-DOCSTART- NN O O

Recognition NN O I-Task
of NN O I-Task
proper NN O I-OtherScientificTerm
nouns NN O I-OtherScientificTerm
in NN O O
Japanese NN O I-Material
text NN O I-Material
has NN O O
been NN O O
studied NN O O
as NN O O
a NN O O
part NN O O
of NN O O
the NN O O
more NN O O
general NN O O
problem NN O O
of NN O O
morphological NN O I-Task
analysis NN O I-Task
in NN O O
Japanese NN O I-Task
text NN O I-Task
processing NN O I-Task
-LRB- NN O O
-LSB- NN O O
1 NN O O
-RSB- NN O O
-LSB- NN O O
2 NN O O
-RSB- NN O O
-RRB- NN O O
. NN O O

It NN O I-Generic
has NN O O
also NN O O
been NN O O
studied NN O O
in NN O O
the NN O O
framework NN O O
of NN O O
Japanese NN O I-Task
information NN O I-Task
extraction NN O I-Task
-LRB- NN O O
-LSB- NN O O
3 NN O O
-RSB- NN O O
-RRB- NN O O
in NN O O
recent NN O O
years NN O O
. NN O O

Our NN O O
approach NN O I-Generic
to NN O O
the NN O O
Multi-lingual NN O I-Task
Evaluation NN O I-Task
Task NN O I-Task
-LRB- NN O I-Task
MET NN O I-Task
-RRB- NN O I-Task
for NN O O
Japanese NN O I-Material
text NN O I-Material
is NN O O
to NN O O
consider NN O O
the NN O O
given NN O O
task NN O I-Generic
as NN O O
a NN O O
morphological NN O I-Task
analysis NN O I-Task
problem NN O I-Task
in NN O O
Japanese NN O I-Material
. NN O O

Our NN O O
morphological NN O I-Method
analyzer NN O I-Method
has NN O O
done NN O O
all NN O O
the NN O O
necessary NN O O
work NN O O
for NN O O
the NN O O
recognition NN O I-Task
and NN O I-Task
classification NN O I-Task
of NN O I-Task
proper NN O I-OtherScientificTerm
names NN O I-OtherScientificTerm
, NN O I-OtherScientificTerm
numerical NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
temporal NN O I-OtherScientificTerm
expressions NN O I-OtherScientificTerm
, NN O O
i.e. NN O O
Named NN O I-OtherScientificTerm
Entity NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
NE NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
items NN O I-OtherScientificTerm
in NN O O
the NN O O
Japanese NN O I-Material
text NN O I-Material
. NN O O

The NN O O
analyzer NN O I-Generic
is NN O O
called NN O O
`` NN O I-Method
Amorph NN O I-Method
'' NN O I-Method
. NN O O

Amorph NN O I-Method
recognizes NN O O
NE NN O I-OtherScientificTerm
items NN O I-OtherScientificTerm
in NN O O
two NN O O
stages NN O O
: NN O O
dictionary NN O I-Method
lookup NN O I-Method
and NN O O
rule NN O I-Method
application NN O I-Method
. NN O O

First NN O O
, NN O O
it NN O I-Generic
uses NN O O
several NN O O
kinds NN O O
of NN O O
dictionaries NN O I-OtherScientificTerm
to NN O O
segment NN O O
and NN O O
tag NN O O
Japanese NN O I-OtherScientificTerm
character NN O I-OtherScientificTerm
strings NN O I-OtherScientificTerm
. NN O O

Second NN O O
, NN O O
based NN O O
on NN O O
the NN O O
information NN O O
resulting NN O O
from NN O O
the NN O O
dictionary NN O I-Method
lookup NN O I-Method
stage NN O I-Method
, NN O O
a NN O O
set NN O O
of NN O O
rules NN O I-OtherScientificTerm
is NN O O
applied NN O O
to NN O O
the NN O O
segmented NN O O
strings NN O O
in NN O O
order NN O O
to NN O O
identify NN O O
NE NN O I-OtherScientificTerm
items NN O I-OtherScientificTerm
. NN O O

When NN O O
a NN O O
segment NN O O
is NN O O
found NN O O
to NN O O
be NN O O
an NN O O
NE NN O I-OtherScientificTerm
item NN O I-OtherScientificTerm
, NN O O
this NN O O
information NN O O
is NN O O
added NN O O
to NN O O
the NN O O
segment NN O O
and NN O O
it NN O O
is NN O O
used NN O O
to NN O O
generate NN O O
the NN O O
final NN O O
output NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
propose NN O O
to NN O O
incorporate NN O O
a NN O O
priori NN O I-OtherScientificTerm
geometric NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
in NN O O
a NN O O
3 NN O I-Method
-- NN O I-Method
D NN O I-Method
stereo NN O I-Method
reconstruction NN O I-Method
scheme NN O I-Method
to NN O O
cope NN O O
with NN O O
the NN O O
many NN O O
cases NN O O
where NN O O
image NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
alone NN O O
is NN O O
not NN O O
sufficient NN O O
to NN O O
accurately NN O O
recover NN O O
3 NN O I-OtherScientificTerm
-- NN O I-OtherScientificTerm
D NN O I-OtherScientificTerm
shape NN O I-OtherScientificTerm
. NN O O

Our NN O O
approach NN O I-Generic
is NN O O
based NN O O
on NN O O
the NN O O
iterative NN O I-Method
deformation NN O I-Method
of NN O I-Method
a NN O I-Method
3 NN O I-Method
-- NN O I-Method
D NN O I-Method
surface NN O I-Method
mesh NN O I-Method
to NN O O
minimize NN O O
an NN O O
objective NN O I-OtherScientificTerm
function NN O I-OtherScientificTerm
. NN O O

We NN O O
show NN O O
that NN O O
combining NN O O
anisotropic NN O I-Method
meshing NN O I-Method
with NN O O
a NN O O
non-quadratic NN O I-Method
approach NN O I-Method
to NN O O
regularization NN O I-OtherScientificTerm
enables NN O O
us NN O O
to NN O O
obtain NN O O
satisfactory NN O O
reconstruction NN O I-Task
results NN O O
using NN O O
triangulations NN O I-OtherScientificTerm
with NN O O
few NN O O
vertices NN O I-OtherScientificTerm
. NN O O

Structural NN O I-OtherScientificTerm
or NN O I-OtherScientificTerm
numerical NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
can NN O O
then NN O O
be NN O O
added NN O O
locally NN O O
to NN O O
the NN O O
reconstruction NN O I-Method
process NN O I-Method
through NN O O
a NN O O
constrained NN O I-Method
optimization NN O I-Method
scheme NN O I-Method
. NN O O

They NN O I-Generic
improve NN O O
the NN O O
reconstruction NN O I-Task
results NN O O
and NN O O
enforce NN O O
their NN O O
consistency NN O O
with NN O O
a NN O O
priori NN O I-OtherScientificTerm
knowledge NN O I-OtherScientificTerm
about NN O O
object NN O I-OtherScientificTerm
shape NN O I-OtherScientificTerm
. NN O O

The NN O O
strong NN O O
description NN O O
and NN O O
modeling NN O I-OtherScientificTerm
properties NN O I-OtherScientificTerm
of NN O O
differential NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
make NN O O
them NN O I-Generic
useful NN O O
tools NN O O
that NN O O
can NN O O
be NN O O
efficiently NN O O
used NN O O
as NN O O
constraints NN O O
for NN O O
3 NN O I-Task
-- NN O I-Task
D NN O I-Task
reconstruction NN O I-Task
. NN O O

-DOCSTART- NN O O

This NN O O
work NN O O
proposes NN O O
a NN O O
new NN O O
research NN O O
direction NN O O
to NN O O
address NN O O
the NN O O
lack NN O I-Task
of NN O I-Task
structures NN O I-Task
in NN O I-Task
traditional NN O I-Task
n-gram NN O I-Method
models NN O I-Method
. NN O O

It NN O O
is NN O O
based NN O O
on NN O O
a NN O O
weakly NN O I-Task
supervised NN O I-Task
dependency NN O I-Task
parser NN O I-Task
that NN O O
can NN O O
model NN O O
speech NN O I-OtherScientificTerm
syntax NN O I-OtherScientificTerm
without NN O O
relying NN O O
on NN O O
any NN O O
annotated NN O I-Material
training NN O I-Material
corpus NN O I-Material
. NN O O

Labeled NN O I-OtherScientificTerm
data NN O I-OtherScientificTerm
is NN O O
replaced NN O O
by NN O O
a NN O O
few NN O O
hand-crafted NN O I-OtherScientificTerm
rules NN O I-OtherScientificTerm
that NN O O
encode NN O O
basic NN O O
syntactic NN O I-OtherScientificTerm
knowledge NN O I-OtherScientificTerm
. NN O O

