<s>
Explanation-based	B-General_Concept
learning	I-General_Concept
(	O
EBL	O
)	O
is	O
a	O
form	O
of	O
machine	O
learning	O
that	O
exploits	O
a	O
very	O
strong	O
,	O
or	O
even	O
perfect	O
,	O
domain	B-General_Concept
theory	O
(	O
i.e.	O
</s>
<s>
a	O
formal	O
theory	O
of	O
an	O
application	O
domain	B-General_Concept
akin	O
to	O
a	O
domain	B-Application
model	I-Application
in	O
ontology	O
engineering	O
,	O
not	O
to	O
be	O
confused	O
with	O
Scott	O
's	O
domain	B-General_Concept
theory	O
)	O
in	O
order	O
to	O
make	O
generalizations	O
or	O
form	O
concepts	O
from	O
training	O
examples	O
.	O
</s>
<s>
An	O
example	O
of	O
EBL	O
using	O
a	O
perfect	O
domain	B-General_Concept
theory	O
is	O
a	O
program	O
that	O
learns	O
to	O
play	O
chess	B-Application
through	O
example	O
.	O
</s>
<s>
A	O
specific	O
chess	B-Application
position	O
that	O
contains	O
an	O
important	O
feature	O
such	O
as	O
"	O
Forced	O
loss	O
of	O
black	O
queen	O
in	O
two	O
moves	O
"	O
includes	O
many	O
irrelevant	O
features	O
,	O
such	O
as	O
the	O
specific	O
scattering	O
of	O
pawns	O
on	O
the	O
board	O
.	O
</s>
<s>
A	O
domain	B-General_Concept
theory	O
is	O
perfect	O
or	O
complete	O
if	O
it	O
contains	O
,	O
in	O
principle	O
,	O
all	O
information	O
needed	O
to	O
decide	O
any	O
question	O
about	O
the	O
domain	B-General_Concept
.	O
</s>
<s>
For	O
example	O
,	O
the	O
domain	B-General_Concept
theory	O
for	O
chess	B-Application
is	O
simply	O
the	O
rules	O
of	O
chess	B-Application
.	O
</s>
<s>
EBL	O
uses	O
training	O
examples	O
to	O
make	O
searching	O
for	O
deductive	O
consequences	O
of	O
a	O
domain	B-General_Concept
theory	O
efficient	O
in	O
practice	O
.	O
</s>
<s>
In	O
essence	O
,	O
an	O
EBL	O
system	O
works	O
by	O
finding	O
a	O
way	O
to	O
deduce	O
each	O
training	O
example	O
from	O
the	O
system	O
's	O
existing	O
database	O
of	O
domain	B-General_Concept
theory	O
.	O
</s>
<s>
Having	O
a	O
short	O
proof	O
of	O
the	O
training	O
example	O
extends	O
the	O
domain-theory	O
database	O
,	O
enabling	O
the	O
EBL	O
system	O
to	O
find	O
and	O
classify	O
future	O
examples	O
that	O
are	O
similar	O
to	O
the	O
training	O
example	O
very	O
quickly	O
.	O
</s>
<s>
operationality	O
criteria	O
(	O
criteria	O
for	O
determining	O
which	O
features	O
in	O
the	O
domain	B-General_Concept
are	O
efficiently	O
recognizable	O
,	O
e.g.	O
</s>
<s>
An	O
especially	O
good	O
application	O
domain	B-General_Concept
for	O
an	O
EBL	O
is	O
natural	O
language	O
processing	O
(	O
NLP	O
)	O
.	O
</s>
<s>
Here	O
a	O
rich	O
domain	B-General_Concept
theory	O
,	O
i.e.	O
,	O
a	O
natural	O
language	O
grammar	O
—	O
although	O
neither	O
perfect	O
nor	O
complete	O
,	O
is	O
tuned	O
to	O
a	O
particular	O
application	O
or	O
particular	O
language	O
usage	O
,	O
using	O
a	O
treebank	O
(	O
training	O
examples	O
)	O
.	O
</s>
<s>
The	O
method	O
has	O
been	O
successfully	O
applied	O
to	O
several	O
large-scale	O
natural	O
language	O
parsing	O
systems	O
,	O
where	O
the	O
utility	O
problem	O
was	O
solved	O
by	O
omitting	O
the	O
original	O
grammar	O
(	O
domain	B-General_Concept
theory	O
)	O
and	O
using	O
specialized	O
LR-parsing	O
techniques	O
,	O
resulting	O
in	O
huge	O
speed-ups	O
,	O
at	O
a	O
cost	O
in	O
coverage	O
,	O
but	O
with	O
a	O
gain	O
in	O
disambiguation	O
.	O
</s>
<s>
Note	O
how	O
the	O
utility	O
problem	O
,	O
first	O
exposed	O
by	O
Minton	O
,	O
was	O
solved	O
by	O
discarding	O
the	O
original	O
grammar/domain	O
theory	O
,	O
and	O
that	O
the	O
quoted	O
articles	O
tend	O
to	O
contain	O
the	O
phrase	O
grammar	O
specialization	O
—	O
quite	O
the	O
opposite	O
of	O
the	O
original	O
term	O
explanation-based	O
generalization	O
.	O
</s>
