e
m
ac
h
in
e
"d
ec
id
es"
o
n
th
e
b
est
p
iece
wis
e
tr
an
s
latio
n
co
m
b
in
in
g
wo
r
d
s
an
d
p
h
r
ases
b
ased
o
n
s
t
a
ti
s
t
i
c
al
l
i
k
e
li
h
o
o
d
(
S
n
o
v
e
r
e
t
a
l
.
T
h
o
s
e
p
e
r
c
e
n
ta
g
e
s
h
a
v
e
to
d
o
w
i
t
h
t
h
e
p
a
r
t
i
c
u
l
a
r
m
at
h
o
f
S
M
T
,
w
h
i
c
h
I
in
t
h
e
o
r
y
d
e
t
e
r
m
i
n
e
s
h
o
w
l
i
k
el
y
a
g
i
v
e
n
t
r
a
n
s
l
a
ti
o
n
p
r
o
d
u
c
e
d
b
y
t
h
e
s
y
s
t
e
m
is
t
o
b
e
r
i
g
h
t
,
v
e
r
s
u
s
a
b
s
o
l
u
te
g
i
b
b
e
r
i
s
h
.
T
h
i
s
p
r
o
c
ess
a
l
l
o
ws
S
M
T
t
o
h
a
n
d
l
e
s
t
r
u
c
t
u
r
e
d
t
r
a
n
s
l
a
ti
o
n
s
p
r
e
tt
y
w
e
l
l
.
N
e
v
e
r
t
h
e
le
s
s
,
b
e
ca
u
s
e
i
t
r
e
li
es
s
o
h
e
a
v
i
l
y
o
n
d
a
t
a
a
n
d
s
t
a
ti
s
ti
ca
l
m
o
d
e
l
s
,
i
t
h
a
s
a
h
a
r
d
e
r
t
i
m
e
w
i
t
h
t
h
e
m
o
r
e
c
o
m
p
l
e
x
n
u
a
n
c
es
o
f
l
a
n
g
u
a
g
e
.
Ma
th
em
atica
l E
x
p
r
ess
io
n
[
25
]
:
e
̂
=
a
r
g
ma
x
P
(
e
∣
f
)
=
a
r
g
ma
x
P
(
f
∣
e
)
P
(
e
)
(
1
)
W
h
er
e,
P
(
e
∣
f
)
is
th
e
p
r
o
b
ab
ilit
y
o
f
a
tar
g
et
s
en
ten
ce
b
ein
g
g
iv
en
a
s
o
u
r
ce
s
en
ten
ce
.
(
∣
)
I
s
th
e
p
r
o
b
ab
ilit
y
o
f
a
s
o
u
r
ce
s
en
ten
ce
g
iv
en
a
tar
g
et
s
en
ten
ce
.
2
.
3
.
2
.
Neura
l
m
a
chine t
ra
ns
la
t
io
n
MT
u
s
e
o
cc
u
r
s
in
au
t
o
m
ated
tr
an
s
latio
n
to
o
ls
a
n
d
p
r
o
g
r
a
m
s
(
in
clu
d
in
g
b
u
t
n
o
t
lim
ited
to
SMT
)
,
but
r
ec
e
n
tly
NM
T
h
as
b
ec
o
m
e
th
e
in
d
u
s
tr
y
s
tan
d
ar
d
f
o
r
ad
v
an
cin
g
th
e
p
e
r
f
o
r
m
an
ce
o
f
MT
t
h
r
o
u
g
h
d
ee
p
lear
n
in
g
tech
n
o
lo
g
y
.
I
n
s
tead
o
f
p
iece
s
,
th
at
wo
r
d
o
r
p
h
r
ase
y
o
u
u
s
e
is
f
ed
to
NM
T
a
n
d
ev
e
r
y
th
in
g
tr
a
n
s
lates a
t
o
n
ce
tak
in
g
th
e
f
u
ll
co
n
tex
t
in
to
ac
co
u
n
t.
E
n
c
o
d
er
tak
es
t
h
e
in
p
u
t
s
en
ten
ce
as
in
p
u
t
an
d
g
en
er
ates
a
f
ix
ed
en
co
d
in
g
wh
ich
is
th
en
u
s
ed
b
y
th
e
d
ec
o
d
er
t
o
p
r
e
d
ict
its
s
ec
o
n
d
lan
g
u
ag
e
tr
an
s
latio
n
.
Att
en
tio
n
is
o
n
e
o
f
th
e
k
ey
f
ea
tu
r
es
o
f
NM
T
an
d
g
r
an
ts
th
e
m
o
d
el
with
an
a
b
ilit
y
to
f
o
cu
s
o
n
im
p
o
r
tan
t
wo
r
d
s
/p
h
r
ases
in
a
s
en
ten
ce
,
esp
ec
ially
cr
u
cial
wh
en
tr
an
s
latin
g
lo
n
g
er
an
d
m
o
r
e
s
y
n
tactica
lly
co
m
p
lex
s
en
ten
ce
s
wh
ich
u
ltima
tely
en
ab
les
u
s
to
b
etter
p
ass
in
g
lo
n
g
-
ra
n
g
e
d
ep
e
n
d
en
cies
f
r
o
m
in
p
u
ts
s
o
we
g
en
er
ate
ap
p
licab
le
tr
an
s
latio
n
s
th
at
ar
e
n
o
t
o
n
ly
n
atu
r
al
s
o
u
n
d
in
g
b
u
t
co
n
tex
tu
ally
p
er
tin
en
t
to
o
.
Ad
d
itio
n
ally
,
NM
T
u
s
es
lan
g
u
ag
e
m
o
d
els
to
en
s
u
r
e
th
at
tr
an
s
latio
n
s
ar
e
als
o
s
y
n
tactica
lly
an
d
s
ty
lis
tically
co
r
r
ec
t.
An
o
th
er
b
en
ef
it
o
f
s
u
b
wo
r
d
N
MT
s
y
s
tem
is
th
at
th
e
y
ca
n
p
r
o
d
u
ce
id
io
m
atic
a
n
d
n
atu
r
al,
c
o
n
tex
tu
al
tr
an
s
latio
n
.
NM
T
s
:
So
,
W
h
ile
NM
T
Has
L
im
itatio
n
s
.
T
o
p
e
r
f
o
r
m
well,
it
r
eq
u
ir
es
a
lar
g
e
am
o
u
n
t
o
f
tr
ai
n
in
g
d
ata
wh
ich
is
d
if
f
icu
lt
to
co
m
e
b
y
f
o
r
lo
w
-
r
eso
u
r
ce
d
la
n
g
u
ag
es.
I
n
ad
d
itio
n
,
th
ey
a
r
e
co
m
p
u
tatio
n
ally
in
ten
s
iv
e
an
d
n
o
t
all
th
e
r
esear
ch
g
r
o
u
p
s
m
ay
h
a
v
e
an
a
b
ilit
y
to
b
u
ild
it
o
r
ap
p
ly
in
r
e
-
s
o
u
r
ce
lim
ited
en
v
ir
o
n
m
en
ts
.
Desp
ite
th
e
m
e
n
tio
n
ed
c
h
allen
g
es,
NM
T
n
o
wad
ay
s
is
th
e
s
tate
-
of
-
th
e
-
ar
t
ap
p
r
o
ac
h
to
o
b
tain
h
ig
h
q
u
ality
MT
f
o
r
v
ar
io
u
s
lan
g
u
ag
es
[
3
5
]
,
[
3
6
]
.
T
h
e
b
as
ic
ar
ch
itectu
r
e
o
f
NM
T
a
p
p
li
es
th
e
en
co
d
e
-
a
n
d
-
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
B
r
id
g
in
g
th
e
lin
g
u
is
tic
d
ivid
e:
r
ec
en
t d
ev
elo
p
men
ts
in
ma
ch
i
n
e
tr
a
n
s
la
tio
n
fo
r
… (
Ja
y
a
n
a
n
d
A
.
K
a
mb
le
)
1277
d
ec
o
d
e
(
E
ND)
m
ec
h
a
n
is
m
an
d
atten
tio
n
lib
e
r
ally
to
d
ed
u
ce
th
e
m
ap
p
in
g
f
r
o
m
a
s
o
u
r
ce
s
en
ten
ce
to
its
m
o
s
t
s
im
ilar
p
ar
ts
.
NM
T
ad
d
itio
n
ally
u
s
es
ad
v
an
ce
d
n
e
u
r
al
n
ets
lik
e
R
NNs,
L
STM
s
an
d
GR
Us
to
h
an
d
le
s
eq
u
en
tial
d
ataa
n
d
k
ee
p
th
e
c
o
n
tex
t
with
in
lo
n
g
er
s
en
ten
ce
s
.
T
h
ese
th
in
g
s
co
m
b
in
ed
allo
w
f
o
r
tr
a
n
s
latio
n
s
to
b
e
f
lu
id
an
d
s
en
s
ib
le
ev
en
i
f
th
e
s
en
ten
ce
s
ar
e
co
m
p
lex
.
