<s>
Natural	B-General_Concept
language	I-General_Concept
generation	I-General_Concept
(	O
NLG	O
)	O
is	O
a	O
software	B-General_Concept
process	O
that	O
produces	O
natural	O
language	O
output	O
.	O
</s>
<s>
Common	O
applications	O
of	O
NLG	O
methods	O
include	O
the	O
production	O
of	O
various	O
reports	O
,	O
for	O
example	O
weather	O
and	O
patient	O
reports	O
;	O
image	O
captions	O
;	O
and	O
chatbots	B-Application
.	O
</s>
<s>
NLG	O
systems	O
can	O
also	O
be	O
compared	O
to	O
translators	B-Application
of	O
artificial	O
computer	O
languages	O
,	O
such	O
as	O
decompilers	B-Application
or	O
transpilers	B-Language
,	O
which	O
also	O
produce	O
human-readable	O
code	O
generated	O
from	O
an	O
intermediate	B-Application
representation	I-Application
.	O
</s>
<s>
NLG	O
may	O
be	O
viewed	O
as	O
complementary	O
to	O
natural-language	B-General_Concept
understanding	I-General_Concept
(	O
NLU	O
)	O
:	O
whereas	O
in	O
natural-language	B-General_Concept
understanding	I-General_Concept
,	O
the	O
system	O
needs	O
to	O
disambiguate	O
the	O
input	O
sentence	O
to	O
produce	O
the	O
machine	O
representation	O
language	O
,	O
in	O
NLG	O
the	O
system	O
needs	O
to	O
make	O
decisions	O
about	O
how	O
to	O
put	O
a	O
representation	O
into	O
words	O
.	O
</s>
<s>
NLG	O
has	O
existed	O
since	O
ELIZA	B-Application
was	O
developed	O
in	O
the	O
mid	O
1960s	O
,	O
but	O
the	O
methods	O
were	O
first	O
used	O
commercially	O
in	O
the	O
1990s	O
.	O
</s>
<s>
NLG	O
techniques	O
range	O
from	O
simple	O
template-based	O
systems	O
like	O
a	O
mail	B-Application
merge	I-Application
that	O
generates	O
form	O
letters	O
,	O
to	O
systems	O
that	O
have	O
a	O
complex	O
understanding	O
of	O
human	O
grammar	O
.	O
</s>
<s>
For	O
example	O
,	O
using	O
the	O
historical	O
data	O
for	O
July	O
1	O
,	O
2005	O
,	O
the	O
software	B-General_Concept
produces	O
:	O
</s>
<s>
The	O
typical	O
stages	O
of	O
natural-language	B-General_Concept
generation	I-General_Concept
,	O
as	O
proposed	O
by	O
Dale	O
and	O
Reiter	O
,	O
are	O
:	O
</s>
<s>
Content	B-General_Concept
determination	I-General_Concept
:	O
Deciding	O
what	O
information	O
to	O
mention	O
in	O
the	O
text	O
.	O
</s>
<s>
Document	B-General_Concept
structuring	I-General_Concept
:	O
Overall	O
organisation	O
of	O
the	O
information	O
to	O
convey	O
.	O
</s>
<s>
Aggregation	B-General_Concept
:	O
Merging	O
of	O
similar	O
sentences	O
to	O
improve	O
readability	O
and	O
naturalness	O
.	O
</s>
<s>
Lexical	B-General_Concept
choice	I-General_Concept
:	O
Putting	O
words	O
to	O
the	O
concepts	O
.	O
</s>
<s>
Referring	B-General_Concept
expression	I-General_Concept
generation	I-General_Concept
:	O
Creating	O
referring	O
expressions	O
that	O
identify	O
objects	O
and	O
regions	O
.	O
</s>
<s>
syntax	B-Application
,	O
morphology	O
,	O
and	O
orthography	O
.	O
</s>
<s>
In	O
other	O
words	O
,	O
we	O
build	O
an	O
NLG	O
system	O
by	O
training	O
a	O
machine	O
learning	O
algorithm	O
(	O
often	O
an	O
LSTM	B-Algorithm
)	O
on	O
a	O
large	O
data	O
set	O
of	O
input	O
data	O
and	O
corresponding	O
(	O
human-written	O
)	O
output	O
texts	O
.	O
</s>
<s>
The	O
end-to-end	O
approach	O
has	O
perhaps	O
been	O
most	O
successful	O
in	O
image	B-Application
captioning	I-Application
