<s>
Automated	B-General_Concept
essay	I-General_Concept
scoring	I-General_Concept
(	O
AES	O
)	O
is	O
the	O
use	O
of	O
specialized	O
computer	O
programs	O
to	O
assign	O
grades	O
to	O
essays	O
written	O
in	O
an	O
educational	O
setting	O
.	O
</s>
<s>
It	O
is	O
a	O
form	O
of	O
educational	O
assessment	O
and	O
an	O
application	O
of	O
natural	B-Language
language	I-Language
processing	I-Language
.	O
</s>
<s>
Therefore	O
,	O
it	O
can	O
be	O
considered	O
a	O
problem	O
of	O
statistical	B-General_Concept
classification	I-General_Concept
.	O
</s>
<s>
Educational	O
Testing	O
Service	O
offers	O
"	O
e-rater	O
"	O
,	O
an	O
automated	B-General_Concept
essay	I-General_Concept
scoring	I-General_Concept
program	O
.	O
</s>
<s>
In	O
2012	O
,	O
the	O
Hewlett	O
Foundation	O
sponsored	O
a	O
competition	O
on	O
Kaggle	B-Application
called	O
the	O
Automated	O
Student	O
Assessment	O
Prize	O
(	O
ASAP	O
)	O
.	O
</s>
<s>
Although	O
the	O
investigators	O
reported	O
that	O
the	O
automated	B-General_Concept
essay	I-General_Concept
scoring	I-General_Concept
was	O
as	O
reliable	O
as	O
human	O
scoring	O
,	O
this	O
claim	O
was	O
not	O
substantiated	O
by	O
any	O
statistical	O
tests	O
because	O
some	O
of	O
the	O
vendors	O
required	O
that	O
no	O
such	O
tests	O
be	O
performed	O
as	O
a	O
precondition	O
for	O
their	O
participation	O
.	O
</s>
<s>
Results	O
of	O
supervised	B-General_Concept
learning	I-General_Concept
demonstrate	O
that	O
the	O
automatic	O
systems	O
perform	O
well	O
when	O
marking	O
by	O
different	O
human	O
teachers	O
is	O
in	O
good	O
agreement	O
.	O
</s>
<s>
Unsupervised	O
clustering	B-Algorithm
of	O
answers	O
showed	O
that	O
excellent	O
papers	O
and	O
weak	O
papers	O
formed	O
well-defined	O
clusters	O
,	O
and	O
the	O
automated	O
marking	O
rule	O
for	O
these	O
clusters	O
worked	O
well	O
,	O
whereas	O
marks	O
given	O
by	O
human	O
teachers	O
for	O
the	O
third	O
cluster	O
( 	O
 '	O
mixed	O
 '	O
)	O
can	O
be	O
controversial	O
,	O
and	O
the	O
reliability	O
of	O
any	O
assessment	O
of	O
works	O
from	O
the	O
'	O
mixed	O
 '	O
cluster	O
can	O
often	O
be	O
questioned	O
(	O
both	O
human	O
and	O
computer-based	O
)	O
.	O
</s>
<s>
Due	O
to	O
the	O
growing	O
popularity	O
of	O
deep	O
neural	O
networks	O
,	O
deep	O
learning	O
approaches	O
have	O
been	O
adopted	O
for	O
automated	B-General_Concept
essay	I-General_Concept
scoring	I-General_Concept
,	O
generally	O
obtaining	O
superior	O
results	O
,	O
often	O
surpassing	O
inter-human	O
agreement	O
levels	O
.	O
</s>
<s>
Early	O
attempts	O
used	O
linear	B-General_Concept
regression	I-General_Concept
.	O
</s>
<s>
Modern	O
systems	O
may	O
use	O
linear	B-General_Concept
regression	I-General_Concept
or	O
other	O
machine	O
learning	O
techniques	O
often	O
in	O
combination	O
with	O
other	O
statistical	O
techniques	O
such	O
as	O
latent	O
semantic	O
analysis	O
and	O
Bayesian	O
inference	O
.	O
</s>
<s>
The	O
automated	B-General_Concept
essay	I-General_Concept
scoring	I-General_Concept
task	O
has	O
also	O
been	O
studied	O
in	O
the	O
cross-domain	O
setting	O
using	O
machine	O
learning	O
models	O
,	O
where	O
the	O
models	O
are	O
trained	O
on	O
essays	O
written	O
for	O
one	O
prompt	O
(	O
topic	O
)	O
and	O
tested	O
on	O
essays	O
written	O
for	O
another	O
prompt	O
.	O
</s>
<s>
Among	O
them	O
are	O
percent	O
agreement	O
,	O
Scott	O
's	O
π	O
,	O
Cohen	B-General_Concept
's	I-General_Concept
κ	I-General_Concept
,	O
Krippendorf	O
's	O
α	O
,	O
Pearson	O
's	O
correlation	O
coefficient	O
r	O
,	O
Spearman	B-General_Concept
's	I-General_Concept
rank	I-General_Concept
correlation	I-General_Concept
coefficient	I-General_Concept
ρ	O
,	O
and	O
Lin	O
's	O
concordance	O
correlation	O
coefficient	O
.	O
</s>
<s>
Most	O
resources	O
for	O
automated	B-General_Concept
essay	I-General_Concept
scoring	I-General_Concept
are	O
proprietary	O
.	O
</s>
