<s>
Automatic	B-Application
summarization	I-Application
is	O
the	O
process	O
of	O
shortening	O
a	O
set	O
of	O
data	O
computationally	O
,	O
to	O
create	O
a	O
subset	O
(	O
a	O
summary	O
)	O
that	O
represents	O
the	O
most	O
important	O
or	O
relevant	O
information	O
within	O
the	O
original	O
content	O
.	O
</s>
<s>
Artificial	B-Application
intelligence	I-Application
algorithms	O
are	O
commonly	O
developed	O
and	O
employed	O
to	O
achieve	O
this	O
,	O
specialized	O
for	O
different	O
types	O
of	O
data	O
.	O
</s>
<s>
Text	B-Application
summarization	I-Application
is	O
usually	O
implemented	O
by	O
natural	B-Language
language	I-Language
processing	I-Language
methods	O
,	O
designed	O
to	O
locate	O
the	O
most	O
informative	O
sentences	O
in	O
a	O
given	O
document	O
.	O
</s>
<s>
On	O
the	O
other	O
hand	O
,	O
visual	O
content	O
can	O
be	O
summarized	O
using	O
computer	B-Application
vision	I-Application
algorithms	O
.	O
</s>
<s>
Image	B-Application
summarization	I-Application
is	O
the	O
subject	O
of	O
ongoing	O
research	O
;	O
existing	O
approaches	O
typically	O
attempt	O
to	O
display	O
the	O
most	O
representative	O
images	O
from	O
a	O
given	O
image	O
collection	O
,	O
or	O
generate	O
a	O
video	O
that	O
only	O
includes	O
the	O
most	O
important	O
content	O
from	O
the	O
entire	O
collection	O
.	O
</s>
<s>
Video	B-Application
summarization	I-Application
algorithms	O
identify	O
and	O
extract	O
from	O
the	O
original	O
video	O
content	O
the	O
most	O
important	O
frames	O
(	O
key-frames	O
)	O
,	O
and/or	O
the	O
most	O
important	O
video	O
segments	O
(	O
key-shots	O
)	O
,	O
normally	O
in	O
a	O
temporally	O
ordered	O
fashion	O
.	O
</s>
<s>
In	O
2022	O
Google	B-Application
Docs	I-Application
released	O
an	O
automatic	B-Application
summarization	I-Application
feature	O
.	O
</s>
<s>
There	O
are	O
two	O
general	O
approaches	O
to	O
automatic	B-Application
summarization	I-Application
:	O
extraction	B-General_Concept
and	O
abstraction	O
.	O
</s>
<s>
For	O
text	O
,	O
extraction	B-General_Concept
is	O
analogous	O
to	O
the	O
process	O
of	O
skimming	O
,	O
where	O
the	O
summary	O
(	O
if	O
available	O
)	O
,	O
headings	O
and	O
subheadings	O
,	O
figures	O
,	O
the	O
first	O
and	O
last	O
paragraphs	O
of	O
a	O
section	O
,	O
and	O
optionally	O
the	O
first	O
and	O
last	O
sentences	O
in	O
a	O
paragraph	O
are	O
read	O
before	O
one	O
chooses	O
to	O
read	O
the	O
entire	O
document	O
in	O
detail	O
.	O
</s>
<s>
Other	O
examples	O
of	O
extraction	B-General_Concept
that	O
include	O
key	O
sequences	O
of	O
text	O
in	O
terms	O
of	O
clinical	O
relevance	O
(	O
including	O
patient/problem	O
,	O
intervention	O
,	O
and	O
outcome	O
)	O
.	O
</s>
<s>
Abstraction	O
may	O
transform	O
the	O
extracted	O
content	O
by	O
paraphrasing	B-General_Concept
sections	O
of	O
the	O
source	O
document	O
,	O
to	O
condense	O
a	O
text	O
more	O
strongly	O
than	O
extraction	B-General_Concept
.	O
</s>
<s>
Such	O
transformation	O
,	O
however	O
,	O
is	O
computationally	O
much	O
more	O
challenging	O
than	O
extraction	B-General_Concept
,	O
involving	O
both	O
natural	B-Language
language	I-Language
processing	I-Language
and	O
often	O
a	O
deep	O
understanding	O
of	O
the	O
domain	O
of	O
the	O
original	O
text	O
in	O
cases	O
where	O
the	O
original	O
document	O
relates	O
to	O
a	O
special	O
field	O
of	O
knowledge	O
.	O
</s>
<s>
"	O
Paraphrasing	B-General_Concept
"	O
is	O
even	O
more	O
difficult	O
to	O