,	O
that	O
is	O
automatically	O
generating	O
a	O
textual	O
caption	O
for	O
an	O
image	O
.	O
</s>
<s>
systems	O
usually	O
perform	O
data	B-General_Concept
analysis	I-General_Concept
as	O
well	O
as	O
text	B-General_Concept
generation	I-General_Concept
.	O
</s>
<s>
NLG	O
is	O
also	O
being	O
used	O
commercially	O
in	O
automated	B-General_Concept
journalism	I-General_Concept
,	O
chatbots	B-Application
,	O
generating	O
product	O
descriptions	O
for	O
e-commerce	O
sites	O
,	O
summarising	O
medical	O
records	O
,	O
and	O
enhancing	O
accessibility	O
(	O
for	O
example	O
by	O
describing	O
graphs	O
and	O
data	O
sets	O
to	O
blind	O
people	O
)	O
.	O
</s>
<s>
An	O
example	O
of	O
an	O
interactive	O
use	O
of	O
NLG	O
is	O
the	O
WYSIWYM	B-General_Concept
framework	O
.	O
</s>
<s>
It	O
stands	O
for	O
What	B-General_Concept
you	I-General_Concept
see	I-General_Concept
is	I-General_Concept
what	I-General_Concept
you	I-General_Concept
meant	I-General_Concept
and	O
allows	O
users	O
to	O
see	O
and	O
manipulate	O
the	O
continuously	O
rendered	O
view	O
(	O
NLG	O
output	O
)	O
of	O
an	O
underlying	O
formal	O
language	O
document	O
(	O
NLG	O
input	O
)	O
,	O
thereby	O
editing	O
the	O
formal	O
language	O
without	O
learning	O
it	O
.	O
</s>
<s>
Over	O
the	O
past	O
few	O
years	O
,	O
there	O
has	O
been	O
an	O
increased	O
interest	O
in	O
automatically	B-Application
generating	I-Application
captions	I-Application
for	O
images	O
,	O
as	O
part	O
of	O
a	O
broader	O
endeavor	O
to	O
investigate	O
the	O
interface	O
between	O
vision	O
and	O
language	O
.	O
</s>
<s>
A	O
case	O
of	O
data-to-text	O
generation	O
,	O
the	O
algorithm	O
of	O
image	B-Application
captioning	I-Application
(	O
or	O
automatic	O
image	O
description	O
)	O
involves	O
taking	O
an	O
image	O
,	O
analyzing	O
its	O
visual	O
content	O
,	O
and	O
generating	O
a	O
textual	O
description	O
(	O
typically	O
a	O
sentence	O
)	O
that	O
verbalizes	O
the	O
most	O
prominent	O
aspects	O
of	O
the	O
image	O
.	O
</s>
<s>
An	O
image	B-Application
captioning	I-Application
system	O
involves	O
two	O
sub-tasks	O
.	O
</s>
<s>
Recent	O
research	O
utilizes	O
deep	O
learning	O
approaches	O
through	O
features	O
from	O
a	O
pre-trained	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
such	O
as	O
AlexNet	O
,	O
VGG	O
or	O
Caffe	O
,	O
where	O
caption	O
generators	O
use	O
an	O
activation	O
layer	O
from	O
the	O
pre-trained	O
network	O
as	O
their	O
input	O
features	O
.	O
</s>
<s>
Text	B-General_Concept
Generation	I-General_Concept
,	O
the	O
second	O
task	O
,	O
is	O
performed	O
using	O
a	O
wide	O
range	O
of	O
techniques	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
the	O
Midge	O
system	O
,	O
input	O
images	O
are	O
represented	O
as	O
triples	O
consisting	O
of	O
object/stuff	O
detections	O
,	O
action/pose	O
detections	O
and	O
spatial	O
relations	O
.	O
</s>
<s>
Notwithstanding	O
the	O
recent	O
introduction	O
of	O
Flickr30K	O
,	O
MS	O
COCO	O
and	O
other	O
large	O
datasets	O
have	O
enabled	O
the	O
training	O
of	O
more	O
complex	O
models	O
such	O
as	O
neural	O
networks	O
,	O
it	O
has	O
been	O
argued	O
that	O
research	O
in	O
image	B-Application
captioning	I-Application
could	O
benefit	O
from	O
larger	O
and	O
diversified	O