I
n
p
ar
ticu
lar
,
th
e
m
ath
em
a
tical
f
o
r
m
s
o
f
th
e
r
elatio
n
s
h
ip
b
etwe
en
th
e
in
p
u
t
an
d
o
u
tp
u
t
s
eq
u
en
ce
s
g
iv
es
in
s
ig
h
t
in
to
h
o
w
s
u
ch
co
n
tex
tu
ally
s
im
ilar
tr
an
s
latio
n
s
s
h
o
u
ld
b
e
g
en
er
at
ed
b
y
t
h
is
m
o
d
el
ac
co
r
d
in
g
l
y
.
Ma
th
em
atica
l E
x
p
r
ess
io
n
[
25
],
[
37
]
:
N
M
T
(
X
)
=
s
oft
ma
x
(
W
⋅
ht
+
b
)
(
2
)
w
h
er
e
is
o
u
tp
u
t
p
r
o
b
ab
ilit
y
d
i
s
tr
ib
u
tio
n
o
v
e
r
tar
g
et
v
o
ca
b
u
l
ar
y
,
is
weig
h
t
m
atr
ix
,
h
t
is
h
i
d
d
en
s
tate
o
f
n
eu
r
al
n
etwo
r
k
at
tim
e
s
tep
,
b
in
n
o
t
atio
n
s
d
en
o
tes
b
ias
v
ec
to
r
an
d
ac
ts
as
ac
tiv
atio
n
f
u
n
ctio
n
th
at
m
ap
s
in
p
u
t
lo
g
its
to
b
e
a
p
r
o
b
a
b
ilit
y
d
is
tr
ib
u
tio
n
.
2
.
3
.
3
.
B
y
t
e
pa
ir
enco
din
g
(
B
P
E
)
B
PE
[
3
8
]
is
a
p
o
wer
f
u
l
te
x
t
t
o
k
en
izatio
n
alg
o
r
ith
m
,
f
o
r
ch
ar
ac
ter
r
em
o
v
in
g
m
eth
o
d
s
in
NL
P
th
at
lear
n
s
to
em
b
ed
r
ar
e
o
r
u
n
s
ee
n
wo
r
d
s
.
B
PE
s
p
lits
wo
r
d
s
ev
en
f
u
r
t
h
er
in
to
s
m
aller
p
iece
s
,
o
f
ten
ch
a
r
ac
ter
s
o
r
s
u
b
wo
r
d
s
,
wh
ile
f
u
ll
-
w3
o
r
d
s
ca
n
lead
to
o
u
t
-
of
-
v
o
ca
b
u
lr
y
i
n
s
tan
ce
s
(
OOV)
.
Nex
t,
to
r
ea
ch
a
p
r
ed
eter
m
in
ed
v
o
ca
b
u
lar
y
s
ize
it
co
m
b
in
es
t
h
e
lev
els
o
f
h
ig
h
est
c
o
u
n
ts
in
an
iter
ativ
e
way
u
n
til
th
e
d
esi
r
ed
s
ize
is
r
ea
ch
ed
.
T
h
is
is
a
n
ice
tr
a
d
e
-
o
f
f
:
co
m
m
o
n
wo
r
d
s
r
em
ai
n
p
r
esen
t
(
s
o
ar
e
ea
s
y
wo
r
d
s
s
tay
in
tact)
,
b
u
t
less
co
m
m
o
n
(
o
r
h
a
r
d
er
)
o
n
es a
r
e
b
r
o
k
en
in
t
o
ea
s
ier
ch
u
n
k
s
.
Su
ch
an
ap
p
r
o
ac
h
wo
r
k
s
well
f
o
r
NM
T
wh
ich
is
a
task
th
at
u
tili
ze
s
lar
g
e
v
o
ca
b
u
lar
y
to
tr
a
n
s
late
tex
t
an
d
th
e
m
o
d
el
s
h
o
u
ld
n
o
t
b
e
b
iased
b
y
u
n
s
ee
n
wo
r
d
s
.
On
e
d
o
wn
s
id
e
th
o
u
g
h
,
is
th
at,
as
we
s
et
t
o
p
v
o
ca
b
u
lar
y
s
ize
m
an
u
ally
,
a
wr
o
n
g
ch
o
i
ce
ca
n
lead
t
o
p
o
o
r
p
er
f
o
r
m
an
ce
.
T
h
e
r
ef
o
r
e
,
wh
ile
B
PE
is
a
p
o
wer
f
u
l
an
d
im
p
o
r
tan
t
p
ar
t
o
f
th
e
e
n
co
d
e
r
th
at
b
o
o
s
ts
m
o
d
el
p
e
r
f
o
r
m
a
n
ce
,
an
o
p
tim
al
v
o
ca
b
u
lar
y
s
ize
s
h
o
u
ld
also
b
e
f
in
etu
n
ed
.
B
PE
wo
r
k
s
b
y
iter
ativ
ely
r
ep
l
ac
in
g
m
o
s
t
f
r
eq
u
e
n
t
ad
jace
n
t
ch
ar
ac
ter
s
o
r
ch
ar
ac
ter
p
air
s
(
u
p
to
a
lim
it
in
s
ize
d
eg
r
ee
)
u
n
til
th
e
tar
g
et
p
r
ep
r
o
ce
s
s
ed
v
o
ca
b
u
la
r
y
s
ize
is
r
ea
ch
ed
.
T
h
o
u
g
h
it
allo
ws
a
m
o
d
el
to
b
etter
lear
n
f
r
o
m
d
if
f
er
e
n
t
lin
g
u
is
tic
d
o
m
ain
s
,
its
p
er
f
o
r
m
a
n
ce
is
s
en
s
it
iv
e
b
o
t
h
to
o
s
m
all
an
d
to
o
lar
g
e
o
f
v
o
ca
b
u
lar
y
s
izes (
wh
ich
im
p
a
cts tr
an
s
latio
n
q
u
ality
)
[
3
9
]
.
Ma
th
em
atica
lly
,
E
x
p
r
ess
io
n
[
40
]
:
B
PE
(
in
p
ut
)
=
ite
r
a
tive
r
e
pl
a
c
e
me
n
t
of
most
fr
e
q
ue
n
t
b
yte
pa
ir
s
(
3
)
2
.
3
.
4
.
At
t
ent
io
n
m
ec
ha
nis
m
T
h
e
atten
tio
n
m
ec
h
a
n
is
m
is
a
h
u
g
e
a
d
v
an
ce
f
o
r
th
e
n
eu
r
al
n
etwo
r
k
o
f
esp
ec
ially
f
o
r
MT
.
T
h
is
en
ab
les
m
o
d
els
to
f
o
cu
s
m
o
r
e
o
n
r
elev
a
n
t
co
m
p
o
n
en
ts
o
f
th
eir
i
n
p
u
t
r
ath
er
th
a
n
tr
ea
tin
g
ev
er
y
t
h
in
g
in
th
e
in
p
u
t
as
eq
u
ally
r
elev
en
t.
I
t
d
o
es
th
is
b
y
ass
ig
n
in
g
atte
n
tio
n
weig
h
ts
wh
ich
h
elp
s
th
e
m
o
d
el
to
d
eter
m
in
e
th
at
p
ar
t
o
f
th
e
in
p
u
t
o
n
wh
ich
it
s
h
o
u
ld
f
o
cu
s
(
0
f
o
r
f
o
r
g
ettin
g
,
1
f
o
r
r
em
em
b
er
i
n
g
)
,
allo
win
g
it
to
lear
n
co
m
p
lex
s
en
ten
ce
s
an
d
lo
n
g
-
r
an
g
e
wo
r
d
d
ep
e
n
d
en
cies
m
o
r
e
ef
f
icien
tly
.
I
t
is
th
is
atten
tio
n
th
at
allo
ws
f
o
r
m
o
r
e
ac
cu
r
ate
m
o
d
els,
an
d
a
g
r
ea
ter
co
n
tex
t
awa
r
e
n
ess
esp
ec
ially
wh
en
th
er
e
ar
e
ce
r
tain
wo
r
d
s
th
at
s
ig
n
if
y
th
e
im
p
o
r
tan
ce
m
o
r
e
th
a
n
o
t
h
er
s
.
T
h
ey
b
asically
aid
f
o
r
th
e
b
etter
an
d
p
r
ec
is
e
tr
an
s
latio
n
s
b
ec
au
s
e
th
ey
ca
p
tu
r
e
th
e
in
-
p
lace
p
ec
u
liar
ities
o
f
lan
g
u
ag
e.
T
h
e
m
ath
e
m
atica
l
ex
p
r
ess
io
n
s
co
n
ce
r
n
i
n
g
th
is
p
r
o
ce
s
s
ar
e
m
en
tio
n
ed
as f
o
llo
w
[
4
1
]
.
Ma
th
em
atica
l E
x
p
r
ess
io
n
[
42
]
:
A
tte
n
t
ion
(
Q
,
K
,
V
)
=
s
oft
ma
x
(
√
)
(
4
)
Q
is
th
e
q
u
er
y
m
atr
ix
,
K
is
th
e
k
ey
m
atr
ix
,
V
is
th
e
v
alu
e
m
atr
i
x
,
an
d
is
th
e
d
im
en
s
io
n
o
f
th
e
k
e
y
s
.