apply	O
to	O
image	O
and	O
video	O
,	O
which	O
is	O
why	O
most	O
summarization	O
systems	O
are	O
extractive	O
.	O
</s>
<s>
The	O
second	O
is	O
query	B-Library
relevant	O
summarization	O
,	O
sometimes	O
called	O
query-based	O
summarization	O
,	O
which	O
summarizes	O
objects	O
specific	O
to	O
a	O
query	B-Library
.	O
</s>
<s>
Summarization	O
systems	O
are	O
able	O
to	O
create	O
both	O
query	B-Library
relevant	O
text	B-Application
summaries	I-Application
and	O
generic	O
machine-generated	O
summaries	O
depending	O
on	O
what	O
the	O
user	O
needs	O
.	O
</s>
<s>
An	O
example	O
of	O
a	O
summarization	O
problem	O
is	O
document	B-Application
summarization	I-Application
,	O
which	O
attempts	O
to	O
automatically	O
produce	O
an	O
abstract	O
from	O
a	O
given	O
document	O
.	O
</s>
<s>
Sometimes	O
one	O
might	O
be	O
interested	O
in	O
generating	O
a	O
summary	O
from	O
a	O
single	O
source	O
document	O
,	O
while	O
others	O
can	O
use	O
multiple	O
source	O
documents	O
(	O
for	O
example	O
,	O
a	O
cluster	B-Algorithm
of	O
articles	O
on	O
the	O
same	O
topic	O
)	O
.	O
</s>
<s>
This	O
problem	O
is	O
called	O
multi-document	B-General_Concept
summarization	I-General_Concept
.	O
</s>
<s>
Image	O
collection	O
summarization	O
is	O
another	O
application	O
example	O
of	O
automatic	B-Application
summarization	I-Application
.	O
</s>
<s>
A	O
summary	O
in	O
this	O
context	O
is	O
useful	O
to	O
show	O
the	O
most	O
representative	O
images	O
of	O
results	O
in	O
an	O
image	B-General_Concept
collection	I-General_Concept
exploration	I-General_Concept
system	O
.	O
</s>
<s>
Video	B-Application
summarization	I-Application
is	O
a	O
related	O
domain	O
,	O
where	O
the	O
system	O
automatically	O
creates	O
a	O
trailer	O
of	O
a	O
long	O
video	O
.	O
</s>
<s>
Query	B-Library
based	O
summarization	O
techniques	O
,	O
additionally	O
model	O
for	O
relevance	O
of	O
the	O
summary	O
with	O
the	O
query	B-Library
.	O
</s>
<s>
Some	O
techniques	O
and	O
algorithms	O
which	O
naturally	O
model	O
summarization	O
problems	O
are	O
TextRank	O
and	O
PageRank	B-Algorithm
,	O
Submodular	B-Algorithm
set	I-Algorithm
function	I-Algorithm
,	O
Determinantal	O
point	O
process	O
,	O
maximal	O
marginal	O
relevance	O
(	O
MMR	O
)	O
etc	O
.	O
</s>
<s>
Abstraction	O
requires	O
a	O
deep	O
understanding	B-General_Concept
of	I-General_Concept
the	I-General_Concept
text	I-General_Concept
,	O
which	O
makes	O
it	O
difficult	O
for	O
a	O
computer	O
system	O
.	O
</s>
<s>
They	O
can	O
enable	O
document	O
browsing	O
by	O
providing	O
a	O
short	O
summary	O
,	O
improve	O
information	B-Library
retrieval	I-Library
(	O
if	O
documents	O
have	O
keyphrases	O
assigned	O
,	O
a	O
user	O
could	O
search	O
by	O
keyphrase	O
to	O
produce	O
more	O
reliable	O
hits	O
than	O
a	O
full-text	B-Application
search	I-Application
)	O
,	O
and	O
be	O
employed	O
in	O
generating	O
index	O
entries	O
for	O
a	O
large	O
text	O
corpus	O
.	O
</s>
<s>
Depending	O
on	O
the	O
different	O
literature	O
and	O
the	O
definition	O
of	O
key	O
terms	O
,	O
words	O
or	O
phrases	O
,	O
keyword	B-General_Concept
extraction	I-General_Concept
is	O
a	O
highly	O
related	O
theme	O
.	O
</s>
<s>
Beginning	O
with	O
the	O
work	O
of	O
Turney	O
,	O
many	O
researchers	O
have	O
approached	O
keyphrase	O
extraction	B-General_Concept
as	O
a	O
supervised	B-General_Concept
machine	I-General_Concept
learning	I-General_Concept
problem	O
.	O
</s>
<s>
Given	O
a	O
