<s>
Question	B-Algorithm
answering	I-Algorithm
(	O
QA	O
)	O
is	O
a	O
computer	O
science	O
discipline	O
within	O
the	O
fields	O
of	O
information	B-Library
retrieval	I-Library
and	O
natural	B-Language
language	I-Language
processing	I-Language
(	O
NLP	B-Language
)	O
that	O
is	O
concerned	O
with	O
building	O
systems	O
that	O
automatically	O
answer	O
questions	O
that	O
are	O
posed	O
by	O
humans	O
in	O
a	O
natural	O
language	O
.	O
</s>
<s>
Some	O
examples	O
of	O
natural	O
language	O
document	O
collections	O
used	O
for	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
include	O
:	O
</s>
<s>
This	O
is	O
a	O
simpler	O
form	O
of	O
question	B-Algorithm
answering	I-Algorithm
,	O
since	O
a	O
given	O
article	O
is	O
relatively	O
short	O
compared	O
to	O
other	O
types	O
of	O
question	B-Algorithm
answering	I-Algorithm
problems	O
that	O
deal	O
with	O
a	O
larger	O
domain	O
of	O
information	O
.	O
</s>
<s>
Closed-book	O
question	B-Algorithm
answering	I-Algorithm
:	O
a	O
system	O
has	O
memorized	O
some	O
facts	O
during	O
training	O
and	O
can	O
answer	O
questions	O
without	O
explicitly	O
being	O
given	O
a	O
context	O
.	O
</s>
<s>
Closed-domain	O
question	B-Algorithm
answering	I-Algorithm
deals	O
with	O
questions	O
under	O
a	O
specific	O
domain	O
(	O
for	O
example	O
,	O
medicine	O
or	O
automotive	O
maintenance	O
)	O
and	O
can	O
exploit	O
domain-specific	O
knowledge	O
frequently	O
formalized	O
in	O
ontologies	O
.	O
</s>
<s>
Question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
machine	B-General_Concept
reading	I-General_Concept
applications	O
have	O
also	O
been	O
constructed	O
in	O
the	O
medical	O
domain	O
,	O
for	O
instance	O
Alzheimer	O
's	O
disease	O
.	O
</s>
<s>
Systems	O
designed	O
for	O
open-domain	B-Algorithm
question	I-Algorithm
answering	I-Algorithm
usually	O
have	O
much	O
more	O
data	O
available	O
from	O
which	O
to	O
extract	O
the	O
answer	O
.	O
</s>
<s>
hybrid	B-General_Concept
systems	I-General_Concept
.	O
</s>
<s>
Hybrid	B-General_Concept
systems	I-General_Concept
use	O
a	O
combination	O
of	O
rule-based	O
and	O
statistical	O
methods	O
.	O
</s>
<s>
Two	O
early	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
were	O
BASEBALL	O
and	O
LUNAR	O
.	O
</s>
<s>
Both	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
were	O
very	O
effective	O
in	O
their	O
chosen	O
domains	O
.	O
</s>
<s>
Further	O
restricted-domain	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
were	O
developed	O
in	O
the	O
following	O
years	O
.	O
</s>
<s>
The	O
language	O
abilities	O
of	O
BASEBALL	O
and	O
LUNAR	O
used	O
techniques	O
similar	O
to	O
ELIZA	B-Application
and	O
DOCTOR	O
,	O
the	O
first	O
chatterbot	B-Protocol
programs	O
.	O
</s>
<s>
The	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
developed	O
to	O
interface	O
with	O
these	O
expert	B-Application
systems	I-Application
produced	O
and	O
valid	O
responses	O
to	O
questions	O
within	O
an	O
area	O
of	O
knowledge	O
.	O
</s>
<s>
These	O
expert	B-Application
systems	I-Application
closely	O
resembled	O
modern	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
except	O
in	O
their	O
internal	O
architecture	O
.	O
</s>
<s>
Expert	B-Application
systems	I-Application
rely	O
heavily	O
on	O
expert-constructed	O