Bayesian NN O I-Method
inference NN O I-Method
then NN O O
samples NN O O
the NN O O
rules NN O I-OtherScientificTerm
, NN O O
disambiguating NN O O
and NN O O
combining NN O O
them NN O I-Generic
to NN O O
create NN O O
complex NN O I-OtherScientificTerm
tree NN O I-OtherScientificTerm
structures NN O I-OtherScientificTerm
that NN O O
maximize NN O O
a NN O O
discriminative NN O I-OtherScientificTerm
model NN O I-OtherScientificTerm
's NN O I-OtherScientificTerm
posterior NN O I-OtherScientificTerm
on NN O O
a NN O O
target NN O O
unlabeled NN O I-Material
corpus NN O I-Material
. NN O O

This NN O O
posterior NN O I-Generic
encodes NN O O
sparse NN O I-OtherScientificTerm
se-lectional NN O I-OtherScientificTerm
preferences NN O I-OtherScientificTerm
between NN O O
a NN O O
head NN O O
word NN O O
and NN O O
its NN O O
dependents NN O O
. NN O O

The NN O O
model NN O I-Generic
is NN O O
evaluated NN O O
on NN O O
English NN O I-Material
and NN O I-Material
Czech NN O I-Material
newspaper NN O I-Material
texts NN O I-Material
, NN O O
and NN O O
is NN O O
then NN O O
validated NN O O
on NN O O
French NN O I-Material
broadcast NN O I-Material
news NN O I-Material
transcriptions NN O I-Material
. NN O O

-DOCSTART- NN O O

Listen-Communicate-Show NN O I-Task
-LRB- NN O I-Task
LCS NN O I-Task
-RRB- NN O I-Task
is NN O O
a NN O O
new NN O O
paradigm NN O O
for NN O O
human NN O I-Task
interaction NN O I-Task
with NN O I-Task
data NN O I-Task
sources NN O I-Task
. NN O O

We NN O O
integrate NN O O
a NN O O
spoken NN O I-Method
language NN O I-Method
understanding NN O I-Method
system NN O I-Method
with NN O O
intelligent NN O I-Method
mobile NN O I-Method
agents NN O I-Method
that NN O O
mediate NN O O
between NN O O
users NN O O
and NN O O
information NN O I-Material
sources NN O I-Material
. NN O O

We NN O O
have NN O O
built NN O O
and NN O O
will NN O O
demonstrate NN O O
an NN O O
application NN O O
of NN O O
this NN O O
approach NN O I-Generic
called NN O O
LCS-Marine NN O I-Task
. NN O O

Using NN O O
LCS-Marine NN O I-Task
, NN O O
tactical NN O O
personnel NN O O
can NN O O
converse NN O O
with NN O O
their NN O O
logistics NN O O
system NN O O
to NN O O
place NN O O
a NN O O
supply NN O O
or NN O O
information NN O O
request NN O O
. NN O O

The NN O O
request NN O O
is NN O O
passed NN O O
to NN O O
a NN O O
mobile NN O I-Method
, NN O I-Method
intelligent NN O I-Method
agent NN O I-Method
for NN O O
execution NN O O
at NN O O
the NN O O
appropriate NN O O
database NN O O
. NN O O

Requestors NN O O
can NN O O
also NN O O
instruct NN O O
the NN O O
system NN O I-Generic
to NN O O
notify NN O O
them NN O O
when NN O O
the NN O O
status NN O O
of NN O O
a NN O O
request NN O O
changes NN O O
or NN O O
when NN O O
a NN O O
request NN O O
is NN O O
complete NN O O
. NN O O

We NN O O
have NN O O
demonstrated NN O O
this NN O O
capability NN O O
in NN O O
several NN O O
field NN O O
exercises NN O O
with NN O O
the NN O O
Marines NN O O
and NN O O
are NN O O
currently NN O O
developing NN O O
applications NN O O
of NN O O
this NN O O
technology NN O O
in NN O O
new NN O I-Material
domains NN O I-Material
. NN O O

-DOCSTART- NN O O

A NN O O
domain NN O I-Method
independent NN O I-Method
model NN O I-Method
is NN O O
proposed NN O O
for NN O O
the NN O O
automated NN O I-Task
interpretation NN O I-Task
of NN O I-Task
nominal NN O I-OtherScientificTerm
compounds NN O I-OtherScientificTerm
in NN O O
English NN O I-Material
. NN O O

This NN O O
model NN O I-Generic
is NN O O
meant NN O O
to NN O O
account NN O O
for NN O O
productive NN O I-OtherScientificTerm
rules NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
interpretation NN O I-OtherScientificTerm
which NN O O
are NN O O
inferred NN O O
from NN O O
the NN O O
morpho-syntactic NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
semantic NN O I-OtherScientificTerm
characteristics NN O I-OtherScientificTerm
of NN O O
the NN O O
nominal NN O I-OtherScientificTerm
constituents NN O I-OtherScientificTerm
. NN O O

In NN O O
particular NN O O
, NN O O
we NN O O
make NN O O
extensive NN O O
use NN O O
of NN O O
Pustejovsky NN O I-Method
's NN O I-Method
principles NN O I-Method
concerning NN O O
the NN O O
predicative NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
associated NN O O
with NN O O
nominals NN O I-OtherScientificTerm
. NN O O

We NN O O
argue NN O O
that NN O O
it NN O O
is NN O O
necessary NN O O
to NN O O
draw NN O O
a NN O O
line NN O O
between NN O O
generalizable NN O I-OtherScientificTerm
semantic NN O I-OtherScientificTerm
principles NN O I-OtherScientificTerm
and NN O O
domain-specific NN O I-OtherScientificTerm
semantic NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
. NN O O

We NN O O
explain NN O O
this NN O O
distinction NN O O
and NN O O
we NN O O
show NN O O
how NN O O
this NN O O
model NN O I-Generic
may NN O O
be NN O O
applied NN O O
to NN O O
the NN O O
interpretation NN O I-Task
of NN O I-Task
compounds NN O I-Task
in NN O O
real NN O O
texts NN O O
, NN O O
provided NN O O
that NN O O
complementary NN O O
semantic NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
are NN O O
retrieved NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
present NN O O
a NN O O
new NN O O
method NN O I-Generic
for NN O O
detecting NN O I-Task
interest NN O I-Task
points NN O I-Task
using NN O O
histogram NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
. NN O O

Unlike NN O O
existing NN O O
interest NN O I-Method
point NN O I-Method
detectors NN O I-Method
, NN O O
which NN O O
measure NN O O
pixel-wise NN O I-Metric
differences NN O I-Metric
in NN O I-Metric
image NN O I-Metric
intensity NN O I-Metric
, NN O O
our NN O O
detectors NN O I-Generic
incorporate NN O O
histogram-based NN O I-Method
representations NN O I-Method
, NN O O
and NN O O
thus NN O O
can NN O O
find NN O O
image NN O O
regions NN O O
that NN O O
present NN O O
a NN O O
distinct NN O O
distribution NN O O
in NN O O
the NN O O
neighborhood NN O O
. NN O O

The NN O O
proposed NN O O
detectors NN O I-Generic
are NN O O
able NN O O
to NN O O
capture NN O O
large-scale NN O I-OtherScientificTerm
structures NN O I-OtherScientificTerm
and NN O O
distinctive NN O I-OtherScientificTerm
textured NN O I-OtherScientificTerm
patterns NN O I-OtherScientificTerm
, NN O O
and NN O O
exhibit NN O O
strong NN O O
invariance NN O O
to NN O O
rotation NN O I-OtherScientificTerm
, NN O O
illumination NN O I-OtherScientificTerm
variation NN O I-OtherScientificTerm
, NN O O
and NN O O
blur NN O I-OtherScientificTerm
. NN O O

The NN O O
experimental NN O O
results NN O O
show NN O O
that NN O O
the NN O O
proposed NN O O
histogram-based NN O I-Method
interest NN O I-Method
point NN O I-Method
detectors NN O I-Method
perform NN O O
particularly NN O O
well NN O O
for NN O O
the NN O O
tasks NN O O
of NN O O
matching NN O I-Task
textured NN O I-Task
scenes NN O I-Task
under NN O O
blur NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
illumination NN O I-OtherScientificTerm
changes NN O I-OtherScientificTerm
, NN O O
in NN O O
terms NN O O
of NN O O
repeatability NN O I-Metric
and NN O O
distinctiveness NN O I-Metric
. NN O O

An NN O O
extension NN O O
of NN O O
our NN O O
method NN O I-Generic
to NN O O
space-time NN O I-Task
interest NN O I-Task
point NN O I-Task
detection NN O I-Task
for NN O O
action NN O I-Task
classification NN O I-Task
is NN O O
also NN O O
presented NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
have NN O O
implemented NN O O
a NN O O
restricted NN O I-Method
domain NN O I-Method
parser NN O I-Method
called NN O O
Plume NN O I-Method
. NN O O