document	O
,	O
we	O
construct	O
an	O
example	O
for	O
each	O
unigram	B-Language
,	O
bigram	O
,	O
and	O
trigram	O
found	O
in	O
the	O
text	O
(	O
though	O
other	O
text	O
units	O
are	O
also	O
possible	O
,	O
as	O
discussed	O
below	O
)	O
.	O
</s>
<s>
Some	O
classifiers	O
make	O
a	O
binary	B-General_Concept
classification	I-General_Concept
for	O
a	O
test	O
example	O
,	O
while	O
others	O
assign	O
a	O
probability	O
of	O
being	O
a	O
keyphrase	O
.	O
</s>
<s>
We	O
can	O
determine	O
the	O
keyphrases	O
by	O
looking	O
at	O
binary	B-General_Concept
classification	I-General_Concept
decisions	O
or	O
probabilities	O
returned	O
from	O
our	O
learned	O
model	O
.	O
</s>
<s>
Designing	O
a	O
supervised	O
keyphrase	O
extraction	B-General_Concept
system	O
involves	O
deciding	O
on	O
several	O
choices	O
(	O
some	O
of	O
these	O
apply	O
to	O
unsupervised	O
,	O
too	O
)	O
.	O
</s>
<s>
Turney	O
and	O
others	O
have	O
used	O
all	O
possible	O
unigrams	B-Language
,	O
bigrams	O
,	O
and	O
trigrams	O
without	O
intervening	O
punctuation	O
and	O
after	O
removing	O
stopwords	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
we	O
use	O
only	O
unigrams	B-Language
,	O
bigrams	O
,	O
and	O
trigrams	O
,	O
then	O
we	O
will	O
never	O
be	O
able	O
to	O
extract	O
a	O
known	O
keyphrase	O
containing	O
four	O
words	O
.	O
</s>
<s>
Hulth	O
uses	O
a	O
reduced	O
set	O
of	O
features	O
,	O
which	O
were	O
found	O
most	O
successful	O
in	O
the	O
KEA	O
(	O
Keyphrase	O
Extraction	B-General_Concept
Algorithm	O
)	O
work	O
derived	O
from	O
Turney	O
's	O
seminal	O
paper	O
.	O
</s>
<s>
Hulth	O
used	O
a	O
single	O
binary	B-General_Concept
classifier	I-General_Concept
so	O
the	O
learning	O
algorithm	O
implicitly	O
determines	O
the	O
appropriate	O
number	O
.	O
</s>
<s>
Virtually	O
any	O
supervised	B-General_Concept
learning	I-General_Concept
algorithm	O
could	O
be	O
used	O
,	O
such	O
as	O
decision	O
trees	O
,	O
Naive	B-General_Concept
Bayes	I-General_Concept
,	O
and	O
rule	O
induction	O
.	O
</s>
<s>
In	O
the	O
case	O
of	O
Turney	O
's	O
GenEx	O
algorithm	O
,	O
a	O
genetic	B-Algorithm
algorithm	I-Algorithm
is	O
used	O
to	O
learn	O
parameters	O
for	O
a	O
domain-specific	O
keyphrase	O
extraction	B-General_Concept
algorithm	O
.	O
</s>
<s>
The	O
genetic	B-Algorithm
algorithm	I-Algorithm
optimizes	O
parameters	O
for	O
these	O
heuristics	O
with	O
respect	O
to	O
performance	O
on	O
training	O
documents	O
with	O
known	O
key	O
phrases	O
.	O
</s>
<s>
Another	O
keyphrase	O
extraction	B-General_Concept
algorithm	O
is	O
TextRank	O
.	O
</s>
<s>
Furthermore	O
,	O
training	O
on	O
a	O
specific	O
domain	O
tends	O
to	O
customize	O
the	O
extraction	B-General_Concept
process	O
to	O
that	O
domain	O
,	O
so	O
the	O
resulting	O
classifier	O
is	O
not	O
necessarily	O
portable	O
,	O
as	O
some	O
of	O
Turney	O
's	O
results	O
demonstrate	O
.	O
</s>
<s>
Unsupervised	O
keyphrase	O
extraction	B-General_Concept
removes	O
the	O
need	O
for	O
training	O
data	O
.	O
</s>
<s>
Instead	O
of	O
trying	O
to	O
learn	O
explicit	O
features	O
that	O
characterize	O
keyphrases	O
,	O
the	O
TextRank	O
algorithm	O
exploits	O
the	O
structure	O
of	O
the	O
text	O
itself	O
to	O
determine	O
keyphrases	O
that	O
appear	O
"	O
central	O
"	O
to	O
the	O
text	O
in	O
the	O
same	O
way	O
that	O
PageRank	B-Algorithm
selects	O
important	O
Web	O
pages	O
.	O
</s>
<s>
TextRank	O
is	O
a	O
general	O
purpose	O
graph-based	O
ranking	O
algorithm	O
for	O