and	O
organized	O
knowledge	O
bases	O
,	O
whereas	O
many	O
modern	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
rely	O
on	O
statistical	O
processing	O
of	O
a	O
large	O
,	O
unstructured	O
,	O
natural	O
language	O
text	O
corpus	O
.	O
</s>
<s>
The	O
1970s	O
and	O
1980s	O
saw	O
the	O
development	O
of	O
comprehensive	O
theories	O
in	O
computational	O
linguistics	O
,	O
which	O
led	O
to	O
the	O
development	O
of	O
ambitious	O
projects	O
in	O
text	O
comprehension	O
and	O
question	B-Algorithm
answering	I-Algorithm
.	O
</s>
<s>
One	O
example	O
was	O
the	O
Unix	B-Application
Consultant	O
(	O
UC	O
)	O
,	O
developed	O
by	O
Robert	O
Wilensky	O
at	O
U.C.	O
</s>
<s>
The	O
system	O
answered	O
questions	O
pertaining	O
to	O
the	O
Unix	B-Application
operating	I-Application
system	I-Application
.	O
</s>
<s>
Another	O
project	O
was	O
LILOG	O
,	O
a	O
text-understanding	B-General_Concept
system	O
that	O
operated	O
on	O
the	O
domain	O
of	O
tourism	O
information	O
in	O
a	O
German	O
city	O
.	O
</s>
<s>
Specialized	O
natural-language	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
have	O
been	O
developed	O
,	O
such	O
as	O
EAGLi	O
for	O
health	O
and	O
life	O
scientists	O
.	O
</s>
<s>
For	O
example	O
,	O
modern	O
open-domain	B-Algorithm
question	I-Algorithm
answering	I-Algorithm
systems	O
may	O
use	O
a	O
retriever-reader	O
architecture	O
.	O
</s>
<s>
Systems	O
such	O
as	O
GPT-3	B-General_Concept
,	O
T5	O
,	O
and	O
BART	O
use	O
an	O
end-to-end	O
architecture	O
in	O
which	O
a	O
transformer-based	O
architecture	O
stores	O
large-scale	O
textual	O
data	O
in	O
the	O
underlying	O
parameters	O
.	O
</s>
<s>
Question	B-Algorithm
answering	I-Algorithm
is	O
dependent	O
on	O
a	O
good	O
search	O
corpus	O
;	O
without	O
documents	O
containing	O
the	O
answer	O
,	O
there	O
is	O
little	O
any	O
question	B-Algorithm
answering	I-Algorithm
system	I-Algorithm
can	O
do	O
.	O
</s>
<s>
Larger	O
collections	O
generally	O
mean	O
better	O
question	B-Algorithm
answering	I-Algorithm
performance	O
,	O
unless	O
the	O
question	O
domain	O
is	O
orthogonal	O
to	O
the	O
collection	O
.	O
</s>
<s>
Data	B-Application
redundancy	I-Application
in	O
massive	O
collections	O
,	O
such	O
as	O
the	O
web	O
,	O
means	O
that	O
nuggets	O
of	O
information	O
are	O
likely	O
to	O
be	O
phrased	O
in	O
many	O
different	O
ways	O
in	O
differing	O
contexts	O
and	O
documents	O
,	O
leading	O
to	O
two	O
benefits	O
:	O
</s>
<s>
If	O
the	O
right	O
information	O
appears	O
in	O
many	O
forms	O
,	O
the	O
question	B-Algorithm
answering	I-Algorithm
system	I-Algorithm
needs	O
to	O
perform	O
fewer	O
complex	O
NLP	B-Language
techniques	O
to	O
understand	O
the	O
text	O
.	O
</s>
<s>
Some	O
question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
rely	O
heavily	O
on	O
automated	O
reasoning	O
.	O
</s>
<s>
In	O
information	B-Library
retrieval	I-Library
,	O
an	O
open-domain	B-Algorithm
question	I-Algorithm
answering	I-Algorithm
system	O
tries	O
to	O
return	O
an	O
answer	O
in	O
response	O
to	O
the	O
user	O
's	O
question	O
.	O
</s>
<s>
The	O
system	O
finds	O
answers	O
by	O
using	O
a	O
combination	O
of	O
techniques	O
from	O
computational	O
linguistics	O
,	O
information	B-Library
retrieval	I-Library
,	O
and	O
knowledge	O