2
.
3
.
5
.
T
ra
ns
f
o
rm
er
a
rc
hite
ct
ure
T
h
e
T
r
an
s
f
o
r
m
er
in
n
atu
r
al
la
n
g
u
ag
e
p
r
o
ce
s
s
in
g
elim
in
ates
th
e
s
eq
u
en
tial
o
r
d
er
o
f
th
in
g
s
an
d
ca
n
b
e
d
o
n
e
o
n
b
atch
m
u
ch
f
aster
,
m
o
r
e
ef
f
icien
tly
in
p
ar
allel.
Un
lik
e
lan
g
u
a
g
e
m
o
d
els
s
u
c
h
as
R
NNs
th
at
ar
e
p
r
o
ce
s
s
ed
to
k
e
n
b
y
to
k
en
,
th
e
T
r
an
s
f
o
r
m
e
r
p
r
o
ce
s
s
es
all
to
k
en
s
at
o
n
ce
an
d
th
is
is
wh
y
it
ca
p
tu
r
es
lo
n
g
r
a
n
g
e
d
ep
en
d
e
n
cies
m
u
ch
b
etter
th
a
n
an
y
o
th
er
ar
c
h
itectu
r
e
ev
er
cr
ea
ted
.
T
h
e
f
o
u
n
d
atio
n
f
o
r
th
is
is
s
elf
-
atten
tio
n
,
wh
ich
allo
ws
ea
ch
to
k
en
to
“lo
o
k
at”
ev
er
y
o
th
e
r
to
k
e
n
in
an
in
p
u
t
s
eq
u
en
ce
an
d
g
et
b
ac
k
a
co
n
tex
t
r
ep
r
esen
tatio
n
o
f
th
e
e
n
tire
co
n
tex
t f
o
r
ea
ch
wo
r
d
.
T
h
e
co
r
e
co
m
p
o
n
en
ts
o
f
th
e
T
r
an
s
f
o
r
m
er
ar
e
its
s
elf
-
atte
n
tio
n
lay
er
s
an
d
p
o
s
itio
n
al
en
co
d
in
g
s
,
wh
ich
allo
w
it
t
o
k
n
o
w
h
o
w
to
k
en
s
r
e
late
to
t
o
n
e
an
o
t
h
er
as
well
as
w
o
r
d
o
r
d
er
.
I
n
ad
d
itio
n
,
it
u
s
es
m
u
lti
-
h
ea
d
atten
tio
n
th
at
allo
ws
th
e
m
o
d
el
to
p
ay
atten
tio
n
to
d
i
f
f
er
en
t
p
o
s
itio
n
s
o
f
th
e
i
n
p
u
t
s
eq
u
en
ce
at
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
1
2
7
2
-
1
2
8
9
1278
s
am
e
tim
e
m
ak
in
g
it
b
etter
a
b
le
to
f
in
d
co
m
p
lex
p
atter
n
s
.
T
h
ese
s
et
o
f
f
ea
tu
r
e
s
en
ab
le
th
e
tr
an
s
f
o
r
m
er
to
ex
ce
l
in
lik
e
tr
an
s
latio
n
,
s
u
m
m
ar
ized
it
a
n
d
a
n
s
wer
an
y
q
u
esti
o
n
.
Ov
er
v
iew
o
f
t
h
e
m
ath
e
m
atica
l
f
o
r
m
u
latio
n
s
o
f
s
elf
-
atten
tio
n
in
tr
an
s
f
o
r
m
e
r
s
[
4
3
]
,
[
4
4
]
.
Tr
a
n
s
for
me
r
(
X
)
=
M
ul
tiHe
a
d
(
X
)
+
FFN
(
M
ul
tiHe
a
d
(
X
)
)
(
5
)
M
ul
tiHe
a
d
(
X
)
r
ep
r
esen
ts
t
h
e
m
u
lti
-
h
ea
d
atten
tio
n
m
ec
h
an
is
m
a
n
d
r
ep
r
es
en
ts
a
f
ee
d
-
f
o
r
war
d
n
etwo
r
k
t
h
at
is
ap
p
lied
to
th
e
o
u
tp
u
t o
f
m
u
lti
-
h
ea
d
atten
tio
n
.
2
.
3
.
6
.
L
a
rg
e
la
n
g
ua
g
e
m
o
del
s
LLMs
ar
e
r
o
b
u
s
t
n
eu
r
al
n
etw
o
r
k
s
d
esig
n
ed
to
s
o
lv
e
v
ar
i
o
u
s
k
in
d
s
o
f
lan
g
u
ag
e
task
s
s
u
ch
as
tex
t
g
en
er
atio
n
,
tr
an
s
latio
n
,
q
u
esti
o
n
-
an
s
wer
in
g
an
d
s
u
m
m
ar
izat
io
n
.
GPT,
B
E
R
T
an
d
L
L
aM
A
ar
e
all
ex
am
p
les
o
f
lan
g
u
ag
e
m
o
d
els
tr
ain
e
d
b
y
d
e
ep
lear
n
in
g
o
v
e
r
v
e
r
y
lar
g
e
s
ets
o
f
wr
itten
/ty
p
ed
d
ata;
th
ey
u
tili
ze
th
eir
tr
ain
in
g
d
ata
to
g
e
n
er
ate
h
u
m
an
-
s
o
u
n
d
in
g
r
esp
o
n
s
es
with
r
e
s
p
ec
t
to
a
g
iv
en
p
r
o
m
p
t.
L
L
Ms
a
r
e
d
if
f
er
en
t
b
ec
a
u
s
e
th
e
y
lev
er
ag
e
s
elf
-
atten
tio
n
an
d
tr
a
n
s
f
o
r
m
er
ar
c
h
itectu
r
es
to
ac
h
i
ev
e
co
n
tex
t
o
v
er
lo
n
g
er
d
is
tan
ce
s
,
wh
ich
en
ab
les
th
em
to
b
etter
p
r
o
ce
s
s
an
d
g
en
er
ate
tex
t.
L
L
Ms
ar
e
b
asically
d
ef
in
ed
as
tr
an
s
f
o
r
m
er
s
(
An
o
th
er
Dim
en
s
io
n
f
o
r
T
r
a
n
s
f
o
r
m
er
s
an
d
L
an
g
u
ag
e
Mo
d
els)
o
r
th
e
u
s
e
o
f
lar
g
e
-
s
ca
le
p
r
etr
ain
in
g
a
n
d
f
in
etu
n
i
n
g
to
ac
h
iev
e
s
tate
-
of
-
th
e
-
a
r
t
r
esu
lts
o
n
lan
g
u
ag
e
task
s
.
T
h
is
m
ak
es
f
o
r
p
o
wer
f
u
l
m
o
d
els
th
at
ar
e
tr
ain
ed
o
n
lar
g
e
d
atasets
f
o
llo
wed
b
y
a
f
in
e
-
tu
n
in
g
p
h
a
s
e
o
n
m
an
y
d
o
wn
s
tr
ea
m
task
s
,
p
r
o
d
u
cin
g
v
er
y
ef
f
ec
tiv
e
m
o
d
els
ev
en
with
lim
ited
d
ata
a
v
ai
lab
le.
W
h
ile
th
ese
m
o
d
els ca
n
d
o
wo
n
d
er
wo
r
k
,
t
h
ey
also
r
eq
u
ir
e
s
ig
n
i
f
ican
t c
o
m
p
u
te
r
eso
u
r
ce
s
to
tr
ain
a
n
d
r
u
n
.
Her
e
we
s
im
p
ly
d
ef
in
e
th
e
m
at
h
s
f
o
r
t
h
eir
m
ai
n
o
p
er
atio
n
(
m
o
s
tly
s
elf
-
atten
tio
n
)
im
p
lem
e
n
tatio
n
o
f
co
r
e
p
ar
t.
Ma
th
em
atica
l e
x
p
r
ess
io
n
[
45
]
,
(
|
1
,
2
,
…
…
…
,
−
1
)
=
(
1
,
2
,
…
…
−
1
)
∑
(
1
,
2
,
…
…
)
(
6)
2
.
4
.
E
v
a
lua
t
i
o
n m
et
rics
Au
to
m
atic
ev
alu
atio
n
m
etr
ics
:
au
to
m
atic
ev
alu
atio
n
m
etr
i
cs
p
lay
a
cr
itical
r
o
le
in
MT
s
y
s
tem
s
,
ev
alu
atin
g
th
e
p
r
o
d
u
ce
d
o
u
tp
u
t
ag
ain
s
t
h
u
m
an
r
ef
er
e
n
ce
tr
an
s
latio
n
s
an
d
m
ea
s
u
r
in
g
h
o
w
clo
s
ely
it
ap
p
r
o
x
im
ates
th
em
[
4
6
]
.
