<s>
Various	O
methods	O
for	O
the	O
evaluation	O
for	O
machine	B-Application
translation	I-Application
have	O
been	O
employed	O
.	O
</s>
<s>
This	O
article	O
focuses	O
on	O
the	O
evaluation	O
of	O
the	O
output	O
of	O
machine	B-Application
translation	I-Application
,	O
rather	O
than	O
on	O
performance	O
or	O
usability	O
evaluation	O
.	O
</s>
<s>
A	O
typical	O
way	O
for	O
lay	O
people	O
to	O
assess	O
machine	B-Application
translation	I-Application
quality	O
is	O
to	O
translate	O
from	O
a	O
source	O
language	O
to	O
a	O
target	O
language	O
and	O
back	O
to	O
the	O
source	O
language	O
with	O
the	O
same	O
engine	O
.	O
</s>
<s>
While	O
round-trip	O
translation	O
may	O
be	O
useful	O
to	O
generate	O
a	O
"	O
surplus	O
of	O
fun	O
,	O
"	O
the	O
methodology	O
is	O
deficient	O
for	O
serious	O
study	O
of	O
machine	B-Application
translation	I-Application
quality	O
.	O
</s>
<s>
This	O
section	O
covers	O
two	O
of	O
the	O
large	O
scale	O
evaluation	O
studies	O
that	O
have	O
had	O
significant	O
impact	O
on	O
the	O
field	O
—	O
the	O
ALPAC	B-General_Concept
1966	O
study	O
and	O
the	O
ARPA	O
study	O
.	O
</s>
<s>
One	O
of	O
the	O
constituent	O
parts	O
of	O
the	O
ALPAC	B-General_Concept
report	I-General_Concept
was	O
a	O
study	O
comparing	O
different	O
levels	O
of	O
human	O
translation	O
with	O
machine	B-Application
translation	I-Application
output	O
,	O
using	O
human	O
subjects	O
as	O
judges	O
.	O
</s>
<s>
The	O
study	O
concluded	O
that	O
,	O
"	O
highly	O
reliable	O
assessments	O
can	O
be	O
made	O
of	O
the	O
quality	O
of	O
human	O
and	O
machine	B-Application
translations	I-Application
"	O
.	O
</s>
<s>
As	O
part	O
of	O
the	O
Human	O
Language	O
Technologies	O
Program	O
,	O
the	O
Advanced	O
Research	O
Projects	O
Agency	O
(	O
ARPA	O
)	O
created	O
a	O
methodology	O
to	O
evaluate	O
machine	B-Application
translation	I-Application
systems	I-Application
,	O
and	O
continues	O
to	O
perform	O
evaluations	O
based	O
on	O
this	O
methodology	O
.	O
</s>
<s>
These	O
articles	O
were	O
translated	O
by	O
professional	O
translators	O
into	O
a	O
series	O
of	O
language	O
pairs	O
,	O
and	O
then	O
translated	O
back	O
into	O
English	O
using	O
the	O
machine	B-Application
translation	I-Application
systems	I-Application
.	O
</s>
<s>
This	O
was	O
good	O
from	O
the	O
point	O
of	O
view	O
that	O
the	O
metric	O
was	O
"	O
externally	O
motivated	O
"	O
,	O
since	O
it	O
was	O
not	O
specifically	O
developed	O
for	O
machine	B-Application
translation	I-Application
.	O
</s>
<s>
A	O
metric	O
that	O
evaluates	O
machine	B-Application
translation	I-Application
output	O
represents	O
the	O
quality	O
of	O
the	O
output	O
.	O
</s>
<s>
The	O
aim	O
of	O
this	O
subsection	O
is	O
to	O
give	O
an	O
overview	O
of	O
the	O
state	O
of	O
the	O
art	O
in	O
automatic	O
metrics	O
for	O
evaluating	O
machine	B-Application
translation	I-Application
.	O
</s>
<s>
machine	B-Application
