<s>
Relevance	B-Application
feedback	I-Application
is	O
a	O
feature	O
of	O
some	O
information	B-Library
retrieval	I-Library
systems	I-Library
.	O
</s>
<s>
The	O
idea	O
behind	O
relevance	B-Application
feedback	I-Application
is	O
to	O
take	O
the	O
results	O
that	O
are	O
initially	O
returned	O
from	O
a	O
given	O
query	B-Library
,	O
to	O
gather	O
user	O
feedback	O
,	O
and	O
to	O
use	O
information	O
about	O
whether	O
or	O
not	O
those	O
results	O
are	O
relevant	O
to	O
perform	O
a	O
new	O
query	B-Library
.	O
</s>
<s>
Explicit	O
feedback	O
is	O
obtained	O
from	O
assessors	O
of	O
relevance	O
indicating	O
the	O
relevance	O
of	O
a	O
document	O
retrieved	O
for	O
a	O
query	B-Library
.	O
</s>
<s>
Binary	O
relevance	B-Application
feedback	I-Application
indicates	O
that	O
a	O
document	O
is	O
either	O
relevant	O
or	O
irrelevant	O
for	O
a	O
given	O
query	B-Library
.	O
</s>
<s>
Graded	O
relevance	B-Application
feedback	I-Application
indicates	O
the	O
relevance	O
of	O
a	O
document	O
to	O
a	O
query	B-Library
on	O
a	O
scale	O
using	O
numbers	O
,	O
letters	O
,	O
or	O
descriptions	O
(	O
such	O
as	O
"	O
not	O
relevant	O
"	O
,	O
"	O
somewhat	O
relevant	O
"	O
,	O
"	O
relevant	O
"	O
,	O
or	O
"	O
very	O
relevant	O
"	O
)	O
.	O
</s>
<s>
An	O
example	O
of	O
this	O
would	O
be	O
the	O
SearchWiki	B-Application
feature	O
implemented	O
by	O
Google	B-Application
on	O
their	O
search	O
website	O
.	O
</s>
<s>
The	O
relevance	B-Application
feedback	I-Application
information	O
needs	O
to	O
be	O
interpolated	O
with	O
the	O
original	O
query	B-Library
to	O
improve	O
retrieval	O
performance	O
,	O
such	O
as	O
the	O
well-known	O
Rocchio	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
<s>
A	O
performance	O
metric	O
which	O
became	O
popular	O
around	O
2005	O
to	O
measure	O
the	O
usefulness	O
of	O
a	O
ranking	O
algorithm	O
based	O
on	O
the	O
explicit	O
relevance	B-Application
feedback	I-Application
is	O
normalized	O
discounted	O
cumulative	O
gain	O
.	O
</s>
<s>
The	O
key	O
differences	O
of	O
implicit	O
relevance	B-Application
feedback	I-Application
from	O
that	O
of	O
explicit	O
include	O
:	O
</s>
<s>
An	O
example	O
of	O
this	O
is	O
dwell	B-Application
time	I-Application
,	O
which	O
is	O
a	O
measure	O
of	O
how	O
long	O
a	O
user	O
spends	O
viewing	O
the	O
page	O
linked	O
to	O
in	O
a	O
search	O
result	O
.	O
</s>
<s>
It	O
is	O
an	O
indicator	O
of	O
how	O
well	O
the	O
search	O
result	O
met	O
the	O
query	B-Library
intent	O
of	O
the	O
user	O
,	O
and	O
is	O
used	O
as	O
a	O
feedback	O
mechanism	O
to	O
improve	O
search	O
results	O
.	O
</s>
<s>
Pseudo	O
relevance	B-Application
feedback	I-Application
,	O
also	O
known	O
as	O
blind	O
relevance	B-Application
feedback	I-Application
,	O
provides	O
a	O
method	O
for	O
automatic	O
local	O
analysis	O
.	O
</s>
<s>
It	O
automates	O
the	O
manual	O
part	O
of	O
relevance	B-Application
feedback	I-Application
,	O
so	O
that	O
the	O
user	O
gets	O
improved	O
retrieval	O
performance	O
without	O
an	O
extended	O
interaction	O
.	O
</s>
<s>
The	O
method	O