On
e
p
o
p
u
lar
m
etr
ic
is
ca
lled
b
ilin
g
u
al
e
v
alu
atio
n
u
n
d
er
s
tu
d
y
(
B
L
E
U
)
th
at
q
u
an
tifie
s
wh
at
p
er
ce
n
t o
f
th
e
n
g
r
a
m
s
in
y
o
u
r
MT
an
d
a
r
ef
e
r
en
ce
tr
an
s
latio
n
o
v
er
lap
with
ea
ch
o
th
er
,
as a
m
ea
s
u
r
em
en
t
o
f
f
lu
e
n
cy
a
n
d
ac
c
u
r
ac
y
.
T
h
e
s
ec
o
n
d
m
etic
is
th
e
m
etr
ic
f
o
r
ev
alu
atio
n
o
f
tr
an
s
latio
n
wit
h
ex
p
licit
o
r
d
e
r
in
g
(
ME
T
E
OR
)
,
wh
ich
also
f
o
c
u
s
es
o
n
w
o
r
d
s
in
b
o
th
r
ef
er
en
ce
an
d
tr
an
s
latio
n
b
u
t
also
ac
co
u
n
ts
f
o
r
co
n
tex
tu
a
l
m
ea
n
in
g
,
r
e
d
u
ce
s
in
f
lectio
n
al
v
ar
iab
ilit
y
.
L
astl
y
,
tr
an
s
latio
n
er
r
o
r
r
ate
(
T
E
R
)
g
ets
th
e
m
in
im
u
m
n
u
m
b
er
o
f
ed
its
th
at
ar
e
r
e
q
u
ir
e
d
to
m
atc
h
a
m
ac
h
in
e
o
u
tp
u
t
tr
an
s
latio
n
to
t
h
e
r
e
f
er
en
ce
tr
an
s
latio
n
an
d
f
o
cu
s
es
o
n
ly
o
n
th
o
s
e
s
eg
m
en
ts
wh
er
e
tr
an
s
la
ted
tex
t
m
is
m
atch
es
h
u
m
an
.
T
h
ese
d
im
en
s
io
n
s
r
ep
r
esen
t
a
h
o
lis
tic
in
-
d
e
p
th
p
r
o
f
ilin
g
o
f
th
e
tr
an
s
latio
n
q
u
a
lity
b
ey
o
n
d
ad
e
q
u
ac
y
a
n
d
p
o
s
t
-
ed
itin
g
ef
f
o
r
t o
f
MT
o
u
tp
u
t.
B
L
E
U
s
co
r
e:
B
L
E
U
s
co
r
e
is
a
co
m
m
o
n
ly
u
s
ed
m
etr
ic
b
y
Pap
in
en
i
et
a
l.
[
47
]
to
ev
alu
ate
m
ac
h
in
e
-
g
e
n
er
ated
tr
a
n
s
latio
n
s
.
I
t
in
d
icate
s
h
o
w
m
u
ch
a
tr
a
n
s
latio
n
d
ev
iates
f
r
o
m
a
s
et
o
f
h
u
m
a
n
r
e
f
er
en
ce
tr
an
s
latio
n
s
,
r
an
g
in
g
b
etwe
en
0
an
d
1
with
h
ig
h
e
r
v
alu
e
r
ep
r
esen
tin
g
b
etter
alig
n
m
e
n
t
.
n
-
g
r
am
p
r
ec
is
io
n
m
ea
s
u
r
es
th
e
o
v
er
lap
o
f
n
-
g
r
am
s
in
m
ac
h
in
e
o
u
tp
u
t
an
d
r
ef
er
en
ce
s
,
co
m
p
u
tin
g
B
L
E
U
s
co
r
e.
A
b
r
ev
ity
p
en
alty
(
B
P)
is
ac
co
u
n
ted
f
o
r
in
th
e
f
in
al
co
m
p
u
tatio
n
to
av
o
id
f
av
o
r
in
g
s
h
o
r
ter
tr
an
s
la
tio
n
s
,
an
d
p
en
al
ize
tr
an
s
latio
n
s
th
at
ar
e
ex
ce
s
s
iv
ely
s
h
o
r
t
wh
ich
a
h
ig
h
p
r
ec
i
s
io
n
o
n
l
y
wo
u
ld
r
ewa
r
d
.
B
L
E
U
d
ev
elo
p
ed
as
a
s
tan
d
ar
d
m
etr
ic
f
o
r
au
to
m
atic
m
ac
h
in
e
tr
an
s
latio
n
ev
alu
ati
o
n
d
u
e
to
its
o
b
jectiv
e
,
r
ep
r
o
d
u
cib
le
co
m
p
u
tatio
n
an
d
its
g
o
o
d
s
ca
lin
g
p
r
o
p
er
tie
s
ac
r
o
s
s
lan
g
u
a
g
es
an
d
m
o
d
el
s
.
T
h
e
f
o
r
m
u
la
in
cl
u
d
es
a
b
r
e
v
ity
p
e
n
alty
wh
ic
h
co
n
s
tr
ain
s
th
e
len
g
th
o
f
th
e
tr
a
n
s
latio
n
:
=
⋅
(
∑
=
1
)
(
7
)
w
h
er
e:
BP
is
th
e
b
r
ev
ity
p
en
alty
.
is
th
e
p
r
ec
is
io
n
o
f
n
-
g
r
am
s
.
is
th
e
weig
h
t f
o
r
ea
c
h
n
-
g
r
am
len
g
th
(
u
s
u
ally
u
n
if
o
r
m
)
.
T
h
ese
co
n
ce
p
ts
h
elp
in
th
e
o
r
ig
in
o
f
m
o
d
er
n
MT
m
et
h
o
d
s
,
f
r
o
m
h
is
to
r
ical
v
iews
to
co
n
tem
p
o
r
ar
y
tech
n
iq
u
es
[
4
8
]
.
ME
T
E
OR
Sco
r
e:
A
s
co
r
in
g
m
eth
o
d
t
h
at
allo
ws
f
o
r
a
m
o
r
e
n
u
an
ce
d
co
m
p
ar
is
o
n
o
f
th
e
MT
with
B
L
E
U.
T
h
e
f
ir
s
t,
ME
T
E
OR
,
f
o
cu
s
es
m
o
r
e
o
n
m
ea
n
in
g
an
d
s
y
n
o
n
y
m
s
th
an
n
-
g
r
am
o
v
e
r
lap
.
T
h
e
m
eth
o
d
s
ch
ar
ac
ter
is
e
th
e
m
ea
n
in
g
an
d
r
elatio
n
s
b
etwe
en
wo
r
d
s
to
a
m
o
r
e
a
cc
u
r
ate
d
eg
r
ee
,
b
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
B
r
id
g
in
g
th
e
lin
g
u
is
tic
d
ivid
e:
r
ec
en
t d
ev
elo
p
men
ts
in
ma
ch
i
n
e
tr
a
n
s
la
tio
n
fo
r
… (
Ja
y
a
n
a
n
d
A
.
K
a
mb
le
)
1279
r
ep
r
esen
tin
g
n
o
t
o
n
ly
ex
ac
t
wo
r
d
p
atter
n
s
b
u
t
also
m
atch
es
th
at
ar
e
em
b
ed
d
ed
to
g
et
h
er
in
ce
r
tain
way
s
.
T
h
e
s
co
r
e
v
a
r
ies b
etwe
en
0
an
d
1
,
with
a
h
ig
h
er
v
alu
e
in
d
ica
tin
g
a
b
etter
co
v
er
ag
e
o
f
t
h
e
h
u
m
an
tr
an
s
latio
n
s
.
I
n
ter
m
s
o
f
co
m
p
u
tatio
n
,
M
E
T
E
OR
'
s
ca
lcu
latio
n
d
o
es
n
o
t
r
ely
o
n
ex
ac
t
m
atc
h
es
as
d
o
B
L
E
U.
T
h
is
co
n
s
is
ts
o
f
d
if
f
er
en
t
ty
p
e
s
o
f
wo
r
d
m
atch
es,
ex
ac
t
m
at
ch
,
s
tem
m
ed
m
atch
an
d
s
y
n
o
n
y
m
-
b
ased
m
atch
as
well
as a
cc
o
u
n
tin
g
f
o
r
th
e
o
r
d
er
in
wh
ich
wo
r
d
s
wer
e
wr
itten
.
An
d
we
p
en
alize
a
tr
an
s
latio
n
if
it p
lace
s
wo
r
d
s
in
awk
war
d
s
p
o
ts
.
ME
T
E
OR
ag
g
r
eg
ates
p
r
ec
is
io
n
(
am
o
u
n
t
o
f
MT
th
at
co
r
r
esp
o
n
d
s
to
h
u
m
an
r
ef
e
r
en
ce
)
an
d
r
ec
all
(
h
o
w
m
u
ch
th
e
h
u
m
an
r
ef
er
en
ce
is
ca
p
tu
r
e
d
)
,
b
o
th
ar
e
eq
u
ally
weig
h
ted
.
T
h
is
,
a
n
d
its
wo
r
d
o
r
d
er
p
en
alty
m
ak
e
it
a
p
o
wer
f
u
l
m
ea
s
u
r
e
f
o
r
ass
ess
in
g
th
e
q
u
ality
o
f
tr
an
s
latio
n
.