datasets	O
.	O
</s>
<s>
Other	O
open	O
challenges	O
include	O
visual	O
question-answering	B-Algorithm
(	O
VQA	O
)	O
,	O
as	O
well	O
as	O
the	O
construction	O
and	O
evaluation	O
multilingual	O
repositories	O
for	O
image	O
description	O
.	O
</s>
<s>
Another	O
area	O
where	O
NLG	O
has	O
been	O
widely	O
applied	O
is	O
automated	O
dialogue	O
systems	O
,	O
frequently	O
in	O
the	O
form	O
of	O
chatbots	B-Application
.	O
</s>
<s>
A	O
chatbot	B-Application
or	O
chatterbot	O
is	O
a	O
software	B-General_Concept
application	O
used	O
to	O
conduct	O
an	O
on-line	O
chat	O
conversation	O
via	O
text	O
or	O
text-to-speech	B-Application
,	O
in	O
lieu	O
of	O
providing	O
direct	O
contact	O
with	O
a	O
live	O
human	O
agent	O
.	O
</s>
<s>
While	O
natural	B-Language
language	I-Language
processing	I-Language
(	O
NLP	B-Language
)	O
techniques	O
are	O
applied	O
in	O
deciphering	O
human	O
input	O
,	O
NLG	O
informs	O
the	O
output	O
part	O
of	O
the	O
chatbot	B-Application
algorithms	O
in	O
facilitating	O
real-time	O
dialogues	O
.	O
</s>
<s>
Early	O
chatbot	B-Application
systems	O
,	O
including	O
Cleverbot	B-Protocol
created	O
by	O
Rollo	O
Carpenter	O
in	O
1988	O
and	O
published	O
in	O
1997	O
,	O
reply	O
to	O
questions	O
by	O
identifying	O
how	O
a	O
human	O
has	O
responded	O
to	O
the	O
same	O
question	O
in	O
a	O
conversation	O
database	O
using	O
information	B-Library
retrieval	I-Library
(	O
IR	B-Application
)	O
techniques	O
.	O
</s>
<s>
Modern	O
chatbot	B-Application
systems	O
predominantly	O
rely	O
on	O
machine	O
learning	O
(	O
ML	O
)	O
models	O
,	O
such	O
as	O
sequence-to-sequence	O
learning	O
and	O
reinforcement	O
learning	O
to	O
generate	O
natural	O
language	O
output	O
.	O
</s>
<s>
For	O
example	O
,	O
the	O
Alibaba	O
shopping	O
assistant	O
first	O
uses	O
an	O
IR	B-Application
approach	O
to	O
retrieve	O
the	O
best	O
candidates	O
from	O
the	O
knowledge	O
base	O
,	O
then	O
uses	O
the	O
ML-driven	O
seq2seq	O
model	O
re-rank	O
the	O
candidate	O
responses	O
and	O
generate	O
the	O
answer	O
.	O
</s>
<s>
Creative	O
language	B-General_Concept
generation	I-General_Concept
by	O
NLG	O
has	O
been	O
hypothesized	O
since	O
the	O
field	O
’s	O
origins	O
.	O
</s>
<s>
Metrics	O
:	O
compare	O
generated	O
texts	O
to	O
texts	O
written	O
by	O
people	O
from	O
the	O
same	O
input	O
data	O
,	O
using	O
an	O
automatic	O
metric	O
such	O
as	O
BLEU	O
,	O
METEOR	B-Language
,	O
ROUGE	B-General_Concept
and	O
LEPOR	O
.	O
</s>
<s>
Hence	O
(	O
as	O
in	O
other	O
areas	O
of	O
NLP	B-Language
)	O
task-based	O
evaluations	O
are	O
the	O
exception	O
,	O
not	O
the	O
norm	O
.	O
</s>
<s>
In	O
any	O
case	O
,	O
human	O
ratings	O
are	O
the	O
most	O
popular	O
evaluation	O
technique	O
in	O
NLG	O
;	O
this	O
is	O
contrast	O
to	O
machine	B-General_Concept
translation	I-General_Concept
,	O
where	O
metrics	O
are	O
widely	O
used	O
.	O
</s>
<s>
A	O
confident	O
but	O
unfaithful	O
response	O
is	O
a	O
hallucination	B-General_Concept
.	O
</s>
<s>
In	O
Natural	B-Language
Language	I-Language
Processing	I-Language
,	O
a	O
hallucination	B-General_Concept
is	O
often	O
defined	O
as	O
"	O
generated	O
content	O
that	O
is	O
nonsensical	O
or	O
unfaithful	O
to	O
the	O
provided	O
source	O
content	O
"	O
.	O
</s>