representation	O
.	O
</s>
<s>
It	O
then	O
transforms	O
this	O
input	O
sentence	O
into	O
a	O
query	B-Library
in	O
its	O
logical	O
form	O
.	O
</s>
<s>
Assigning	O
a	O
question	O
type	O
to	O
the	O
question	O
is	O
a	O
crucial	O
task	O
;	O
the	O
entire	O
answer	O
extraction	B-General_Concept
process	O
relies	O
on	O
finding	O
the	O
correct	O
question	O
type	O
and	O
hence	O
the	O
correct	O
answer	O
type	O
.	O
</s>
<s>
Keyword	O
extraction	B-General_Concept
is	O
the	O
first	O
step	O
in	O
identifying	O
the	O
input	O
question	O
type	O
.	O
</s>
<s>
A	O
lexical	O
dictionary	O
such	O
as	O
WordNet	B-General_Concept
can	O
be	O
used	O
for	O
understanding	O
the	O
context	O
.	O
</s>
<s>
Once	O
the	O
system	O
identifies	O
the	O
question	O
type	O
,	O
it	O
uses	O
an	O
information	B-Library
retrieval	I-Library
system	I-Library
to	O
find	O
a	O
set	O
of	O
documents	O
that	O
contain	O
the	O
correct	O
keywords	O
.	O
</s>
<s>
For	O
questions	O
such	O
as	O
"	O
Who	O
"	O
or	O
"	O
Where	O
"	O
,	O
a	O
named-entity	B-General_Concept
recogniser	I-General_Concept
finds	O
relevant	O
"	O
Person	O
"	O
and	O
"	O
Location	O
"	O
names	O
from	O
the	O
retrieved	O
documents	O
.	O
</s>
<s>
An	O
open-source	B-License
,	O
math-aware	O
,	O
question	B-Algorithm
answering	I-Algorithm
system	I-Algorithm
called	O
MathQA	O
,	O
based	O
on	O
Ask	O
Platypus	O
and	O
Wikidata	O
,	O
was	O
published	O
in	O
2018	O
.	O
</s>
<s>
One	O
possible	O
approach	O
is	O
to	O
perform	O
supervised	O
annotation	O
via	O
Entity	B-General_Concept
Linking	I-General_Concept
.	O
</s>
<s>
The	O
lab	O
was	O
motivated	O
by	O
the	O
fact	O
that	O
20%	O
of	O
mathematical	O
queries	B-Library
in	O
general-purpose	O
search	O
engines	O
are	O
expressed	O
as	O
well-formed	O
questions	O
.	O
</s>
<s>
The	O
inverse	O
of	O
mathematical	O
question	B-Algorithm
answering	I-Algorithm
—	O
mathematical	O
question	O
generation	O
—	O
has	O
also	O
been	O
researched	O
.	O
</s>
<s>
Question	B-Algorithm
answering	I-Algorithm
systems	I-Algorithm
have	O
been	O
extended	O
in	O
recent	O
years	O
to	O
encompass	O
additional	O
domains	O
of	O
knowledge	O
For	O
example	O
,	O
systems	O
have	O
been	O
developed	O
to	O
automatically	O
answer	O
temporal	O
and	O
geospatial	O
questions	O
,	O
questions	O
of	O
definition	O
and	O
terminology	O
,	O
biographical	O
questions	O
,	O
multilingual	O
questions	O
,	O
and	O
questions	O
about	O
the	O
content	O
of	O
audio	O
,	O
images	O
,	O
and	O
video	O
.	O
</s>
<s>
Current	O
question	B-Algorithm
answering	I-Algorithm
research	O
topics	O
include	O
:	O
</s>
<s>
In	O
2011	O
,	O
Watson	B-Application
,	O
a	O
question	B-Algorithm
answering	I-Algorithm
computer	O
system	O
developed	O
by	O
IBM	O
,	O
competed	O
in	O
two	O
exhibition	O
matches	O
of	O
Jeopardy	O
!	O
</s>
<s>
Facebook	O
Research	O
made	O
their	O
DrQA	O
system	O
available	O
under	O
an	O
open	B-License
source	I-License
license	O
.	O
</s>
<s>
The	O
open	B-License
source	I-License
framework	O
Haystack	O
by	O
deepset	B-General_Concept
combines	O
open-domain	B-Algorithm
question	I-Algorithm
answering	I-Algorithm
with	O
generative	O
question	B-Algorithm
answering	I-Algorithm
and	O
supports	O
the	O
of	O
the	O
language	B-Language
models	I-Language
for	O
.	O
</s>