T
h
e
ME
T
E
OR
f
o
r
m
u
latio
n
is
g
iv
en
b
el
o
w
[
49
]
.
M
E
TEOR
=
10
⋅
P
⋅
R
R
+
9P
(
1
−
)
(
8
)
w
h
er
e:
wh
er
e
th
e
u
n
ig
r
am
r
ec
all
an
d
p
r
ec
is
io
n
ar
e
g
iv
e
n
b
y
R
an
d
P,
r
esp
ec
tiv
ely
.
T
h
e
b
r
e
v
ity
p
en
alty
MN
is
d
eter
m
in
ed
b
y
:
=
0
.
5
(
)
(
9
)
wh
er
e
QR
is
th
e
n
u
m
b
er
o
f
m
atch
in
g
u
n
ig
r
am
s
,
an
d
C
is
th
e
m
in
im
u
m
n
u
m
b
er
o
f
p
h
r
ase
s
r
eq
u
ir
ed
to
m
atch
u
n
ig
r
am
s
in
t
h
e
SMT
o
u
tp
u
t
with
th
o
s
e
f
o
u
n
d
in
th
e
r
ef
er
e
n
ce
tr
an
s
latio
n
s
.
2
.
5
.
Rev
iew
o
f
ex
is
t
ing
m
et
h
o
do
lo
g
ies
E
ar
ly
MT
s
y
s
tem
s
u
s
ed
r
u
le
-
b
ased
ap
p
r
o
ac
h
es
t
h
at
h
o
we
v
er
ar
e
co
n
s
tr
ain
ed
b
y
th
e
a
m
o
u
n
t
o
f
lin
g
u
is
tic
r
eso
u
r
ce
s
r
e
q
u
ir
ed
.
T
h
ese
s
y
s
tem
s
ex
p
lo
ited
h
an
d
-
cr
af
ted
lin
g
u
is
tic
r
u
les
to
tr
a
n
s
late
b
u
t
th
e
y
ten
d
to
b
e
b
r
ittl
e
an
d
d
id
n
o
t
g
e
n
er
alize
well
to
n
o
v
el
s
en
ten
ce
s
[
50
]
.
W
e
n
o
w
p
r
esen
t
r
elate
d
wo
r
k
f
o
r
t
h
e
task
o
f
E
n
g
lis
h
to
Ma
r
ath
i M
T
.
I
n
2
0
1
7
,
a
r
esear
ch
s
tu
d
ied
b
y
Ku
n
ch
u
k
u
ttan
an
d
B
h
attac
h
ar
y
y
a
[
51
]
s
h
o
wed
th
at
B
PE
co
u
ld
f
ac
ilit
ate
tr
an
s
latio
n
b
etwe
en
s
im
ilar
la
n
g
u
ag
es.
T
h
eir
aim
was
to
en
h
an
ce
tr
an
s
latio
n
b
y
b
ein
g
ca
p
ab
le
o
f
lear
n
in
g
v
ar
iab
le
-
len
g
th
s
u
b
-
wo
r
d
u
n
its
,
en
ab
lin
g
an
im
p
r
o
v
ed
h
an
d
lin
g
o
f
v
o
ca
b
u
lar
y
an
d
t
h
e
ad
d
r
ess
in
g
co
m
m
o
n
is
s
u
es
s
u
ch
as
v
o
ca
b
u
lar
y
s
p
ar
s
ity
o
r
o
u
t
-
of
-
v
o
ca
b
u
lar
y
(
OOV)
wo
r
d
s
.
B
PE
p
r
o
ce
e
d
s
b
y
b
e
g
in
n
in
g
with
a
v
o
ca
b
u
lar
y
o
f
w
o
r
d
s
a
n
d
iter
ativ
ely
m
er
g
i
n
g
t
h
e
m
o
s
t
f
r
eq
u
en
t p
air
o
f
ch
ar
ac
ter
s
o
r
s
y
m
b
o
ls
.
I
t
r
ep
ea
ts
th
e
p
r
o
ce
s
s
u
n
til a
f
i
x
ed
n
u
m
b
er
o
f
m
e
r
g
es o
r
s
o
m
e
v
o
ca
b
u
l
ar
y
s
ize
is
o
b
tain
ed
.
I
n
th
is
way
,
B
PE
d
ec
r
ea
s
es
v
o
ca
b
u
lar
y
s
ize
wh
en
r
etain
in
g
s
en
s
e
an
d
allo
ws
m
o
r
e
ef
f
ec
tiv
e
tr
an
s
latio
n
b
etwe
en
s
im
ilar
l
an
g
u
ag
es
(
co
m
p
ar
e
T
ab
le
3
)
.
T
h
e
ap
p
r
o
ac
h
n
atu
r
ally
s
o
lv
es
th
e
p
r
o
b
lem
o
f
OOV,
also
tu
n
es
th
e
SMT
s
y
s
tem
f
o
r
lan
g
u
ag
es
with
s
im
ilar
lex
ical
s
tr
u
ctu
r
e.
I
n
th
e
en
d
,
th
is
h
elp
s
m
ak
e
tr
an
s
lat
io
n
s
m
o
r
e
ef
f
icien
t
in
m
an
ag
in
g
v
o
ca
b
u
lar
y
an
d
s
o
is
well
-
s
u
ited
to
th
e
o
p
tim
izatio
n
o
f
r
elate
d
lan
g
u
ag
e
p
air
s
.
T
ab
le
3
.
T
h
e
p
ar
allel
co
r
p
o
r
a'
s
tr
ain
,
test
,
an
d
tu
n
e
s
p
lits
[
52
]
La
n
g
u
a
g
e
p
a
i
r
Tr
a
i
n
Tu
n
e
Te
st
b
e
n
-
h
i
n
,
p
a
n
-
h
i
n
4
4
7
7
7
1
0
0
0
2
0
0
0
k
o
k
-
mar,
ma
l
-
t
a
m,t
e
l
-
m
a
l
,
h
i
n
-
m
a
l
,
ma
l
-
h
i
n
4
4
7
7
7
1
0
0
0
2
0
0
0
u
r
d
-
h
i
n
,
b
e
n
-
u
r
d
,
u
r
d
-
mal
,
ma
l
-
u
r
d
3
8
1
6
2
8
4
3
1
7
0
7
b
u
l
-
ma
c
1
5
0
0
0
1
0
0
0
2
0
0
0
d
a
n
-
sw
e
1
5
0
0
0
0
1
0
0
0
2
0
0
0
may
-
i
n
d
1
3
7
0
0
0
1
0
0
0
2
0
0
0
kor
-
j
p
n
,
j
p
n
-
kor
6
9
8
0
9
1
0
0
0
2
0
0
0
2
.
5
.
1
.
Da
t
a
s
et
We
em
p
lo
y
ed
th
e
d
ataset
d
es
cr
ib
ed
i
n
T
ab
le
3
to
tr
ain
,
f
in
etu
n
e
an
d
test
MM
T
m
o
d
els
f
o
r
v
a
r
io
u
s
lan
g
u
ag
e
p
air
s
co
m
p
r
is
in
g
I
n
d
ian
lan
g
u
a
g
es
as
well
a
s
n
o
n
-
I
n
d
ian
lan
g
u
ag
es.
T
h
e
ap
p
r
o
ac
h
en
s
u
r
es
r
esp
o
n
s
ib
le
m
o
d
el
o
p
tim
izati
o
n
an
d
esti
m
atio
n
b
y
m
ain
t
ain
in
g
b
ala
n
ce
d
s
p
lits
f
o
r
tr
ain
in
g
,
tu
n
in
g
,
an
d
test
in
g
.
C
o
llectiv
ely
,
it
in
clu
d
es
I
n
do
-
Ar
y
an
an
d
Dr
av
i
d
ian
lan
g
u
ag
es
f
r
o
m
th
e
I
L
C
I
co
r
p
u
s
[
1
7
]
an
d
p
air
s
o
f
wo
r
ld
lan
g
u
ag
es
f
r
o
m
th
e
Op
en
Su
b
titl
es2
0
1
6
co
r
p
u
s
[
1
8
]
,
r
esu
ltin
g
in
a
c
o
m
p
r
e
h
en
s
iv
e
c
h
allen
g
e
d
ataset
f
o
r
cr
o
s
s
-
lin
g
u
al
tr
an
s
latio
n
s
y
s
tem
s
.
T
h
is
also
p
o
s
es
th
e
in
tr
in
s
ic
c
h
allen
g
e
o
f
lin
g
u
is
tic
h
eter
o
g
en
eity
,
in
clu
d
in
g
v
o
ca
b
u
lar
y
s
p
ar
s
en
ess
in
d
u
ce
d
tr
a
n
s
latio
n
er
r
o
r
s
in
s
u
ch
d
atasets
.
B
L
E
U
is
s
c
o
r
e
wh
ich
m
ea
s
u
r
es
th
e
q
u
a
lity
o
f
a
m
o
d
el
b
y
co
m
p
ar
in
g
tr
an
s
latio
n
s
to
h
u
m
an
r
ef
er
en
ce
s
s
o
th
at
we
ca
n
e
s
tim
ate
h
o
w
g
o
o
d
ar
e
with
o
u
r
tr
an
s
lat
io
n
.