Building NN O O
on NN O O
previous NN O O
work NN O O
at NN O O
Carnegie-Mellon NN O O
University NN O O
e.g. NN O O
-LSB- NN O O
4 NN O O
, NN O O
5 NN O O
, NN O O
8 NN O O
-RSB- NN O O
, NN O O
Plume NN O I-Method
's NN O I-Method
approach NN O I-Method
to NN O O
parsing NN O I-Task
is NN O O
based NN O O
on NN O O
semantic NN O I-OtherScientificTerm
caseframe NN O I-OtherScientificTerm
instantiation NN O I-OtherScientificTerm
. NN O O

This NN O O
has NN O O
the NN O O
advantages NN O O
of NN O O
efficiency NN O O
on NN O O
grammatical NN O I-OtherScientificTerm
input NN O I-OtherScientificTerm
, NN O O
and NN O O
robustness NN O I-Metric
in NN O O
the NN O O
face NN O O
of NN O O
ungrammatical NN O I-OtherScientificTerm
input NN O I-OtherScientificTerm
. NN O O

While NN O O
Plume NN O I-Method
is NN O O
well NN O O
adapted NN O O
to NN O O
simple NN O O
declarative NN O I-Material
and NN O I-Material
imperative NN O I-Material
utterances NN O I-Material
, NN O O
it NN O I-Generic
handles NN O O
passives NN O I-OtherScientificTerm
, NN O O
relative NN O I-OtherScientificTerm
clauses NN O I-OtherScientificTerm
and NN O O
interrogatives NN O I-OtherScientificTerm
in NN O O
an NN O O
ad NN O O
hoc NN O O
manner NN O O
leading NN O O
to NN O O
patchy NN O I-OtherScientificTerm
syntactic NN O I-OtherScientificTerm
coverage NN O I-OtherScientificTerm
. NN O O

This NN O O
paper NN O O
outlines NN O O
Plume NN O I-Method
as NN O O
it NN O I-Generic
currently NN O O
exists NN O O
and NN O O
describes NN O O
our NN O O
detailed NN O O
design NN O O
for NN O O
extending NN O O
Plume NN O I-Method
to NN O O
handle NN O O
passives NN O I-OtherScientificTerm
, NN O O
relative NN O I-OtherScientificTerm
clauses NN O I-OtherScientificTerm
, NN O O
and NN O O
interrogatives NN O I-OtherScientificTerm
in NN O O
a NN O O
general NN O O
manner NN O O
. NN O O

-DOCSTART- NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
present NN O O
an NN O O
unlexicalized NN O I-Method
parser NN O I-Method
for NN O O
German NN O I-Material
which NN O O
employs NN O O
smoothing NN O I-Method
and NN O O
suffix NN O I-Method
analysis NN O I-Method
to NN O O
achieve NN O O
a NN O O
labelled NN O I-Metric
bracket NN O I-Metric
F-score NN O I-Metric
of NN O O
76.2 NN O O
, NN O O
higher NN O O
than NN O O
previously NN O O
reported NN O O
results NN O O
on NN O O
the NN O O
NEGRA NN O I-Material
corpus NN O I-Material
. NN O O

In NN O O
addition NN O O
to NN O O
the NN O O
high NN O O
accuracy NN O I-Metric
of NN O O
the NN O O
model NN O I-Generic
, NN O O
the NN O O
use NN O O
of NN O O
smoothing NN O I-Method
in NN O O
an NN O O
unlexicalized NN O I-Method
parser NN O I-Method
allows NN O O
us NN O O
to NN O O
better NN O O
examine NN O O
the NN O O
interplay NN O O
between NN O O
smoothing NN O I-Method
and NN O O
parsing NN O I-Method
results NN O O
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
presents NN O O
an NN O O
unsupervised NN O I-Method
learning NN O I-Method
approach NN O I-Method
to NN O O
disambiguate NN O O
various NN O O
relations NN O I-OtherScientificTerm
between NN O I-OtherScientificTerm
named NN O I-OtherScientificTerm
entities NN O I-OtherScientificTerm
by NN O O
use NN O O
of NN O O
various NN O O
lexical NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
syntactic NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
from NN O O
the NN O O
contexts NN O O
. NN O O

It NN O I-Generic
works NN O O
by NN O O
calculating NN O O
eigenvectors NN O I-OtherScientificTerm
of NN O O
an NN O O
adjacency NN O I-OtherScientificTerm
graph NN O I-OtherScientificTerm
's NN O I-OtherScientificTerm
Laplacian NN O I-OtherScientificTerm
to NN O O
recover NN O O
a NN O O
submanifold NN O I-OtherScientificTerm
of NN O O
data NN O O
from NN O O
a NN O O
high NN O I-OtherScientificTerm
dimensionality NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
and NN O O
then NN O O
performing NN O O
cluster NN O I-Task
number NN O I-Task
estimation NN O I-Task
on NN O O
the NN O O
eigenvectors NN O I-OtherScientificTerm
. NN O O

Experiment NN O O
results NN O O
on NN O O
ACE NN O I-Material
corpora NN O I-Material
show NN O O
that NN O O
this NN O O
spectral NN O I-Method
clustering NN O I-Method
based NN O I-Method
approach NN O I-Method
outperforms NN O O
the NN O O
other NN O O
clustering NN O I-Method
methods NN O I-Method
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
proposes NN O O
a NN O O
generic NN O O
mathematical NN O I-Method
formalism NN O I-Method
for NN O O
the NN O O
combination NN O O
of NN O O
various NN O O
structures NN O I-Generic
: NN O O
strings NN O I-OtherScientificTerm
, NN O O
trees NN O I-OtherScientificTerm
, NN O O
dags NN O I-OtherScientificTerm
, NN O O
graphs NN O I-OtherScientificTerm
, NN O O
and NN O O
products NN O O
of NN O O
them NN O I-Generic
. NN O O

The NN O O
polarization NN O I-OtherScientificTerm
of NN O O
the NN O O
objects NN O O
of NN O O
the NN O O
elementary NN O I-OtherScientificTerm
structures NN O I-OtherScientificTerm
controls NN O O
the NN O O
saturation NN O O
of NN O O
the NN O O
final NN O O
structure NN O O
. NN O O

This NN O O
formalism NN O I-Generic
is NN O O
both NN O O
elementary NN O O
and NN O O
powerful NN O O
enough NN O O
to NN O O
strongly NN O O
simulate NN O O
many NN O O
grammar NN O I-Method
formalisms NN O I-Method
, NN O O
such NN O O
as NN O O
rewriting NN O I-Method
systems NN O I-Method
, NN O O
dependency NN O I-Method
grammars NN O I-Method
, NN O O
TAG NN O I-OtherScientificTerm
, NN O O
HPSG NN O I-OtherScientificTerm
and NN O O
LFG NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

A NN O O
mixed-signal NN O I-Method
paradigm NN O I-Method
is NN O O
presented NN O O
for NN O O
high-resolution NN O I-Task
parallel NN O I-Task
inner-product NN O I-Task
computation NN O I-Task
in NN O O
very NN O O
high NN O O
dimensions NN O O
, NN O O
suitable NN O O
for NN O O
efficient NN O O
implementation NN O O
of NN O O
kernels NN O I-Method
in NN O O
image NN O I-Task
processing NN O I-Task
. NN O O

At NN O O
the NN O O
core NN O O
of NN O O
the NN O O
externally NN O I-Method
digital NN O I-Method
architecture NN O I-Method
is NN O O
a NN O O
high-density NN O I-OtherScientificTerm
, NN O I-OtherScientificTerm
low-power NN O I-OtherScientificTerm
analog NN O I-OtherScientificTerm
array NN O I-OtherScientificTerm
performing NN O O
binary-binary NN O I-Method
partial NN O I-Method
matrix-vector NN O I-Method
multiplication NN O I-Method
. NN O O

Full NN O I-Metric
digital NN O I-Metric
resolution NN O I-Metric
is NN O O
maintained NN O O
even NN O O
with NN O O
low-resolution NN O I-Method
analog-to-digital NN O I-Method
conversion NN O I-Method
, NN O O
owing NN O O
to NN O O
random NN O I-OtherScientificTerm
statistics NN O I-OtherScientificTerm
in NN O O
the NN O O
analog NN O I-OtherScientificTerm
summation NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
binary NN O I-OtherScientificTerm
products NN O I-OtherScientificTerm
. NN O O

A NN O O
random NN O I-Method
modulation NN O I-Method
scheme NN O I-Method
produces NN O O
near-Bernoulli NN O I-OtherScientificTerm
statistics NN O I-OtherScientificTerm
even NN O O
for NN O O
highly NN O I-Material
correlated NN O I-Material
inputs NN O I-Material
. NN O O

The NN O O
approach NN O I-Generic
is NN O O
validated NN O O
with NN O O
real NN O I-Material
image NN O I-Material
data NN O I-Material
, NN O O
and NN O O
with NN O O
experimental NN O O
results NN O O
from NN O O
a NN O O
CID/DRAM NN O I-OtherScientificTerm
analog NN O I-OtherScientificTerm
array NN O I-OtherScientificTerm
prototype NN O I-OtherScientificTerm
in NN O O
0.5 NN O O
cents NN O O
m NN O O
CMOS NN O O
. NN O O