is	O
to	O
do	O
normal	O
retrieval	O
to	O
find	O
an	O
initial	O
set	O
of	O
most	O
relevant	O
documents	O
,	O
to	O
then	O
assume	O
that	O
the	O
top	O
"	O
k	O
"	O
ranked	O
documents	O
are	O
relevant	O
,	O
and	O
finally	O
to	O
do	O
relevance	B-Application
feedback	I-Application
as	O
before	O
under	O
this	O
assumption	O
.	O
</s>
<s>
Take	O
the	O
results	O
returned	O
by	O
initial	O
query	B-Library
as	O
relevant	O
results	O
(	O
only	O
top	O
k	O
with	O
k	O
being	O
between	O
10	O
and	O
50	O
in	O
most	O
experiments	O
)	O
.	O
</s>
<s>
Do	O
Query	B-Library
Expansion	O
,	O
add	O
these	O
terms	O
to	O
query	B-Library
,	O
and	O
then	O
match	O
the	O
returned	O
documents	O
for	O
this	O
query	B-Library
and	O
finally	O
return	O
the	O
most	O
relevant	O
documents	O
.	O
</s>
<s>
Through	O
a	O
query	B-Library
expansion	O
,	O
some	O
relevant	O
documents	O
missed	O
in	O
the	O
initial	O
round	O
can	O
then	O
be	O
retrieved	O
to	O
improve	O
the	O
overall	O
performance	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
the	O
query	B-Library
is	O
about	O
copper	O
mines	O
and	O
the	O
top	O
several	O
documents	O
are	O
all	O
about	O
mines	O
in	O
Chile	O
,	O
then	O
there	O
may	O
be	O
query	B-Library
drift	O
in	O
the	O
direction	O
of	O
documents	O
on	O
Chile	O
.	O
</s>
<s>
In	O
addition	O
,	O
if	O
the	O
words	O
added	O
to	O
the	O
original	O
query	B-Library
are	O
unrelated	O
to	O
the	O
query	B-Library
topic	O
,	O
the	O
quality	O
of	O
the	O
retrieval	O
is	O
likely	O
to	O
be	O
degraded	O
,	O
especially	O
in	O
Web	O
search	O
,	O
where	O
web	O
documents	O
often	O
cover	O
multiple	O
different	O
topics	O
.	O
</s>
<s>
To	O
improve	O
the	O
quality	O
of	O
expansion	O
words	O
in	O
pseudo-relevance	O
feedback	O
,	O
a	O
positional	O
relevance	B-Application
feedback	I-Application
for	O
pseudo-relevance	O
feedback	O
has	O
been	O
proposed	O
to	O
select	O
from	O
feedback	O
documents	O
those	O
words	O
that	O
are	O
focused	O
on	O
the	O
query	B-Library
topic	O
based	O
on	O
positions	O
of	O
words	O
in	O
feedback	O
documents	O
.	O
</s>
<s>
Specifically	O
,	O
the	O
positional	O
relevance	O
model	O
assigns	O
more	O
weights	O
to	O
words	O
occurring	O
closer	O
to	O
query	B-Library
words	O
based	O
on	O
the	O
intuition	O
that	O
words	O
closer	O
to	O
query	B-Library
words	O
are	O
more	O
likely	O
to	O
be	O
related	O
to	O
the	O
query	B-Library
topic	O
.	O
</s>
<s>
Blind	O
feedback	O
automates	O
the	O
manual	O
part	O
of	O
relevance	B-Application
feedback	I-Application
and	O
has	O
the	O
advantage	O
that	O
assessors	O
are	O
not	O
required	O
.	O
</s>
<s>
Relevance	O
information	O
is	O
utilized	O
by	O
using	O
the	O
contents	O
of	O
the	O
relevant	O
documents	O
to	O
either	O
adjust	O
the	O
weights	O
of	O
terms	O
in	O
the	O
original	O
query	B-Library
,	O
or	O
by	O
using	O
those	O
contents	O
to	O
add	O
words	O
to	O
the	O
query	B-Library
.	O
</s>
<s>
Relevance	B-Application
feedback	I-Application
is	O
often	O
implemented	O
using	O
the	O
Rocchio	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