B
L
E
U
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
1
2
7
2
-
1
2
8
9
1280
s
co
r
e
h
ig
h
lig
h
ts
th
e
ar
ea
s
t
h
at
n
ee
d
im
p
r
o
v
e
m
en
t,
a
n
d
e
n
co
u
r
ag
es
th
o
s
e
co
n
tex
tu
ally
r
elev
an
t
tr
an
s
latio
n
s
ca
n
b
e
p
r
o
d
u
ce
d
f
lu
en
tly
b
y
MT
m
o
d
els ac
r
o
s
s
d
if
f
er
e
n
t la
n
g
u
a
g
e
p
air
s
.
An
o
th
er
s
tu
d
y
b
y
J
h
a
an
d
B
h
a
ttach
ar
y
y
a
[
5
3
]
in
u
n
d
er
to
o
k
t
h
e
co
n
s
tr
u
c
tio
n
a
n
d
r
eso
u
r
ce
-
b
u
ild
in
g
o
f
NM
T
s
y
s
tem
s
f
o
r
I
n
d
ian
lan
g
u
ag
es
with
d
if
f
icu
lties
s
u
ch
as
r
ich
m
o
r
p
h
o
l
o
g
y
a
n
d
r
elativ
ely
m
o
r
e
co
m
p
lex
s
y
n
tax
.
E
v
er
y
o
n
e
o
f
t
h
ese
lin
g
u
is
tic
elem
en
t
m
a
k
es
tr
an
s
lat
io
n
in
c
r
ea
s
in
g
ly
c
o
m
p
le
x
an
d
ca
lls
f
o
r
a
d
d
itio
n
al
r
ef
in
ed
m
o
d
el
s
to
ch
ec
k
with
it.
T
h
e
au
th
o
r
s
ex
a
m
in
e
a
n
u
m
b
er
o
f
NM
T
ar
ch
itectu
r
es,
with
p
ar
ticu
lar
em
p
h
asis
o
n
t
r
an
s
f
o
r
m
e
r
m
o
d
els
th
at
u
s
e
s
elf
-
atten
tio
n
to
l
ea
r
n
lo
n
g
-
ter
m
ter
m
d
ep
e
n
d
e
n
cies
in
s
en
ten
ce
s
.
T
h
ey
u
s
e
tr
an
s
f
er
-
lear
n
in
g
f
r
o
m
b
etter
-
r
eso
u
r
ce
d
lan
g
u
ag
es
o
r
s
tr
en
g
th
en
ac
cu
r
ac
y
o
n
lo
w
-
r
eso
u
r
ce
o
n
es with
p
r
e
-
tr
ain
ed
m
o
d
els
as
well.
T
h
e
d
em
o
n
s
tr
ated
ex
p
e
r
im
en
t
o
f
n
m
t
in
T
a
b
le
4
th
at,
b
u
ilt
b
etwe
en
I
n
d
ian
lan
g
u
ag
es,
th
e
r
e
p
o
r
ted
i
m
p
r
o
v
em
en
t
is
v
is
ib
le
u
s
in
g
B
L
E
U
s
co
r
es.
Hin
d
i
-
B
en
g
ali,
in
o
th
er
wo
r
d
s
th
e
s
co
r
e
f
o
r
TL
-
less
T
r
an
s
f
o
r
m
e
r
im
p
r
o
v
es
f
r
o
m
2
5
.
4
t
o
3
0
.
7
af
ter
B
PE
(
th
e
v
alu
e
b
ef
o
r
e
was
2
8
.
5
)
Ho
wev
er
,
f
o
r
th
e
T
r
an
s
f
o
r
m
e
r
+
T
L
ca
s
e,
th
is
n
u
m
b
er
h
ad
f
allen
to
2
8
.
4
f
r
o
m
an
in
itial
s
co
r
e
o
f
3
0
.
1
a
f
ter
ap
p
ly
in
g
B
PE
m
ea
n
in
g
th
at
tr
a
n
s
f
er
lea
r
n
in
g
n
ee
d
s
m
o
r
e
f
in
e
-
tu
n
i
n
g
b
ef
o
r
e
it
b
ec
o
m
es
u
s
ef
u
l.
Desp
ite
t
h
e
ad
v
an
ce
m
e
n
t
i
n
tr
an
s
latio
n
q
u
ality
p
r
o
v
id
e
d
b
y
n
m
t,
ac
q
u
ir
es
a
co
n
s
id
er
a
b
l
e
b
o
ttlen
ec
k
o
f
r
eq
u
ir
in
g
lar
g
e
tr
ain
in
g
d
atasets
f
r
o
m
tex
t c
o
r
p
o
r
a,
esp
ec
ially
f
o
r
lo
w
-
r
eso
u
r
ce
lan
g
u
ag
es.
T
ab
le
4
.
T
h
e
p
ar
allel
co
r
p
o
r
a
o
f
I
n
d
ian
lan
g
u
ag
es
[
54
]
D
a
t
a
s
e
t
La
n
g
u
a
g
e
p
a
i
r
N
u
mb
e
r
o
f
se
n
t
e
n
c
e
s
S
o
u
r
c
e
D
a
t
a
s
e
t
1
H
i
n
d
i
-
B
e
n
g
a
l
i
5
0
0
,
0
0
0
P
a
r
a
l
l
e
l
C
o
r
p
u
s
D
a
t
a
s
e
t
2
M
a
r
a
t
h
i
-
Ta
m
i
l
3
0
0
,
0
0
0
P
a
r
a
l
l
e
l
C
o
r
p
u
s
D
a
t
a
s
e
t
3
Te
l
u
g
u
-
G
u
j
a
r
a
t
i
2
0
0
,
0
0
0
P
a
r
a
l
l
e
l
C
o
r
p
u
s
As
f
u
r
th
er
ev
id
en
ce
o
f
im
p
r
o
v
em
en
ts
in
NM
T
s
y
s
tem
s
[
5
5
]
,
Kan
d
im
alla
et
a
l.
[
5
6
]
f
o
cu
s
o
n
tr
an
s
latin
g
f
r
o
m
E
n
g
lis
h
to
v
ar
io
u
s
I
n
d
ian
lan
g
u
ag
es.
T
h
e
y
wr
in
g
s
o
m
e
im
p
r
o
v
em
en
ts
ab
o
v
e
th
e
v
an
illa
atten
tio
n
b
ased
NM
T
im
p
le
m
en
tatio
n
b
y
le
v
er
ag
in
g
tr
an
s
f
er
lear
n
in
g
,
d
o
m
ain
a
d
ap
ta
tio
n
an
d
f
in
et
u
n
in
g
m
eth
o
d
s
.
T
h
ese
m
ec
h
an
is
m
s
ar
e
u
s
ed
with
an
in
ten
t
to
m
o
d
el
th
e
lin
g
u
is
tic
d
if
f
icu
lties
o
f
I
n
d
ian
lan
g
u
ag
es,
th
u
s
h
elp
in
g
o
u
r
s
y
s
tem
in
g
e
n
er
atin
g
co
n
tex
t
s
en
s
itiv
e
an
d
h
ig
h
er
q
u
ality
tr
an
s
latio
n
s
.
T
h
e
r
ig
h
tn
ess
o
f
th
e
o
th
er
p
o
r
tio
n
o
f
th
e
m
o
d
el
o
n
wh
ich
it
is
h
in
g
in
g
is
d
u
e
to
s
p
lice/o
p
tim
ize
i
t
s
o
m
ewh
at
(
th
e
m
o
d
els),
as
p
r
esen
t in
f
o
r
m
u
la
5
,
4
a
n
d
7
r
esp
o
n
s
ib
le
f
o
r
its
g
r
a
d
u
ally
im
p
o
r
tan
t b
u
ild
u
p
in
ca
u
s
e
–
e
f
f
e
ct
ca
lcu
latio
n
s
.
T
h
e
d
ataset
u
s
ed
f
o
r
th
is
wo
r
k
is
th
e
lar
g
e
p
ar
allel
d
ata
f
o
r
E
n
g
lis
h
to
I
n
d
ia
n
lan
g
u
ag
es
p
air
s
,
with
d
o
m
ain
ad
a
p
tio
n
an
d
f
i
n
etu
n
i
n
g
u
s
in
g
m
o
n
o
lin
g
u
al
d
ata.
Su
ch
h
eter
o
g
e
n
eo
u
s
d
atasets
f
ac
ilit
ate
th
e
NM
T
m
o
d
el
to
d
ea
l
with
th
e
in
tr
icat
e
s
y
n
tactic
ch
ar
ac
te
r
is
tics
o
f
I
n
d
ian
la
n
g
u
a
g
es.
I
n
th
e
b
ig
g
e
r
B
L
E
U
s
co
r
e
ca
s
e,
we
m
ak
e
f
u
r
th
er
a
n
aly
s
is
o
n
t
h
e
im
p
ac
t
o
f
Din
g
et
a
l.