-DOCSTART- NN O O

In NN O O
this NN O O
paper NN O O
we NN O O
specialize NN O O
the NN O O
projective NN O I-Method
unifocal NN O I-Method
, NN O I-Method
bifo-cal NN O I-Method
, NN O I-Method
and NN O I-Method
trifocal NN O I-Method
tensors NN O I-Method
to NN O O
the NN O O
affine NN O I-OtherScientificTerm
case NN O I-OtherScientificTerm
, NN O O
and NN O O
show NN O O
how NN O O
the NN O O
tensors NN O I-Generic
obtained NN O O
relate NN O O
to NN O O
the NN O O
registered NN O I-OtherScientificTerm
tensors NN O I-OtherScientificTerm
encountered NN O O
in NN O O
previous NN O O
work NN O O
. NN O O

This NN O O
enables NN O O
us NN O O
to NN O O
obtain NN O O
an NN O O
affine NN O I-OtherScientificTerm
specialization NN O I-OtherScientificTerm
of NN O O
known NN O O
projective NN O O
relations NN O O
connecting NN O O
points NN O O
and NN O O
lines NN O O
across NN O O
two NN O O
or NN O O
three NN O O
views NN O O
. NN O O

In NN O O
the NN O O
simpler NN O O
case NN O O
of NN O O
affine NN O I-Method
cameras NN O I-Method
we NN O O
give NN O O
neccessary NN O O
and NN O O
sufficient NN O O
constraints NN O O
on NN O O
the NN O O
components NN O O
of NN O O
the NN O O
trifocal NN O I-Method
tensor NN O I-Method
, NN O O
together NN O O
with NN O O
a NN O O
simple NN O O
geometric NN O I-OtherScientificTerm
interpretation NN O I-OtherScientificTerm
. NN O O

Finally NN O O
, NN O O
we NN O O
show NN O O
how NN O O
the NN O O
estimation NN O I-Task
of NN O I-Task
the NN O I-Task
tensors NN O I-Generic
from NN O O
point NN O I-OtherScientificTerm
correspondences NN O I-OtherScientificTerm
is NN O O
achieved NN O O
through NN O O
factorization NN O I-Method
, NN O O
and NN O O
discuss NN O O
the NN O O
estimation NN O I-Generic
from NN O O
line NN O I-OtherScientificTerm
correspondences NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
presents NN O O
an NN O O
algorithm NN O O
for NN O O
selecting NN O O
an NN O O
appropriate NN O O
classifier NN O O
word NN O O
for NN O O
a NN O O
noun NN O O
. NN O O

In NN O O
Thai NN O I-Material
language NN O I-Material
, NN O O
it NN O O
frequently NN O O
happens NN O O
that NN O O
there NN O O
is NN O O
fluctuation NN O O
in NN O O
the NN O O
choice NN O O
of NN O O
classifier NN O I-Method
for NN O O
a NN O O
given NN O O
concrete NN O O
noun NN O O
, NN O O
both NN O O
from NN O O
the NN O O
point NN O O
of NN O O
view NN O O
of NN O O
the NN O O
whole NN O O
speech NN O O
community NN O O
and NN O O
individual NN O O
speakers NN O O
. NN O O

Basically NN O O
, NN O O
there NN O O
is NN O O
no NN O O
exact NN O O
rule NN O O
for NN O O
classifier NN O I-Task
selection NN O I-Task
. NN O O

As NN O O
far NN O O
as NN O O
we NN O O
can NN O O
do NN O O
in NN O O
the NN O O
rule-based NN O I-Method
approach NN O I-Method
is NN O O
to NN O O
give NN O O
a NN O O
default NN O O
rule NN O O
to NN O O
pick NN O O
up NN O O
a NN O O
corresponding NN O O
classifier NN O I-Method
of NN O O
each NN O O
noun NN O O
. NN O O

Registration NN O O
of NN O O
classifier NN O I-Method
for NN O O
each NN O O
noun NN O O
is NN O O
limited NN O O
to NN O O
the NN O O
type NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
unit NN O I-OtherScientificTerm
classifier NN O I-OtherScientificTerm
because NN O O
other NN O O
types NN O O
are NN O O
open NN O O
due NN O O
to NN O O
the NN O O
meaning NN O O
of NN O O
representation NN O O
. NN O O

We NN O O
propose NN O O
a NN O O
corpus-based NN O I-Method
method NN O I-Method
-LRB- NN O O
Biber NN O O
,1993 NN O O
; NN O O
Nagao NN O O
,1993 NN O O
; NN O O
Smadja NN O O
,1993 NN O O
-RRB- NN O O
which NN O O
generates NN O O
Noun NN O I-Method
Classifier NN O I-Method
Associations NN O I-Method
-LRB- NN O I-Method
NCA NN O I-Method
-RRB- NN O I-Method
to NN O O
overcome NN O O
the NN O O
problems NN O O
in NN O O
classifier NN O I-Task
assignment NN O I-Task
and NN O O
semantic NN O I-Task
construction NN O I-Task
of NN O I-Task
noun NN O I-Task
phrase NN O I-Task
. NN O O

The NN O O
NCA NN O I-Method
is NN O O
created NN O O
statistically NN O O
from NN O O
a NN O O
large NN O O
corpus NN O O
and NN O O
recomposed NN O O
under NN O O
concept NN O I-OtherScientificTerm
hierarchy NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
and NN O O
frequency NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
occurrences NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

The NN O O
perception NN O I-Task
of NN O I-Task
transparent NN O I-Task
objects NN O I-Task
from NN O O
images NN O I-Material
is NN O O
known NN O O
to NN O O
be NN O O
a NN O O
very NN O O
hard NN O O
problem NN O O
in NN O O
vision NN O O
. NN O O

Given NN O O
a NN O O
single NN O O
image NN O O
, NN O O
it NN O O
is NN O O
difficult NN O O
to NN O O
even NN O O
detect NN O O
the NN O O
presence NN O O
of NN O O
transparent NN O I-OtherScientificTerm
objects NN O I-OtherScientificTerm
in NN O O
the NN O O
scene NN O O
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
explore NN O O
what NN O O
can NN O O
be NN O O
said NN O O
about NN O O
transparent NN O I-OtherScientificTerm
objects NN O I-OtherScientificTerm
by NN O O
a NN O O
moving NN O O
observer NN O O
. NN O O

We NN O O
show NN O O
how NN O O
features NN O I-OtherScientificTerm
that NN O O
are NN O O
imaged NN O O
through NN O O
a NN O O
transparent NN O I-OtherScientificTerm
object NN O I-OtherScientificTerm
behave NN O O
differently NN O O
from NN O O
those NN O I-Generic
that NN O O
are NN O O
rigidly NN O O
attached NN O O
to NN O O
the NN O O
scene NN O O
. NN O O

We NN O O
present NN O O
a NN O O
novel NN O O
model-based NN O I-Method
approach NN O I-Method
to NN O O
recover NN O O
the NN O O
shapes NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
the NN O I-OtherScientificTerm
poses NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
transparent NN O I-OtherScientificTerm
objects NN O I-OtherScientificTerm
from NN O O
known NN O I-OtherScientificTerm
motion NN O I-OtherScientificTerm
. NN O O

The NN O O
objects NN O I-Generic
can NN O O
be NN O O
complex NN O O
in NN O O
that NN O O
they NN O I-Generic
may NN O O
be NN O O
composed NN O O
of NN O O
multiple NN O I-OtherScientificTerm
layers NN O I-OtherScientificTerm
with NN O O
different NN O O
refractive NN O I-OtherScientificTerm
indices NN O I-OtherScientificTerm
. NN O O

We NN O O
have NN O O
conducted NN O O
numerous NN O O
simulations NN O O
to NN O O
verify NN O O
the NN O O
practical NN O O
feasibility NN O O
of NN O O
our NN O O
algorithm NN O I-Generic
. NN O O

We NN O O
have NN O O
applied NN O O
it NN O I-Generic
to NN O O
real NN O I-Material
scenes NN O I-Material
that NN O O
include NN O O
transparent NN O I-OtherScientificTerm
objects NN O I-OtherScientificTerm
and NN O O
recovered NN O O
the NN O O
shapes NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
the NN O I-OtherScientificTerm
objects NN O I-OtherScientificTerm
with NN O O
high NN O O
accuracy NN O I-Metric
. NN O O

-DOCSTART- NN O O

We NN O O
propose NN O O
a NN O O
novel NN O O
probabilistic NN O I-Method
framework NN O I-Method
for NN O O
learning NN O O
visual NN O I-Task
models NN O I-Task
of NN O I-Task
3D NN O I-Task
object NN O I-Task
categories NN O I-Task
by NN O O
combining NN O O
appearance NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
and NN O O
geometric NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
. NN O O