NLP	B-Language
.	O
</s>
<s>
Essentially	O
,	O
it	O
runs	O
PageRank	B-Algorithm
on	O
a	O
graph	B-Application
specially	O
designed	O
for	O
a	O
particular	O
NLP	B-Language
task	O
.	O
</s>
<s>
For	O
keyphrase	O
extraction	B-General_Concept
,	O
it	O
builds	O
a	O
graph	B-Application
using	O
some	O
set	O
of	O
text	O
units	O
as	O
vertices	O
.	O
</s>
<s>
Unlike	O
PageRank	B-Algorithm
,	O
the	O
edges	O
are	O
typically	O
undirected	O
and	O
can	O
be	O
weighted	O
to	O
reflect	O
a	O
degree	O
of	O
similarity	O
.	O
</s>
<s>
Once	O
the	O
graph	B-Application
is	O
constructed	O
,	O
it	O
is	O
used	O
to	O
form	O
a	O
stochastic	O
matrix	O
,	O
combined	O
with	O
a	O
damping	O
factor	O
(	O
as	O
in	O
the	O
"	O
random	O
surfer	O
model	O
"	O
)	O
,	O
and	O
the	O
ranking	O
over	O
vertices	O
is	O
obtained	O
by	O
finding	O
the	O
eigenvector	O
corresponding	O
to	O
eigenvalue	O
1	O
(	O
i.e.	O
,	O
the	O
stationary	O
distribution	O
of	O
the	O
random	O
walk	O
on	O
the	O
graph	B-Application
)	O
.	O
</s>
<s>
Potentially	O
,	O
we	O
could	O
do	O
something	O
similar	O
to	O
the	O
supervised	O
methods	O
and	O
create	O
a	O
vertex	O
for	O
each	O
unigram	B-Language
,	O
bigram	O
,	O
trigram	O
,	O
etc	O
.	O
</s>
<s>
However	O
,	O
to	O
keep	O
the	O
graph	B-Application
small	O
,	O
the	O
authors	O
decide	O
to	O
rank	O
individual	O
unigrams	B-Language
in	O
a	O
first	O
step	O
,	O
and	O
then	O
include	O
a	O
second	O
step	O
that	O
merges	O
highly	O
ranked	O
adjacent	O
unigrams	B-Language
to	O
form	O
multi-word	O
phrases	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
we	O
rank	O
unigrams	B-Language
and	O
find	O
that	O
"	O
advanced	O
"	O
,	O
"	O
natural	O
"	O
,	O
"	O
language	O
"	O
,	O
and	O
"	O
processing	O
"	O
all	O
get	O
high	O
ranks	O
,	O
then	O
we	O
would	O
look	O
at	O
the	O
original	O
text	O
and	O
see	O
that	O
these	O
words	O
appear	O
consecutively	O
and	O
create	O
a	O
final	O
keyphrase	O
using	O
all	O
four	O
together	O
.	O
</s>
<s>
Note	O
that	O
the	O
unigrams	B-Language
placed	O
in	O
the	O
graph	B-Application
can	O
be	O
filtered	O
by	O
part	O
of	O
speech	O
.	O
</s>
<s>
Two	O
vertices	O
are	O
connected	O
by	O
an	O
edge	O
if	O
the	O
unigrams	B-Language
appear	O
within	O
a	O
window	O
of	O
size	O
N	O
in	O
the	O
original	O
text	O
.	O
</s>
<s>
Thus	O
,	O
"	O
natural	O
"	O
and	O
"	O
language	O
"	O
might	O
be	O
linked	O
in	O
a	O
text	O
about	O
NLP	B-Language
.	O
</s>
<s>
The	O
technique	O
chosen	O
is	O
to	O
set	O
a	O
count	O
T	O
to	O
be	O
a	O
user-specified	O
fraction	O
of	O
the	O
total	O
number	O
of	O
vertices	O
in	O
the	O
graph	B-Application
.	O
</s>
<s>
Then	O
the	O
top	O
T	O
vertices/unigrams	O
are	O
selected	O
based	O
on	O
their	O
stationary	O
probabilities	O
.	O
</s>
<s>
A	O
post	O
-	O
processing	O
step	O
is	O
then	O
applied	O
to	O
merge	O
adjacent	O
instances	O
of	O
these	O
T	O
unigrams	B-Language
.	O
</s>
<s>
It	O
is	O
not	O
initially	O
clear	O
why	O
applying	O
PageRank	B-Algorithm
to	O
a	O
co-occurrence	O
graph	B-Application
would	O
produce	O
useful	O
keyphrases	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
a	O
text	O
about	O
machine	O
learning	O
,	O
the	O
unigram	B-Language
"	O
learning	O
"	O
might	O
co-occur	O
with	O
"	O
machine	O
"	O
,	O
"	O
supervised	O
"	O
,	O
"	O
un-supervised	O
"	O
,	O
and	O
"	O
semi-supervised	O
"	O
in	O
four	O
different	O
sentences	O
.	O
</s>
<s>
Running	O
PageRank/TextRank	O
on	O
the	O
graph	B-Application