B
u
t
t
h
e
p
ap
e
r
also
m
en
tio
n
s
th
at
d
ata
av
ailab
ilit
y
an
d
q
u
ality
co
n
tin
u
e
to
b
e
a
h
is
to
r
ical
ch
allen
g
e,
p
ar
ticu
lar
ly
in
lack
o
f
r
eso
u
r
ce
s
I
n
d
ian
l
an
g
u
ag
es,
wh
ich
is
b
o
ttlen
ec
k
f
o
r
th
e
b
etter
p
er
f
o
r
m
an
ce
o
f
o
u
r
m
o
d
els ac
r
o
s
s
all
L
2
s
.
R
am
esh
et
a
l.
[
3
2
]
Pre
s
en
t
S
am
an
an
tar
:
m
o
s
t
lar
g
est
ac
ce
s
s
ib
le
p
u
b
lic
p
ar
allel
co
r
p
u
s
f
o
r
an
y
1
1
I
n
d
ic
lan
g
u
ag
es
as
well
as
Go
ld
en
B
en
ch
m
ar
k
in
g
d
ataset
f
o
r
MT
.
T
h
ese
ar
e
e
x
p
an
d
ed
in
t
o
m
u
ltip
le
s
o
u
r
ce
s
f
r
o
m
th
r
ee
d
if
f
e
r
en
t
ar
ea
s
(
l
eg
al,
r
elig
io
u
s
an
d
g
o
v
e
r
n
m
en
t
an
d
web
)
to
m
ak
e
a
r
ich
d
ata
s
et
f
o
r
MT
d
ev
elo
p
m
e
n
t
in
lan
g
u
ag
es
lik
e
Hin
d
i,
B
en
g
ali,
T
am
il
an
d
T
elu
g
u
.
T
h
ey
f
o
cu
s
o
n
o
n
e
m
illi
o
n
s
en
ten
ce
p
air
s
with
alig
n
m
en
t
an
d
q
u
ality
d
a
ta.
Ho
wev
er
,
th
e
p
a
p
er
cites
o
n
e
d
r
awb
ac
k
,
a
lack
o
f
q
u
ality
co
n
tr
o
l
f
o
r
s
u
ch
a
lar
g
e
an
d
v
ar
ied
d
ataset.
Ho
wev
er
,
Sam
an
an
tar
is
an
im
p
o
r
tan
t
s
tep
to
war
d
s
cr
ea
tin
g
s
u
ch
lar
g
e
-
s
ca
le
r
ep
o
s
ito
r
ies;
an
d
it h
ig
h
lig
h
ts
th
e
is
s
u
es th
at
co
m
e
in
to
p
lay
in
v
ettin
g
th
is
m
u
ltit
u
d
e
o
f
s
o
u
r
ce
s
f
o
r
q
u
ality
.
Statis
t
ics
in
T
ab
le
5
s
h
o
w
i
n
s
ig
h
t
in
to
p
ar
alle
l
c
o
r
p
o
r
a
with
th
eir
co
r
r
esp
o
n
d
in
g
lan
g
u
ag
e
p
air
s
(
f
r
o
m
E
n
g
lis
h
to
b
asic
I
n
d
i
c)
--
all
d
atasets
ar
e
eith
er
ex
tr
ac
ted
o
r
o
r
ig
in
ally
cr
e
ated
.
Fo
r
in
s
tan
ce
,
E
n
g
lis
h
-
Hin
d
i
h
as
1
,
0
0
0
,
0
0
0
s
en
ten
ce
p
air
s
f
r
o
m
g
e
n
er
al
p
ar
allel
co
r
p
o
r
a
an
d
E
n
g
lis
h
-
B
en
g
ali
h
as
800
,
0
0
0
p
air
s
.
0
co
r
p
u
s
with
s
en
ten
ce
p
air
s
d
ef
in
i
n
g
E
n
g
lis
h
–
T
am
il
s
p
ec
if
icity
W
e
also
r
ep
o
r
t
m
o
n
o
lin
g
u
al
co
r
p
o
r
a
f
o
r
Hin
d
i,
B
en
g
ali
an
d
T
am
il
to
o
am
o
u
n
tin
g
to
2
,
0
0
0
,
0
0
0
s
en
ten
ce
s
ea
ch
in
th
e
tab
le.
T
h
e
k
ey
m
ess
ag
e
f
r
o
m
th
is
tab
le,
h
o
wev
er
,
is
th
at
th
er
e
ar
e
v
e
r
y
lar
g
e,
in
f
ac
t
u
n
b
eliev
ab
ly
d
i
v
er
s
e
p
ar
allel
a
n
d
m
o
n
o
lin
g
u
al
d
ata
av
ailab
le
wh
ich
ar
e
im
p
o
r
tan
t
to
b
u
ild
MT
s
y
s
tem
s
f
o
r
th
ese
lan
g
u
ag
es sp
ec
if
ically
ac
r
o
s
s
m
u
ltip
le
d
o
m
ain
s
.
T
ab
le
5
.
Par
allel
co
r
p
o
r
a
o
f
E
n
g
lis
h
to
I
n
d
ian
lan
g
u
ag
e
p
air
s
[
32
]
D
a
t
a
s
e
t
l
a
n
g
u
a
g
e
p
a
i
r
N
u
mb
e
r
o
f
se
n
t
e
n
c
e
s
S
o
u
r
c
e
G
e
n
e
r
a
l
C
o
r
p
u
s
En
g
l
i
sh
-
H
i
n
d
i
1
,
0
0
0
,
0
0
0
P
a
r
a
l
l
e
l
C
o
r
p
u
s
G
e
n
e
r
a
l
C
o
r
p
u
s
En
g
l
i
sh
-
B
e
n
g
a
l
i
8
0
0
,
0
0
0
P
a
r
a
l
l
e
l
C
o
r
p
u
s
D
o
ma
i
n
-
s
p
e
c
i
f
i
c
En
g
l
i
sh
-
Ta
mi
l
2
0
0
,
0
0
0
P
a
r
a
l
l
e
l
C
o
r
p
u
s
M
o
n
o
l
i
n
g
u
a
l
C
o
r
p
u
s
H
i
n
d
i
,
B
e
n
g
a
l
i
,
T
a
mi
l
2
,
0
0
0
,
0
0
0
(
e
a
c
h
)
M
o
n
o
l
i
n
g
u
a
l
C
o
r
p
u
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
B
r
id
g
in
g
th
e
lin
g
u
is
tic
d
ivid
e:
r
ec
en
t d
ev
elo
p
men
ts
in
ma
ch
i
n
e
tr
a
n
s
la
tio
n
fo
r
… (
Ja
y
a
n
a
n
d
A
.
K
a
mb
le
)
1281
Sin
g
h
et
a
l.
[
5
7
]
o
n
h
o
w
B
E
R
T
,
GPT
-
3
(
L
L
Ms)
wo
r
k
we
ll
(
o
r
n
o
t
r
ea
lly
)
at
tr
a
n
s
latio
n
b
etwe
en
E
n
g
lis
h
an
d
I
n
d
ian
lan
g
u
ag
es
in
th
eir
p
ap
e
r
,
E
v
al
u
atin
g
T
r
an
s
latio
n
C
ap
ab
ilit
ies
o
f
L
L
M
s
f
o
r
E
n
g
lis
h
an
d
I
n
d
ian
L
a
n
g
u
a
g
es
co
n
s
id
er
(
b
o
th
au
to
m
ated
an
d
h
u
m
an
)
ev
alu
atio
n
ap
p
r
o
ac
h
es
to
m
e
asu
r
e
th
e
q
u
ality
o
f
tr
an
s
latio
n
s
b
etwe
en
lan
g
u
ag
es.
T
h
ey
ar
e
co
m
b
in
e
d
with
h
u
m
an
f
lu
e
n
cy
,
ac
c
u
r
ac
y
an
d
cu
ltu
r
al
r
ef
er
en
ce
s
cr
iter
ia
b
ased
o
n
au
to
m
atic
m
ea
s
u
r
es
lik
e
B
L
E
U,
T
E
R
,
ME
T
E
OR
.
T
h
e
team
is
em
p
lo
y
in
g
tr
a
n
s
f
o
r
m
e
r
m
o
d
els with
atten
tio
n
to
b
o
o
s
t
tr
an
s
latio
n
q
u
ality
.
T
ab
le
6
s
h
o
ws
th
eir
en
g
-
in
d
p
air
s
an
d
as
s
u
ch
th
eir
p
ar
allel
co
r
p
u
s
[
2
3
]
.
T
h
e
r
es
o
u
r
ce
s
f
o
r
E
n
g
lis
h
-
Hin
d
i
(
2
,
0
0
0
,
0
0
0
s
e
n
ten
ce
p
air
s
)
ar
e
wea
lth
ier
an
d
th
en
f
o
l
-
lo
wed
b
y
E
n
g
lis
h
-
B
en
g
ali
with
1
,
5
0
0
,
0
0
0
.