Objects NN O O
are NN O O
represented NN O O
as NN O O
a NN O O
coherent NN O O
ensemble NN O O
of NN O O
parts NN O O
that NN O O
are NN O O
consistent NN O O
under NN O O
3D NN O I-OtherScientificTerm
viewpoint NN O I-OtherScientificTerm
transformations NN O I-OtherScientificTerm
. NN O O

Each NN O O
part NN O O
is NN O O
a NN O O
collection NN O O
of NN O O
salient NN O I-OtherScientificTerm
image NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
. NN O O

A NN O O
generative NN O I-Method
framework NN O I-Method
is NN O O
used NN O O
for NN O O
learning NN O O
a NN O O
model NN O I-Generic
that NN O O
captures NN O O
the NN O O
relative NN O O
position NN O O
of NN O O
parts NN O O
within NN O O
each NN O O
of NN O O
the NN O O
discretized NN O I-OtherScientificTerm
viewpoints NN O I-OtherScientificTerm
. NN O O

Contrary NN O O
to NN O O
most NN O O
of NN O O
the NN O O
existing NN O O
mixture NN O I-Method
of NN O I-Method
viewpoints NN O I-Method
models NN O I-Method
, NN O O
our NN O O
model NN O I-Generic
establishes NN O O
explicit NN O O
correspondences NN O O
of NN O O
parts NN O O
across NN O O
different NN O O
viewpoints NN O I-OtherScientificTerm
of NN O O
the NN O O
object NN O O
class NN O O
. NN O O

Given NN O O
a NN O O
new NN O O
image NN O I-Material
, NN O O
detection NN O I-Task
and NN O O
classification NN O I-Task
are NN O O
achieved NN O O
by NN O O
determining NN O O
the NN O O
position NN O I-OtherScientificTerm
and NN O O
viewpoint NN O I-OtherScientificTerm
of NN O O
the NN O O
model NN O O
that NN O O
maximize NN O O
recognition NN O I-Metric
scores NN O I-Metric
of NN O O
the NN O O
candidate NN O O
objects NN O O
. NN O O

Our NN O O
approach NN O I-Generic
is NN O O
among NN O O
the NN O O
first NN O O
to NN O O
propose NN O O
a NN O O
generative NN O I-Method
proba-bilistic NN O I-Method
framework NN O I-Method
for NN O O
3D NN O I-Task
object NN O I-Task
categorization NN O I-Task
. NN O O

We NN O O
test NN O O
our NN O O
algorithm NN O I-Generic
on NN O O
the NN O O
detection NN O I-Task
task NN O I-Task
and NN O O
the NN O O
viewpoint NN O I-Task
classification NN O I-Task
task NN O I-Task
by NN O O
using NN O O
'' NN O O
car NN O O
'' NN O O
category NN O O
from NN O O
both NN O O
the NN O O
Savarese NN O O
et NN O O
al. NN O O
2007 NN O O
and NN O O
PASCAL NN O I-Material
VOC NN O I-Material
2006 NN O I-Material
datasets NN O I-Material
. NN O O

We NN O O
show NN O O
promising NN O O
results NN O O
in NN O O
both NN O O
the NN O O
detection NN O I-Task
and NN O I-Task
viewpoint NN O I-Task
classification NN O I-Task
tasks NN O I-Task
on NN O O
these NN O O
two NN O O
challenging NN O O
datasets NN O I-Generic
. NN O O

-DOCSTART- NN O O

We NN O O
present NN O O
an NN O O
application NN O O
of NN O O
ambiguity NN O I-Method
packing NN O I-Method
and NN O I-Method
stochastic NN O I-Method
disambiguation NN O I-Method
techniques NN O I-Method
for NN O O
Lexical-Functional NN O I-Method
Grammars NN O I-Method
-LRB- NN O I-Method
LFG NN O I-Method
-RRB- NN O I-Method
to NN O O
the NN O O
domain NN O O
of NN O O
sentence NN O I-Task
condensation NN O I-Task
. NN O O

Our NN O O
system NN O I-Generic
incorporates NN O O
a NN O O
linguistic NN O I-Method
parser/generator NN O I-Method
for NN O O
LFG NN O I-Method
, NN O O
a NN O O
transfer NN O I-Method
component NN O I-Method
for NN O O
parse NN O I-Task
reduction NN O I-Task
operating NN O O
on NN O O
packed NN O I-OtherScientificTerm
parse NN O I-OtherScientificTerm
forests NN O I-OtherScientificTerm
, NN O O
and NN O O
a NN O O
maximum-entropy NN O I-Method
model NN O I-Method
for NN O O
stochastic NN O I-Task
output NN O I-Task
selection NN O I-Task
. NN O O

Furthermore NN O O
, NN O O
we NN O O
propose NN O O
the NN O O
use NN O O
of NN O O
standard NN O O
parser NN O I-Method
evaluation NN O I-Method
methods NN O I-Method
for NN O O
automatically NN O O
evaluating NN O O
the NN O O
summarization NN O I-Metric
quality NN O I-Metric
of NN O O
sentence NN O I-Method
condensation NN O I-Method
systems NN O I-Method
. NN O O

An NN O O
experimental NN O O
evaluation NN O O
of NN O O
summarization NN O I-Metric
quality NN O I-Metric
shows NN O O
a NN O O
close NN O O
correlation NN O O
between NN O O
the NN O O
automatic NN O I-Method
parse-based NN O I-Method
evaluation NN O I-Method
and NN O O
a NN O O
manual NN O I-Method
evaluation NN O I-Method
of NN O O
generated NN O O
strings NN O O
. NN O O

Overall NN O O
summarization NN O I-Metric
quality NN O I-Metric
of NN O O
the NN O O
proposed NN O O
system NN O I-Generic
is NN O O
state-of-the-art NN O O
, NN O O
with NN O O
guaranteed NN O O
grammaticality NN O I-Metric
of NN O O
the NN O O
system NN O O
output NN O O
due NN O O
to NN O O
the NN O O
use NN O O
of NN O O
a NN O O
constraint-based NN O I-Method
parser/generator NN O I-Method
. NN O O

-DOCSTART- NN O O

The NN O O
robust NN O I-Method
principal NN O I-Method
component NN O I-Method
analysis NN O I-Method
-LRB- NN O I-Method
robust NN O I-Method
PCA NN O I-Method
-RRB- NN O I-Method
problem NN O I-Method
has NN O O
been NN O O
considered NN O O
in NN O O
many NN O O
machine NN O I-Task
learning NN O I-Task
applications NN O I-Task
, NN O O
where NN O O
the NN O O
goal NN O O
is NN O O
to NN O O
decompose NN O O
the NN O O
data NN O I-OtherScientificTerm
matrix NN O I-OtherScientificTerm
to NN O O
a NN O O
low NN O I-OtherScientificTerm
rank NN O I-OtherScientificTerm
part NN O I-OtherScientificTerm
plus NN O O
a NN O O
sparse NN O I-OtherScientificTerm
residual NN O I-OtherScientificTerm
. NN O O

While NN O O
current NN O O
approaches NN O I-Generic
are NN O O
developed NN O O
by NN O O
only NN O O
considering NN O O
the NN O O
low NN O I-OtherScientificTerm
rank NN O I-OtherScientificTerm
plus NN O I-OtherScientificTerm
sparse NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
, NN O O
in NN O O
many NN O O
applications NN O O
, NN O O
side NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
of NN O O
row NN O O
and/or NN O O
column NN O O
entities NN O O
may NN O O
also NN O O
be NN O O
given NN O O
, NN O O
and NN O O
it NN O O
is NN O O
still NN O O
unclear NN O O
to NN O O
what NN O O
extent NN O O
could NN O O
such NN O O
information NN O I-Generic
help NN O O
robust NN O I-Method
PCA NN O I-Method
. NN O O

Thus NN O O
, NN O O
in NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
study NN O O
the NN O O
problem NN O O
of NN O O
robust NN O I-Method
PCA NN O I-Method
with NN O O
side NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
, NN O O
where NN O O
both NN O O
prior NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
and NN O O
features NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
entities NN O I-OtherScientificTerm
are NN O O
exploited NN O O
for NN O O
recovery NN O I-Task
. NN O O

We NN O O
propose NN O O
a NN O O
convex NN O I-Task
problem NN O I-Task
to NN O O
incorporate NN O O
side NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
in NN O O
robust NN O I-Method
PCA NN O I-Method
and NN O O
show NN O O
that NN O O
the NN O O
low NN O I-OtherScientificTerm
rank NN O I-OtherScientificTerm
matrix NN O I-OtherScientificTerm
can NN O O
be NN O O
exactly NN O O
recovered NN O O
via NN O O
the NN O O
proposed NN O O
method NN O I-Generic
under NN O O
certain NN O O
conditions NN O O
. NN O O