is	O
likely	O
to	O
rank	O
"	O
learning	O
"	O
highly	O
.	O
</s>
<s>
Similarly	O
,	O
if	O
the	O
text	O
contains	O
the	O
phrase	O
"	O
supervised	B-General_Concept
classification	I-General_Concept
"	O
,	O
then	O
there	O
would	O
be	O
an	O
edge	O
between	O
"	O
supervised	O
"	O
and	O
"	O
classification	O
"	O
.	O
</s>
<s>
If	O
it	O
ends	O
up	O
with	O
a	O
high	O
rank	O
,	O
it	O
will	O
be	O
selected	O
as	O
one	O
of	O
the	O
top	O
T	O
unigrams	B-Language
,	O
along	O
with	O
"	O
learning	O
"	O
and	O
probably	O
"	O
classification	O
"	O
.	O
</s>
<s>
In	O
the	O
final	O
post-processing	O
step	O
,	O
we	O
would	O
then	O
end	O
up	O
with	O
keyphrases	O
"	O
supervised	B-General_Concept
learning	I-General_Concept
"	O
and	O
"	O
supervised	B-General_Concept
classification	I-General_Concept
"	O
.	O
</s>
<s>
In	O
short	O
,	O
the	O
co-occurrence	O
graph	B-Application
will	O
contain	O
densely	O
connected	O
regions	O
for	O
terms	O
that	O
appear	O
often	O
and	O
in	O
different	O
contexts	O
.	O
</s>
<s>
A	O
random	O
walk	O
on	O
this	O
graph	B-Application
will	O
have	O
a	O
stationary	O
distribution	O
that	O
assigns	O
large	O
probabilities	O
to	O
the	O
terms	O
in	O
the	O
centers	O
of	O
the	O
clusters	O
.	O
</s>
<s>
This	O
is	O
similar	O
to	O
densely	O
connected	O
Web	O
pages	O
getting	O
ranked	O
highly	O
by	O
PageRank	B-Algorithm
.	O
</s>
<s>
This	O
approach	O
has	O
also	O
been	O
used	O
in	O
document	B-Application
summarization	I-Application
,	O
considered	O
below	O
.	O
</s>
<s>
Like	O
keyphrase	O
extraction	B-General_Concept
,	O
document	B-Application
summarization	I-Application
aims	O
to	O
identify	O
the	O
essence	O
of	O
a	O
text	O
.	O
</s>
<s>
Supervised	O
text	B-Application
summarization	I-Application
is	O
very	O
much	O
like	O
supervised	O
keyphrase	O
extraction	B-General_Concept
.	O
</s>
<s>
Note	O
,	O
however	O
,	O
that	O
these	O
natural	O
summaries	O
can	O
still	O
be	O
used	O
for	O
evaluation	O
purposes	O
,	O
since	O
ROUGE-1	O
evaluation	O
only	O
considers	O
unigrams	B-Language
.	O
</s>
<s>
During	O
the	O
DUC	O
2001	O
and	O
2002	O
evaluation	O
workshops	O
,	O
TNO	O
developed	O
a	O
sentence	B-General_Concept
extraction	I-General_Concept
system	O
for	O
multi-document	B-General_Concept
summarization	I-General_Concept
in	O
the	O
news	O
domain	O
.	O
</s>
<s>
The	O
system	O
was	O
based	O
on	O
a	O
hybrid	O
system	O
using	O
a	O
Naive	B-General_Concept
Bayes	I-General_Concept
classifier	I-General_Concept
and	O
statistical	O
language	O
models	O
for	O
modeling	O
salience	O
.	O
</s>
<s>
The	O
unsupervised	O
approach	O
to	O
summarization	O
is	O
also	O
quite	O
similar	O
in	O
spirit	O
to	O
unsupervised	O
keyphrase	O
extraction	B-General_Concept
and	O
gets	O
around	O
the	O
issue	O
of	O
costly	O
training	O
data	O
.	O
</s>
<s>
LexRank	O
is	O
an	O
algorithm	O
essentially	O
identical	O
to	O
TextRank	O
,	O
and	O
both	O
use	O
this	O
approach	O
for	O
document	B-Application
summarization	I-Application
.	O
</s>
<s>
The	O
two	O
methods	O
were	O
developed	O
by	O
different	O
groups	O
at	O
the	O
same	O
time	O
,	O
and	O
LexRank	O
simply	O
focused	O
on	O
summarization	O
,	O
but	O
could	O
just	O
as	O
easily	O
be	O
used	O
for	O
keyphrase	O
extraction	B-General_Concept
or	O
any	O
other	O
NLP	B-Language
ranking	O
task	O
.	O
</s>
<s>
In	O
both	O
LexRank	O
and	O
TextRank	O
,	O
a	O
graph	B-Application
is	O
constructed	O
by	O
creating	O
a	O
vertex	O
for	O
each	O