Ma
r
ath
i
T
am
il
T
el
u
g
u
an
d
Gu
ja
r
ati
ar
e
th
e
o
th
e
r
th
r
ee
wh
ich
h
as
s
ca
le
d
ataset
(
8
0
0
0
0
0
-
1
2
0
0
0
0
0
s
en
ten
ce
p
air
s
)
.
T
h
is
s
p
an
s
h
o
ws th
e
ab
ilit
y
o
f
th
e
m
o
d
el
to
t
ac
k
le
in
ter
-
I
n
d
ian
lan
g
u
ag
e
tr
a
n
s
latab
ilit
y
.
I
n
ter
m
s
o
f
p
er
f
o
r
m
an
ce
(
T
ab
le
7
)
E
n
g
lis
h
-
Hin
d
i
is
th
e
b
est,
with
3
2
.
5
B
L
E
U
s
o
r
e,
an
d
E
n
g
lis
h
-
B
en
g
ali
it
th
e
s
ec
o
n
d
-
b
est
with
2
8
.
9
On
th
e
lo
wer
s
id
e,
we
h
av
e
r
esp
ec
tab
le
B
L
E
U
s
co
r
es
o
f
2
1
.
5
an
d
2
2
.
8
f
o
r
E
n
g
lis
h
-
Ass
am
ese
an
d
E
n
g
lis
h
-
Or
iy
a
s
o
wh
ich
in
d
icate
s
th
at
th
e
q
u
ality
o
f
t
r
an
s
latio
n
is
n
o
t
as
g
r
ea
t
h
e
r
e.
T
h
is
is
also
s
u
p
p
o
r
ted
b
y
T
E
R
s
co
r
es
an
d
b
o
th
I
n
d
o
HM
+I
n
d
o
B
ar
e
m
u
ch
less
er
r
o
n
eo
u
s
co
m
p
ar
ed
to
er
r
o
r
-
p
r
o
n
e
E
n
g
li
s
h
-
Ass
am
ese
an
d
E
n
g
lig
h
-
Or
i
y
a.
Ho
wev
er
,
f
o
r
all
th
at
p
r
o
m
is
e
th
er
e
is
o
n
e
b
ig
d
r
awb
ac
k
as
in
d
icate
d
in
th
is
s
tu
d
y
:
y
o
u
n
ee
d
m
ass
iv
e
co
m
p
u
te
r
eso
u
r
ce
s
to
tan
g
le
with
th
ese
L
L
Ms.
T
h
is
co
u
ld
b
e
a
b
o
ttlen
ec
k
p
ar
ticu
lar
ly
in
r
e
s
o
u
r
ce
co
n
s
tr
ain
ed
s
ettin
g
s
,
wh
ich
co
n
f
in
es
th
e
u
tili
ty
o
f
th
ese
m
o
d
els
f
o
r
r
ea
l
tr
an
s
latio
n
task
s
.
L
astl
y
,
th
e
L
L
Ms
ca
n
b
e
co
n
s
id
er
ed
a
b
asic
s
o
lu
tio
n
to
im
p
r
o
v
e
tr
an
s
latio
n
s
y
s
tem
s
f
o
r
E
n
g
lis
h
-
I
n
d
ian
lan
g
u
ag
e
p
a
ir
s
;
h
o
wev
er
,
th
eir
u
s
ag
e
will r
em
ain
lim
ited
b
ec
a
u
s
e
o
f
th
is
h
ig
h
co
m
p
u
tatio
n
al
r
eq
u
ir
em
e
n
t.
T
ab
le
6
.
Par
allel
co
r
p
o
r
a
f
o
r
E
n
g
lis
h
an
d
I
n
d
ian
lan
g
u
a
g
es
La
n
g
u
a
g
e
p
a
i
r
N
u
mb
e
r
o
f
se
n
t
e
n
c
e
s
En
g
l
i
sh
-
H
i
n
d
i
2
,
0
0
0
,
0
0
0
En
g
l
i
sh
-
B
e
n
g
a
l
i
1
,
5
0
0
,
0
0
0
En
g
l
i
sh
-
M
a
r
a
t
h
i
1
,
2
0
0
,
0
0
0
En
g
l
i
sh
-
T
a
mi
l
1
,
0
0
0
,
0
0
0
En
g
l
i
sh
-
T
e
l
u
g
u
1
,
0
0
0
,
0
0
0
En
g
l
i
sh
-
G
u
j
a
r
a
t
i
8
0
0
,
0
0
0
En
g
l
i
sh
-
K
a
n
n
a
d
a
8
0
0
,
0
0
0
En
g
l
i
sh
-
M
a
l
a
y
a
l
a
m
7
0
0
,
0
0
0
En
g
l
i
sh
-
P
u
n
j
a
b
i
6
0
0
,
0
0
0
En
g
l
i
sh
-
O
r
i
y
a
5
0
0
,
0
0
0
En
g
l
i
sh
-
A
ssa
mese
4
0
0
,
0
0
0
T
ab
le
7
.
Au
to
m
ated
m
etr
ics r
e
s
u
lt
La
n
g
u
a
g
e
p
a
i
r
B
LEU
TER
En
g
l
i
sh
-
H
i
n
d
i
3
2
.
5
5
6
.
2
En
g
l
i
sh
-
B
e
n
g
a
l
i
2
8
.
9
5
9
.
3
En
g
l
i
sh
-
M
a
r
a
t
h
i
2
5
.
7
6
1
.
7
En
g
l
i
sh
-
T
a
mi
l
2
4
.
6
6
2
.
5
En
g
l
i
sh
-
T
e
l
u
g
u
2
4
.
8
6
1
.
9
En
g
l
i
sh
-
G
u
j
a
r
a
t
i
2
7
.
2
6
0
.
1
En
g
l
i
sh
-
K
a
n
n
a
d
a
2
6
.
5
6
0
.
8
En
g
l
i
sh
-
M
a
l
a
y
a
l
a
m
2
3
.
4
63
En
g
l
i
sh
-
P
u
n
j
a
b
i
2
5
.
1
6
2
.
2
En
g
l
i
sh
-
O
r
i
y
a
2
2
.
8
6
4
.
3
En
g
l
i
sh
-
A
ssa
mese
2
1
.
5
65
3.
R
E
SU
L
T
S A
N
D
D
ISC
U
S
SIO
N
3
.
1
.
B
L
E
U
Sco
re
I
m
pro
v
ement
s
f
ro
m
SM
T
+
B
P
E
T
ab
le
8
s
h
o
ws
th
e
B
L
E
U
s
co
r
in
g
o
f
two
tr
an
s
latio
n
m
o
d
els
to
war
d
s
lan
g
u
ag
e
p
air
s
,
with
wh
ich
we
o
b
s
er
v
e
h
o
w
B
PE
in
f
lu
en
ce
s
t
r
an
s
latio
n
q
u
ality
.
T
h
e
s
co
r
es o
f
B
OT
H
th
e
m
o
d
els
in
cr
ea
s
e
d
s
ig
n
if
ican
tly
af
ter
B
PE
a
s
ca
n
b
e
s
ee
n
in
F
ig
u
r
e
2
with
m
o
d
el
1
im
p
r
o
v
in
g
f
r
o
m
2
5
.
3
to
3
0
.
7
an
d
m
o
d
el
2
f
r
o
m
2
2
.
8
to
2
8
.
4
,
d
en
o
tin
g
t
h
at
B
PE
lead
s
to
m
u
ch
im
p
r
o
v
e
d
tr
an
s
latio
n
ac
c
u
r
ac
y
an
d
f
lu
e
n
cy
.
W
h
ich
m
e
an
s
,
o
n
ly
r
u
le
b
ase
d
SMT
h
as
b
ee
n
th
e
p
ar
t
o
f
r
esear
ch
.
I
t
d
o
es
n
o
t
u
s
e
r
ec
en
tly
d
ev
elo
p
ed
m
et
h
o
d
s
in
NM
T
th
at
m
ig
h
t
p
r
o
d
u
c
e
b
etter
m
u
ltil
in
g
u
al
tr
a
n
s
latio
n
s
an
d
m
o
r
e
ef
f
ec
tiv
el
y
ad
d
r
ess
l
in
g
u
is
tic
co
m
p
lex
ities
.
T
ab
le
8
.
B
L
E
U
s
co
r
es in
d
icati
n
g
en
h
an
ce
d
tr
a
n
s
latio
n
q
u
alit
y
[
51
]
M
o
d
e
l
La
n
g
u
a
g
e
p
a
i
r
B
LEU
s
c
o
r
e
(
B
e
f
o
r
e
B
P
E)
B
LEU
s
c
o
r
e
(
A
f
t
e
r
B
P
E)
M
o
d
e
l
1
La
n
g
u
a
g
e
A
-
B
2
5
.
3
3
0
.
7
M
o
d
e
l
2
La
n
g
u
a
g
e
B
-
C
2
2
.
8
2
8
.
4
Fig
u
r
e
2
.
B
L
E
U
s
co
r
e
im
p
r
o
v
em
en
ts
ac
h
iev
ed
b
y
u
s
in
g
B
PE
Evaluation Warning : The document was created with Spire.PDF for Python.