In NN O O
particular NN O O
, NN O O
our NN O O
guarantee NN O O
suggests NN O O
that NN O O
a NN O O
substantial NN O O
amount NN O O
of NN O O
low NN O I-OtherScientificTerm
rank NN O I-OtherScientificTerm
matrices NN O I-OtherScientificTerm
, NN O O
which NN O O
can NN O O
not NN O O
be NN O O
recovered NN O O
by NN O O
standard NN O O
robust NN O I-Method
PCA NN O I-Method
, NN O O
become NN O O
re-coverable NN O O
by NN O O
our NN O O
proposed NN O O
method NN O I-Generic
. NN O O

The NN O O
result NN O O
theoretically NN O O
justifies NN O O
the NN O O
effectiveness NN O O
of NN O O
features NN O I-OtherScientificTerm
in NN O O
robust NN O I-Method
PCA NN O I-Method
. NN O O

In NN O O
addition NN O O
, NN O O
we NN O O
conduct NN O O
synthetic NN O O
experiments NN O O
as NN O O
well NN O O
as NN O O
a NN O O
real NN O O
application NN O O
on NN O O
noisy NN O I-Task
image NN O I-Task
classification NN O I-Task
to NN O O
show NN O O
that NN O O
our NN O O
method NN O I-Generic
also NN O O
improves NN O O
the NN O O
performance NN O O
in NN O O
practice NN O O
by NN O O
exploiting NN O O
side NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
presents NN O O
necessary NN O O
and NN O O
sufficient NN O O
conditions NN O O
for NN O O
the NN O O
use NN O O
of NN O O
demonstrative NN O I-OtherScientificTerm
expressions NN O I-OtherScientificTerm
in NN O O
English NN O I-Material
and NN O O
discusses NN O O
implications NN O I-Generic
for NN O O
current NN O O
discourse NN O I-Method
processing NN O I-Method
algorithms NN O I-Method
. NN O O

We NN O O
examine NN O O
a NN O O
broad NN O O
range NN O O
of NN O O
texts NN O O
to NN O O
show NN O O
how NN O O
the NN O O
distribution NN O O
of NN O O
demonstrative NN O I-OtherScientificTerm
forms NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
functions NN O I-OtherScientificTerm
is NN O O
genre NN O O
dependent NN O O
. NN O O

This NN O O
research NN O O
is NN O O
part NN O O
of NN O O
a NN O O
larger NN O O
study NN O O
of NN O O
anaphoric NN O I-OtherScientificTerm
expressions NN O I-OtherScientificTerm
, NN O O
the NN O O
results NN O O
of NN O O
which NN O O
will NN O O
be NN O O
incorporated NN O O
into NN O O
a NN O O
natural NN O I-Method
language NN O I-Method
generation NN O I-Method
system NN O I-Method
. NN O O

-DOCSTART- NN O O

In NN O O
the NN O O
study NN O O
of NN O O
expressive NN O I-Task
speech NN O I-Task
communication NN O I-Task
, NN O O
it NN O O
is NN O O
commonly NN O O
accepted NN O O
that NN O O
the NN O O
emotion NN O O
perceived NN O O
by NN O O
the NN O O
listener NN O O
is NN O O
a NN O O
good NN O O
approximation NN O O
of NN O O
the NN O O
intended NN O O
emotion NN O O
conveyed NN O O
by NN O O
the NN O O
speaker NN O O
. NN O O

This NN O O
paper NN O O
analyzes NN O O
the NN O O
validity NN O O
of NN O O
this NN O O
assumption NN O O
by NN O O
comparing NN O O
the NN O O
mismatches NN O O
between NN O O
the NN O O
assessments NN O O
made NN O O
by NN O O
na NN O O
¨ NN O O
ıve NN O O
listeners NN O O
and NN O O
by NN O O
the NN O O
speakers NN O O
that NN O O
generated NN O O
the NN O O
data NN O O
. NN O O

The NN O O
analysis NN O O
is NN O O
based NN O O
on NN O O
the NN O O
hypothesis NN O O
that NN O O
people NN O O
are NN O O
better NN O O
decoders NN O O
of NN O O
their NN O O
own NN O O
emotions NN O O
. NN O O

Therefore NN O O
, NN O O
self-assessments NN O O
will NN O O
be NN O O
closer NN O O
to NN O O
the NN O O
intended NN O O
emotions NN O O
. NN O O

Using NN O O
the NN O O
IEMOCAP NN O I-Material
database NN O I-Material
, NN O O
discrete NN O I-Task
-LRB- NN O I-Task
categorical NN O I-Task
-RRB- NN O I-Task
and NN O I-Task
continuous NN O I-Task
-LRB- NN O I-Task
attribute NN O I-Task
-RRB- NN O I-Task
emotional NN O I-Task
assessments NN O I-Task
evaluated NN O O
by NN O O
the NN O O
actors NN O O
and NN O O
na NN O O
¨ NN O O
ıve NN O O
listeners NN O O
are NN O O
compared NN O O
. NN O O

The NN O O
results NN O O
indicate NN O O
that NN O O
there NN O O
is NN O O
a NN O O
mismatch NN O O
between NN O O
the NN O O
expression NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
perception NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
emotion NN O I-OtherScientificTerm
. NN O O

The NN O O
speakers NN O O
in NN O O
the NN O O
database NN O I-Generic
assigned NN O O
their NN O O
own NN O O
emotions NN O O
to NN O O
more NN O O
specific NN O O
emotional NN O O
categories NN O O
, NN O O
which NN O O
led NN O O
to NN O O
more NN O O
extreme NN O O
values NN O O
in NN O O
the NN O O
activation-valence NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

The NN O O
problem NN O O
of NN O O
blind NN O I-Task
separation NN O I-Task
of NN O I-Task
underdetermined NN O I-Task
instantaneous NN O I-Task
mixtures NN O I-Task
of NN O I-Task
independent NN O I-Task
signals NN O I-Task
is NN O O
addressed NN O O
through NN O O
a NN O O
method NN O I-Generic
relying NN O O
on NN O O
nonstationarity NN O I-OtherScientificTerm
of NN O O
the NN O O
original NN O I-OtherScientificTerm
signals NN O I-OtherScientificTerm
. NN O O

The NN O O
signals NN O I-Generic
are NN O O
assumed NN O O
to NN O O
be NN O O
piecewise NN O O
stationary NN O O
with NN O O
varying NN O O
variances NN O O
in NN O O
different NN O O
epochs NN O O
. NN O O

In NN O O
comparison NN O O
with NN O O
previous NN O O
works NN O O
, NN O O
in NN O O
this NN O O
paper NN O O
it NN O O
is NN O O
assumed NN O O
that NN O O
the NN O O
signals NN O I-Generic
are NN O O
not NN O O
i.i.d. NN O O
in NN O O
each NN O O
epoch NN O O
, NN O O
but NN O O
obey NN O O
a NN O O
first-order NN O I-Method
autoregressive NN O I-Method
model NN O I-Method
. NN O O

This NN O O
model NN O I-Generic
was NN O O
shown NN O O
to NN O O
be NN O O
more NN O O
appropriate NN O O
for NN O O
blind NN O I-Task
separation NN O I-Task
of NN O I-Task
natural NN O I-Task
speech NN O I-Task
signals NN O I-Task
. NN O I-Task

A NN O O
separation NN O I-Method
method NN O I-Method
is NN O O
proposed NN O O
that NN O O
is NN O O
nearly NN O O
statistically NN O O
efficient NN O O
-LRB- NN O O
approaching NN O O
the NN O O
corresponding NN O O
Cramér-Rao NN O I-OtherScientificTerm
lower NN O I-OtherScientificTerm
bound NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
, NN O O
if NN O O
the NN O O
separated NN O O
signals NN O O
obey NN O O
the NN O O
assumed NN O I-Generic
model NN O I-Generic
. NN O O

In NN O O
the NN O O
case NN O O
of NN O O
natural NN O I-OtherScientificTerm
speech NN O I-OtherScientificTerm
signals NN O I-OtherScientificTerm
, NN O O
the NN O O
method NN O I-Generic
is NN O O
shown NN O O
to NN O O
have NN O O
separation NN O I-Metric
accuracy NN O I-Metric
better NN O O
than NN O O
the NN O O
state-of-the-art NN O O
methods NN O I-Generic
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
proposes NN O O
to NN O O
use NN O O
a NN O O
convolution NN O I-Method
kernel NN O I-Method
over NN O I-Method
parse NN O I-Method
trees NN O I-Method
to NN O O
model NN O O
syntactic NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
for NN O O
relation NN O I-Task
extraction NN O I-Task
. NN O O

Our NN O O
study NN O O
reveals NN O O
that NN O O
the NN O O
syntactic NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
embedded NN O O
in NN O O
a NN O O
parse NN O I-OtherScientificTerm
tree NN O I-OtherScientificTerm
are NN O O
very NN O O
effective NN O O
for NN O O
relation NN O I-Task
extraction NN O I-Task
and NN O O
these NN O O
features NN O I-Generic
can NN O O
be NN O O
well NN O O
captured NN O O
by NN O O
the NN O O
convolution NN O I-Method
tree NN O I-Method
kernel NN O I-Method
. NN O O