sentence	O
in	O
the	O
document	O
.	O
</s>
<s>
While	O
LexRank	O
uses	O
cosine	O
similarity	O
of	O
TF-IDF	O
vectors	O
,	O
TextRank	O
uses	O
a	O
very	O
similar	O
measure	O
based	O
on	O
the	O
number	O
of	O
words	O
two	O
sentences	O
have	O
in	O
common	O
(	O
normalized	B-General_Concept
by	O
the	O
sentences	O
 '	O
lengths	O
)	O
.	O
</s>
<s>
In	O
both	O
algorithms	O
,	O
the	O
sentences	O
are	O
ranked	O
by	O
applying	O
PageRank	B-Algorithm
to	O
the	O
resulting	O
graph	B-Application
.	O
</s>
<s>
Unlike	O
TextRank	O
,	O
LexRank	O
has	O
been	O
applied	O
to	O
multi-document	B-General_Concept
summarization	I-General_Concept
.	O
</s>
<s>
Multi-document	B-General_Concept
summarization	I-General_Concept
is	O
an	O
automatic	O
procedure	O
aimed	O
at	O
extraction	B-General_Concept
of	O
information	O
from	O
multiple	O
texts	O
written	O
about	O
the	O
same	O
topic	O
.	O
</s>
<s>
Resulting	O
summary	O
report	O
allows	O
individual	O
users	O
,	O
such	O
as	O
professional	O
information	O
consumers	O
,	O
to	O
quickly	O
familiarize	O
themselves	O
with	O
information	O
contained	O
in	O
a	O
large	O
cluster	B-Algorithm
of	O
documents	O
.	O
</s>
<s>
In	O
such	O
a	O
way	O
,	O
multi-document	B-General_Concept
summarization	I-General_Concept
systems	O
are	O
complementing	O
the	O
news	B-Application
aggregators	I-Application
performing	O
the	O
next	O
step	O
down	O
the	O
road	O
of	O
coping	O
with	O
information	O
overload	O
.	O
</s>
<s>
Multi-document	B-General_Concept
summarization	I-General_Concept
may	O
also	O
be	O
done	O
in	O
response	O
to	O
a	O
question	O
.	O
</s>
<s>
Multi-document	B-General_Concept
summarization	I-General_Concept
creates	O
information	O
reports	O
that	O
are	O
both	O
concise	O
and	O
comprehensive	O
.	O
</s>
<s>
A	O
related	O
method	O
is	O
Maximal	O
Marginal	O
Relevance	O
(	O
MMR	O
)	O
,	O
which	O
uses	O
a	O
general-purpose	O
graph-based	O
ranking	O
algorithm	O
like	O
Page/Lex/TextRank	O
that	O
handles	O
both	O
"	O
centrality	O
"	O
and	O
"	O
diversity	O
"	O
in	O
a	O
unified	O
mathematical	O
framework	O
based	O
on	O
absorbing	O
Markov	O
chain	O
random	O
walks	O
(	O
a	O
random	O
walk	O
where	O
certain	O
states	O
end	O
the	O
walk	O
)	O
.	O
</s>
<s>
The	O
state	O
of	O
the	O
art	O
results	O
for	O
multi-document	B-General_Concept
summarization	I-General_Concept
are	O
obtained	O
using	O
mixtures	O
of	O
submodular	B-Algorithm
functions	I-Algorithm
.	O
</s>
<s>
These	O
methods	O
have	O
achieved	O
the	O
state	O
of	O
the	O
art	O
results	O
for	O
Document	B-Application
Summarization	I-Application
Corpora	O
,	O
DUC	O
04	O
-	O
07	O
.	O
</s>
<s>
Similar	O
results	O
were	O
achieved	O
with	O
the	O
use	O
of	O
determinantal	O
point	O
processes	O
(	O
which	O
are	O
a	O
special	O
case	O
of	O
submodular	B-Algorithm
functions	I-Algorithm
)	O
for	O
DUC-04	O
.	O
</s>
<s>
A	O
new	O
method	O
for	O
multi-lingual	O
multi-document	B-General_Concept
summarization	I-General_Concept
that	O
avoids	O
redundancy	O
generates	O
ideograms	O
to	O
represent	O
the	O
meaning	O
of	O
each	O
sentence	O
in	O
each	O
document	O
,	O
then	O
evaluates	O
similarity	O
by	O
comparing	O
ideogram	O
shape	O
and	O
position	O
.	O
</s>
<s>
The	O
idea	O
of	O
a	O
submodular	B-Algorithm
set	I-Algorithm
function	I-Algorithm
has	O
recently	O
emerged	O
as	O
a	O
powerful	O
modeling	O
tool	O
for	O
various	O
summarization	O
problems	O
.	O
</s>
<s>
Submodular	B-Algorithm
functions	I-Algorithm