translation	I-Application
is	O
to	O
a	O
professional	O
human	O
translation	O
,	O
the	O
better	O
it	O
is	O
"	O
.	O
</s>
<s>
The	O
metric	O
modifies	O
simple	O
precision	O
since	O
machine	B-Application
translation	I-Application
systems	I-Application
have	O
been	O
known	O
to	O
generate	O
more	O
words	O
than	O
appear	O
in	O
a	O
reference	O
text	O
.	O
</s>
<s>
No	O
other	O
machine	B-Application
translation	I-Application
metric	O
is	O
yet	O
to	O
significantly	O
outperform	O
BLEU	O
with	O
respect	O
to	O
correlation	O
with	O
human	O
judgment	O
across	O
language	O
pairs	O
.	O
</s>
<s>
Where	O
BLEU	O
simply	O
calculates	O
n-gram	B-Language
precision	O
adding	O
equal	O
weight	O
to	O
each	O
one	O
,	O
NIST	O
also	O
calculates	O
how	O
informative	O
a	O
particular	O
n-gram	B-Language
is	O
.	O
</s>
<s>
That	O
is	O
to	O
say	O
,	O
when	O
a	O
correct	O
n-gram	B-Language
is	O
found	O
,	O
the	O
rarer	O
that	O
n-gram	B-Language
is	O
,	O
the	O
more	O
weight	O
it	O
is	O
given	O
.	O
</s>
<s>
It	O
was	O
originally	O
used	O
for	O
measuring	O
the	O
performance	O
of	O
speech	B-Application
recognition	I-Application
systems	O
but	O
is	O
also	O
used	O
in	O
the	O
evaluation	B-General_Concept
of	I-General_Concept
machine	I-General_Concept
translation	I-General_Concept
.	O
</s>
<s>
The	O
metric	O
is	O
based	O
on	O
the	O
weighted	O
harmonic	O
mean	O
of	O
unigram	B-Language
precision	O
and	O
unigram	B-Language
recall	O
.	O
</s>
<s>
A	O
new	O
MT	O
evaluation	O
metric	O
LEPOR	O
was	O
proposed	O
as	O
the	O
combination	O
of	O
many	O
evaluation	O
factors	O
including	O
existing	O
ones	O
(	O
precision	O
,	O
recall	O
)	O
and	O
modified	O
ones	O
(	O
sentence-length	O
penalty	O
and	O
n-gram	B-Language
based	O
word	O
order	O
penalty	O
)	O
.	O
</s>
<s>
There	O
are	O
some	O
machine	B-Application
translation	I-Application
evaluation	O
survey	O
works	O
,	O
where	O
people	O
introduced	O
more	O
details	O
about	O
what	O
kinds	O
of	O
human	O
evaluation	O
methods	O
they	O
used	O
and	O
how	O
they	O
work	O
,	O
such	O
as	O
the	O
intelligibility	O
,	O
fidelity	O
,	O
fluency	O
,	O
adequacy	O
,	O
comprehension	O
,	O
and	O
informativeness	O
,	O
etc	O
.	O
</s>
<s>
Some	O
state-of-the-art	O
overview	O
on	O
both	O
manual	O
and	O
automatic	B-Application
translation	I-Application
evaluation	O
introduced	O
the	O
recently	O
developed	O
translation	O
quality	O
assessment	O
(	O
TQA	O
)	O
methodologies	O
,	O
such	O
as	O
the	O
crowd-sourced	O
intelligence	O
Amazon	O
Mechanical	O
Turk	O
utilization	O
,	O
statistical	O
significance	O
testing	O
,	O
re-visiting	O
traditional	O
criteria	O
with	O
newly	O
designed	O
strategies	O
,	O
as	O
well	O
as	O
MT	O
quality	O
estimation	O
(	O
QE	O
)	O
shared	O
tasks	O
from	O
the	O
annual	O
workshop	O
on	O
MT	O
(	O
WMT	O
)	O
and	O
corresponding	O
models	O
that	O
do	O
not	O
rely	O
on	O
human	O
offered	O
reference	O
translations	O
.	O
</s>