Evaluation NN O O
on NN O O
the NN O O
ACE NN O I-Material
2003 NN O I-Material
corpus NN O I-Material
shows NN O O
that NN O O
the NN O O
convolution NN O I-Method
kernel NN O I-Method
over NN O I-Method
parse NN O I-Method
trees NN O I-Method
can NN O O
achieve NN O O
comparable NN O O
performance NN O O
with NN O O
the NN O O
previous NN O O
best-reported NN O O
feature-based NN O I-Method
methods NN O I-Method
on NN O O
the NN O O
24 NN O O
ACE NN O O
relation NN O O
subtypes NN O O
. NN O O

It NN O O
also NN O O
shows NN O O
that NN O O
our NN O O
method NN O I-Generic
significantly NN O O
outperforms NN O O
the NN O O
previous NN O O
two NN O O
dependency NN O I-Method
tree NN O I-Method
kernels NN O I-Method
on NN O O
the NN O O
5 NN O O
ACE NN O O
relation NN O O
major NN O O
types NN O O
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
presents NN O O
the NN O O
results NN O O
of NN O O
automatically NN O I-Task
inducing NN O I-Task
a NN O I-Task
Combinatory NN O I-OtherScientificTerm
Categorial NN O I-OtherScientificTerm
Grammar NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
CCG NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
lexicon NN O I-OtherScientificTerm
from NN O O
a NN O O
Turkish NN O I-Material
dependency NN O I-Material
treebank NN O I-Material
. NN O O

The NN O O
fact NN O O
that NN O O
Turkish NN O I-Material
is NN O O
an NN O O
agglutinating NN O I-Material
free NN O I-Material
word NN O I-Material
order NN O I-Material
language NN O I-Material
presents NN O O
a NN O O
challenge NN O O
for NN O O
language NN O I-Method
theories NN O I-Method
. NN O O

We NN O O
explored NN O O
possible NN O O
ways NN O O
to NN O O
obtain NN O O
a NN O O
compact NN O I-OtherScientificTerm
lexicon NN O I-OtherScientificTerm
, NN O O
consistent NN O O
with NN O O
CCG NN O I-Method
principles NN O I-Method
, NN O O
from NN O O
a NN O O
treebank NN O I-Generic
which NN O O
is NN O O
an NN O O
order NN O O
of NN O O
magnitude NN O O
smaller NN O O
than NN O O
Penn NN O I-Material
WSJ NN O I-Material
. NN O O

-DOCSTART- NN O O

While NN O O
sentence NN O I-Method
extraction NN O I-Method
as NN O O
an NN O O
approach NN O O
to NN O O
summarization NN O I-Task
has NN O O
been NN O O
shown NN O O
to NN O O
work NN O O
in NN O O
documents NN O O
of NN O O
certain NN O O
genres NN O O
, NN O O
because NN O O
of NN O O
the NN O O
conversational NN O O
nature NN O O
of NN O O
email NN O I-Material
communication NN O I-Material
where NN O O
utterances NN O O
are NN O O
made NN O O
in NN O O
relation NN O O
to NN O O
one NN O O
made NN O O
previously NN O O
, NN O O
sentence NN O I-Method
extraction NN O I-Method
may NN O O
not NN O O
capture NN O O
the NN O O
necessary NN O O
segments NN O O
of NN O O
dialogue NN O O
that NN O O
would NN O O
make NN O O
a NN O O
summary NN O O
coherent NN O O
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
present NN O O
our NN O O
work NN O O
on NN O O
the NN O O
detection NN O I-Task
of NN O I-Task
question-answer NN O I-Task
pairs NN O I-Task
in NN O O
an NN O O
email NN O I-Material
conversation NN O I-Material
for NN O O
the NN O O
task NN O O
of NN O O
email NN O I-Task
summarization NN O I-Task
. NN O O

We NN O O
show NN O O
that NN O O
various NN O O
features NN O I-OtherScientificTerm
based NN O O
on NN O O
the NN O O
structure NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
email-threads NN O I-OtherScientificTerm
can NN O O
be NN O O
used NN O O
to NN O O
improve NN O O
upon NN O O
lexical NN O I-OtherScientificTerm
similarity NN O I-OtherScientificTerm
of NN O O
discourse NN O I-OtherScientificTerm
segments NN O I-OtherScientificTerm
for NN O O
question-answer NN O I-Task
pairing NN O I-Task
. NN O O

-DOCSTART- NN O O

In NN O O
this NN O O
paper NN O O
we NN O O
discuss NN O O
object NN O I-Task
detection NN O I-Task
when NN O O
only NN O O
a NN O O
small NN O O
number NN O O
of NN O O
training NN O O
examples NN O O
are NN O O
given NN O O
. NN O O

Specifically NN O O
, NN O O
we NN O O
show NN O O
how NN O O
to NN O O
incorporate NN O O
a NN O O
simple NN O O
prior NN O I-OtherScientificTerm
on NN O I-OtherScientificTerm
the NN O I-OtherScientificTerm
distribution NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
natural NN O I-OtherScientificTerm
images NN O I-OtherScientificTerm
into NN O O
support NN O I-Method
vector NN O I-Method
machines NN O I-Method
. NN O O

SVMs NN O I-Method
are NN O O
known NN O O
to NN O O
be NN O O
robust NN O O
to NN O O
overfitting NN O I-OtherScientificTerm
; NN O O
however NN O O
, NN O O
a NN O O
few NN O O
training NN O O
examples NN O O
usually NN O O
do NN O O
not NN O O
represent NN O O
well NN O O
the NN O O
structure NN O O
of NN O O
the NN O O
class NN O O
. NN O O

Thus NN O O
the NN O O
resulting NN O O
detectors NN O I-Method
are NN O O
not NN O O
robust NN O O
and NN O O
highly NN O O
depend NN O O
on NN O O
the NN O O
choice NN O O
of NN O O
the NN O O
training NN O O
examples NN O O
. NN O O

We NN O O
incorporate NN O O
the NN O O
prior NN O I-OtherScientificTerm
on NN O I-OtherScientificTerm
natural NN O I-OtherScientificTerm
images NN O I-OtherScientificTerm
by NN O O
requiring NN O O
that NN O O
the NN O O
separating NN O O
hyperplane NN O I-OtherScientificTerm
will NN O O
not NN O O
only NN O O
yield NN O O
a NN O O
wide NN O O
margin NN O O
, NN O O
but NN O O
also NN O O
that NN O O
the NN O O
corresponding NN O O
positive NN O O
half NN O O
space NN O O
will NN O O
have NN O O
a NN O O
low NN O O
probability NN O O
to NN O O
contain NN O O
natural NN O O
images NN O O
-LRB- NN O O
the NN O O
background NN O O
-RRB- NN O O
. NN O O

Our NN O O
experiments NN O O
on NN O O
real NN O I-Material
data NN O I-Material
sets NN O I-Material
show NN O O
that NN O O
the NN O O
resulting NN O O
detector NN O I-Method
is NN O O
more NN O O
robust NN O O
to NN O O
the NN O O
choice NN O O
of NN O O
training NN O O
examples NN O O
, NN O O
and NN O O
substantially NN O O
improves NN O O
both NN O O
linear NN O I-Method
and NN O I-Method
kernel NN O I-Method
SVM NN O I-Method
when NN O O
trained NN O O
on NN O O
10 NN O O
positive NN O O
and NN O O
10 NN O O
negative NN O O
examples NN O O
. NN O O

-DOCSTART- NN O O

Although NN O O
the NN O O
study NN O O
of NN O O
clustering NN O I-Task
is NN O O
centered NN O O
around NN O O
an NN O O
intuitively NN O O
compelling NN O O
goal NN O O
, NN O O
it NN O O
has NN O O
been NN O O
very NN O O
difficult NN O O
to NN O O
develop NN O O
a NN O O
unified NN O I-Method
framework NN O I-Method
for NN O O
reasoning NN O I-Task
about NN O O
it NN O O
at NN O O
a NN O O
technical NN O O
level NN O O
, NN O O
and NN O O
profoundly NN O O
diverse NN O O
approaches NN O O
to NN O O
clustering NN O O
abound NN O O
in NN O O
the NN O O
research NN O O
community NN O O
. NN O O

Here NN O O
we NN O O
suggest NN O O
a NN O O
formal NN O O
perspective NN O O
on NN O O
the NN O O
difficulty NN O O
in NN O O
finding NN O O
such NN O O
a NN O O
unification NN O I-Method
, NN O O
in NN O O
the NN O O
form NN O O
of NN O O
an NN O O
impossibility NN O I-Method
theorem NN O I-Method
: NN O O
for NN O O
a NN O O
set NN O O
of NN O O
three NN O O
simple NN O O
properties NN O O
, NN O O
we NN O O
show NN O O
that NN O O
there NN O O
is NN O O
no NN O O
clustering NN O I-OtherScientificTerm
function NN O I-OtherScientificTerm
satisfying NN O O
all NN O O
three NN O O
. NN O O