naturally	O
model	O
notions	O
of	O
coverage	O
,	O
information	O
,	O
representation	O
and	O
diversity	O
.	O
</s>
<s>
Moreover	O
,	O
several	O
important	O
combinatorial	O
optimization	O
problems	O
occur	O
as	O
special	O
instances	O
of	O
submodular	B-Algorithm
optimization	O
.	O
</s>
<s>
For	O
example	O
,	O
the	O
set	B-Algorithm
cover	I-Algorithm
problem	I-Algorithm
is	O
a	O
special	O
case	O
of	O
submodular	B-Algorithm
optimization	O
,	O
since	O
the	O
set	B-Algorithm
cover	I-Algorithm
function	O
is	O
submodular	B-Algorithm
.	O
</s>
<s>
The	O
set	B-Algorithm
cover	I-Algorithm
function	O
attempts	O
to	O
find	O
a	O
subset	O
of	O
objects	O
which	O
cover	O
a	O
given	O
set	O
of	O
concepts	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
document	B-Application
summarization	I-Application
,	O
one	O
would	O
like	O
the	O
summary	O
to	O
cover	O
all	O
important	O
and	O
relevant	O
concepts	O
in	O
the	O
document	O
.	O
</s>
<s>
This	O
is	O
an	O
instance	O
of	O
set	B-Algorithm
cover	I-Algorithm
.	O
</s>
<s>
Similarly	O
,	O
the	O
facility	O
location	O
problem	O
is	O
a	O
special	O
case	O
of	O
submodular	B-Algorithm
functions	I-Algorithm
.	O
</s>
<s>
Another	O
example	O
of	O
a	O
submodular	B-Algorithm
optimization	O
problem	O
is	O
using	O
a	O
determinantal	O
point	O
process	O
to	O
model	O
diversity	O
.	O
</s>
<s>
Similarly	O
,	O
the	O
Maximum-Marginal-Relevance	O
procedure	O
can	O
also	O
be	O
seen	O
as	O
an	O
instance	O
of	O
submodular	B-Algorithm
optimization	O
.	O
</s>
<s>
All	O
these	O
important	O
models	O
encouraging	O
coverage	O
,	O
diversity	O
and	O
information	O
are	O
all	O
submodular	B-Algorithm
.	O
</s>
<s>
Moreover	O
,	O
submodular	B-Algorithm
functions	I-Algorithm
can	O
be	O
efficiently	O
combined	O
,	O
and	O
the	O
resulting	O
function	O
is	O
still	O
submodular	B-Algorithm
.	O
</s>
<s>
Hence	O
,	O
one	O
could	O
combine	O
one	O
submodular	B-Algorithm
function	I-Algorithm
which	O
models	O
diversity	O
,	O
another	O
one	O
which	O
models	O
coverage	O
and	O
use	O
human	O
supervision	O
to	O
learn	O
a	O
right	O
model	O
of	O
a	O
submodular	B-Algorithm
function	I-Algorithm
for	O
the	O
problem	O
.	O
</s>
<s>
While	O
submodular	B-Algorithm
functions	I-Algorithm
are	O
fitting	O
problems	O
for	O
summarization	O
,	O
they	O
also	O
admit	O
very	O
efficient	O
algorithms	O
for	O
optimization	O
.	O
</s>
<s>
For	O
example	O
,	O
a	O
simple	O
greedy	B-Algorithm
algorithm	I-Algorithm
admits	O
a	O
constant	O
factor	O
guarantee	O
.	O
</s>
<s>
Moreover	O
,	O
the	O
greedy	B-Algorithm
algorithm	I-Algorithm
is	O
extremely	O
simple	O
to	O
implement	O
and	O
can	O
scale	O
to	O
large	O
datasets	O
,	O
which	O
is	O
very	O
important	O
for	O
summarization	O
problems	O
.	O
</s>
<s>
Submodular	B-Algorithm
functions	I-Algorithm
have	O
achieved	O
state-of-the-art	O
for	O
almost	O
all	O
summarization	O
problems	O
.	O
</s>
<s>
For	O
example	O
,	O
work	O
by	O
Lin	O
and	O
Bilmes	O
,	O
2012	O
shows	O
that	O
submodular	B-Algorithm
functions	I-Algorithm
achieve	O
the	O
best	O
results	O
to	O
date	O
on	O
DUC-04	O
,	O
DUC-05	O
,	O
DUC-06	O
and	O
DUC-07	O
systems	O
for	O
document	B-Application
summarization	I-Application
.	O
</s>
<s>
Similarly	O
,	O
work	O
by	O
Lin	O
and	O
Bilmes	O
,	O
2011	O
,	O
shows	O
that	O
many	O
existing	O
systems	O
for	O
automatic	B-Application
summarization	I-Application
are	O
instances	O
of	O
submodular	B-Algorithm
functions	I-Algorithm
.	O
</s>
<s>
This	O
was	O
a	O
breakthrough	O
result	O
establishing	O
submodular	B-Algorithm
functions	I-Algorithm
as	O
the	O
right	O
models	O
for	O
summarization	O
problems	O
.	O
</s>
<s>
Submodular	B-Algorithm
Functions	I-Algorithm
have	O
also	O
been	O
used	O
for	O
other	O
summarization	O
tasks	O
.	O
</s>
<s>
Tschiatschek	O
et	O
al.	O
,	O
2014	O
show	O
that	O
mixtures	O
of	O
submodular	B-Algorithm
functions	I-Algorithm
achieve	O
state-of-the-art	O
results	O
for	O
image	O
collection	O
summarization	O
.	O
</s>
<s>
Similarly	O
,	O
Bairi	O
et	O
al.	O
,	O
2015	O
show	O
the	O
utility	O
of	O
submodular	B-Algorithm
functions	I-Algorithm
for	O
summarizing	O
multi-document	O
topic	O
hierarchies	O
.	O
</s>
<s>
Submodular	B-Algorithm
Functions	I-Algorithm
have	O
also	O
successfully	O
been	O
used	O
for	O
summarizing	O
machine	O
learning	O
datasets	O
.	O
</s>
<s>
Specific	O
applications	O
of	O
automatic	B-Application
summarization	I-Application
include	O
:	O
</s>
<s>
The	O
Reddit	B-Application
bot	B-Protocol
"	O
autotldr	O
"	O
,	O
created	O
in	O
2011	O
summarizes	O
news	O
articles	O
in	O
the	O
comment-section	O
of	O
reddit	B-Application
posts	O
.	O
</s>
<s>
It	O
was	O
found	O
to	O
be	O
very	O
useful	O
by	O
the	O
reddit	B-Application
community	O
which	O
upvoted	O
its	O
summaries	O
hundreds	O
of	O
thousands	O
of	O
times	O
.	O
</s>
<s>
The	O
most	O
common	O
way	O
to	O
evaluate	O
summaries	O
is	O
ROUGE	B-General_Concept
(	O
Recall-Oriented	O
Understudy	O
for	O
Gisting	O
Evaluation	O
)	O
.	O
</s>
<s>
ROUGE	B-General_Concept
is	O
a	O
recall-based	O
measure	O
of	O
how	O
well	O
a	O
summary	O
covers	O
the	O
content	O
of	O
human-generated	O
summaries	O
known	O
as	O
references	O
.	O
</s>
<s>
It	O
calculates	O
n-gram	B-Language
overlaps	O
between	O
automatically	O
generated	O
summaries	O
and	O
previously	O
written	O
human	O
summaries	O
.	O
</s>
<s>
Recall	O
can	O
be	O
computed	O
with	O
respect	O
to	O
unigram	B-Language
,	O
bigram	O
,	O
trigram	O
,	O
or	O
4-gram	B-Language
matching	O
.	O
</s>
<s>
For	O
example	O
,	O
ROUGE-1	O
is	O
the	O
fraction	O
of	O
unigrams	B-Language
that	O
appear	O
in	O
both	O
the	O
reference	O
summary	O
and	O
the	O
automatic	O
summary	O
out	O
of	O
all	O
unigrams	B-Language
in	O
the	O
reference	O
summary	O
.	O
</s>
<s>
ROUGE	B-General_Concept
cannot	O
determine	O
if	O
the	O
result	O
is	O
coherent	O
,	O
that	O
is	O
if	O
sentences	O
flow	O
together	O
in	O
a	O
sensibly	O
.	O
</s>
<s>
High-order	O
n-gram	B-Language
ROUGE	B-General_Concept
measures	O
help	O
to	O
some	O
degree	O
.	O
</s>
<s>
Similarly	O
,	O
for	O
image	B-Application
summarization	I-Application
,	O
Tschiatschek	O
et	O
al.	O
,	O
developed	O
a	O
Visual-ROUGE	O
score	O
which	O
judges	O
the	O
performance	O
of	O
algorithms	O
for	O
image	B-Application
summarization	I-Application
.	O
</s>
<s>
Pattern-based	O
summarization	O
was	O
the	O
most	O
powerful	O
option	O
for	O
multi-document	B-General_Concept
summarization	I-General_Concept
found	O
by	O
2016	O
.	O
</s>
<s>
Recently	O
the	O
rise	O
of	O
transformer	B-Algorithm
models	I-Algorithm
replacing	O
more	O
traditional	O
RNN	B-Algorithm
(	O
LSTM	B-Algorithm
)	O
have	O
provided	O
a	O
flexibility	O
in	O
the	O
mapping	O
of	O
text	O
sequences	O
to	O
text	O
sequences	O
of	O
a	O
different	O
type	O
,	O
which	O
is	O
well	O
suited	O
to	O
automatic	B-Application
summarization	I-Application
.	O
</s>