Relaxations NN O O
of NN O O
these NN O O
properties NN O O
expose NN O O
some NN O O
of NN O O
the NN O O
interesting NN O O
-LRB- NN O O
and NN O O
unavoidable NN O O
-RRB- NN O O
trade-offs NN O O
at NN O O
work NN O O
in NN O O
well-studied NN O I-Method
clustering NN O I-Method
techniques NN O I-Method
such NN O O
as NN O O
single-linkage NN O I-Method
, NN O O
sum-of-pairs NN O I-Method
, NN O O
k-means NN O I-Method
, NN O O
and NN O O
k-median NN O I-Method
. NN O O

-DOCSTART- NN O O

We NN O O
investigate NN O O
independent NN O I-Method
and NN O I-Method
relevant NN O I-Method
event-based NN O I-Method
extractive NN O I-Method
mutli-document NN O I-Method
summarization NN O I-Method
approaches NN O I-Method
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
events NN O O
are NN O O
defined NN O O
as NN O O
event NN O O
terms NN O O
and NN O O
associated NN O O
event NN O O
elements NN O O
. NN O O

With NN O O
independent NN O I-Method
approach NN O I-Method
, NN O O
we NN O O
identify NN O O
important NN O O
contents NN O O
by NN O O
frequency NN O O
of NN O O
events NN O O
. NN O O

With NN O O
relevant NN O I-Method
approach NN O I-Method
, NN O O
we NN O O
identify NN O O
important NN O O
contents NN O O
by NN O O
PageRank NN O I-Method
algorithm NN O I-Method
on NN O O
the NN O O
event NN O I-OtherScientificTerm
map NN O I-OtherScientificTerm
constructed NN O O
from NN O O
documents NN O I-Material
. NN O O

Experimental NN O O
results NN O O
are NN O O
encouraging NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
present NN O O
a NN O O
scanning NN O I-Method
method NN O I-Method
that NN O O
recovers NN O O
dense NN O I-OtherScientificTerm
sub-pixel NN O I-OtherScientificTerm
camera-projector NN O I-OtherScientificTerm
correspondence NN O I-OtherScientificTerm
without NN O O
requiring NN O O
any NN O O
photometric NN O I-OtherScientificTerm
calibration NN O I-OtherScientificTerm
nor NN O O
preliminary NN O O
knowledge NN O O
of NN O O
their NN O O
relative NN O I-OtherScientificTerm
geometry NN O I-OtherScientificTerm
. NN O O

Subpixel NN O I-Metric
accuracy NN O I-Metric
is NN O O
achieved NN O O
by NN O O
considering NN O O
several NN O O
zero-crossings NN O I-OtherScientificTerm
defined NN O O
by NN O O
the NN O O
difference NN O O
between NN O O
pairs NN O O
of NN O O
unstructured NN O I-OtherScientificTerm
patterns NN O I-OtherScientificTerm
. NN O O

We NN O O
use NN O O
gray-level NN O I-OtherScientificTerm
band-pass NN O I-OtherScientificTerm
white NN O I-OtherScientificTerm
noise NN O I-OtherScientificTerm
patterns NN O I-OtherScientificTerm
that NN O O
increase NN O O
robustness NN O I-Metric
to NN O O
indirect NN O I-OtherScientificTerm
lighting NN O I-OtherScientificTerm
and NN O O
scene NN O I-OtherScientificTerm
discontinuities NN O I-OtherScientificTerm
. NN O O

Simulated NN O O
and NN O O
experimental NN O O
results NN O O
show NN O O
that NN O O
our NN O O
method NN O I-Generic
recovers NN O O
scene NN O I-OtherScientificTerm
geometry NN O I-OtherScientificTerm
with NN O O
high NN O O
subpixel NN O I-Metric
precision NN O I-Metric
, NN O O
and NN O O
that NN O O
it NN O I-Generic
can NN O O
handle NN O O
many NN O O
challenges NN O O
of NN O O
active NN O I-Task
reconstruction NN O I-Task
systems NN O I-Task
. NN O O

We NN O O
compare NN O O
our NN O O
results NN O O
to NN O O
state NN O I-Generic
of NN O I-Generic
the NN O I-Generic
art NN O I-Generic
methods NN O I-Generic
such NN O O
as NN O O
mi-cro NN O I-Method
phase NN O I-Method
shifting NN O I-Method
and NN O O
modulated NN O I-Method
phase NN O I-Method
shifting NN O I-Method
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
describes NN O O
a NN O O
novel NN O O
system NN O I-Generic
for NN O O
acquiring NN O I-Task
adjectival NN O I-Task
subcategorization NN O I-OtherScientificTerm
frames NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
scfs NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
and NN O O
associated NN O O
frequency NN O O
information NN O O
from NN O O
English NN O I-Material
corpus NN O I-Material
data NN O I-Material
. NN O O

The NN O O
system NN O I-Generic
incorporates NN O O
a NN O O
decision-tree NN O I-Method
classifier NN O I-Method
for NN O O
30 NN O O
scf NN O O
types NN O O
which NN O O
tests NN O O
for NN O O
the NN O O
presence NN O O
of NN O O
grammatical NN O I-OtherScientificTerm
relations NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
grs NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
in NN O O
the NN O O
output NN O O
of NN O O
a NN O O
robust NN O I-Method
statistical NN O I-Method
parser NN O I-Method
. NN O O

It NN O I-Generic
uses NN O O
a NN O O
powerful NN O O
pattern-matching NN O I-OtherScientificTerm
language NN O I-OtherScientificTerm
to NN O O
classify NN O O
grs NN O I-OtherScientificTerm
into NN O O
frames NN O O
hierarchically NN O O
in NN O O
a NN O O
way NN O O
that NN O O
mirrors NN O O
inheritance-based NN O I-OtherScientificTerm
lexica NN O I-OtherScientificTerm
. NN O O

The NN O O
experiments NN O O
show NN O O
that NN O O
the NN O O
system NN O I-Generic
is NN O O
able NN O O
to NN O O
detect NN O O
scf NN O O
types NN O O
with NN O O
70 NN O O
% NN O O
precision NN O I-Metric
and NN O O
66 NN O O
% NN O O
recall NN O I-Metric
rate NN O O
. NN O O

A NN O O
new NN O O
tool NN O I-Generic
for NN O O
linguistic NN O I-Task
annotation NN O I-Task
of NN O I-Task
scfs NN O I-OtherScientificTerm
in NN O O
corpus NN O O
data NN O O
is NN O O
also NN O O
introduced NN O O
which NN O O
can NN O O
considerably NN O O
alleviate NN O O
the NN O O
process NN O O
of NN O O
obtaining NN O O
training NN O I-Material
and NN O I-Material
test NN O I-Material
data NN O I-Material
for NN O O
subcategorization NN O I-Task
acquisition NN O I-Task
. NN O O

-DOCSTART- NN O O

Machine NN O I-Task
transliteration/back-transliteration NN O I-Task
plays NN O O
an NN O O
important NN O O
role NN O O
in NN O O
many NN O O
multilingual NN O I-Task
speech NN O I-Task
and NN O I-Task
language NN O I-Task
applications NN O I-Task
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
a NN O O
novel NN O O
framework NN O I-Generic
for NN O O
machine NN O I-Task
transliteration/backtransliteration NN O I-Task
that NN O O
allows NN O O
us NN O O
to NN O O
carry NN O O
out NN O O
direct NN O I-Method
orthographical NN O I-Method
mapping NN O I-Method
-LRB- NN O I-Method
DOM NN O I-Method
-RRB- NN O I-Method
between NN O O
two NN O O
different NN O O
languages NN O O
is NN O O
presented NN O O
. NN O O

Under NN O O
this NN O O
framework NN O I-Generic
, NN O O
a NN O O
joint NN O I-Method
source-channel NN O I-Method
transliteration NN O I-Method
model NN O I-Method
, NN O O
also NN O O
called NN O O
n-gram NN O I-Method
transliteration NN O I-Method
model NN O I-Method
-LRB- NN O I-Method
n-gram NN O I-Method
TM NN O I-Method
-RRB- NN O I-Method
, NN O O
is NN O O
further NN O O
proposed NN O O
to NN O O
model NN O O
the NN O O
transliteration NN O I-Method
process NN O I-Method
. NN O O

We NN O O
evaluate NN O O
the NN O O
proposed NN O O
methods NN O I-Generic
through NN O O
several NN O O
transliteration/backtransliteration NN O I-Task
experiments NN O O
for NN O O
English/Chinese NN O I-Material
and NN O I-Material
English/Japanese NN O I-Material
language NN O I-Material
pairs NN O I-Material
. NN O O

Our NN O O
study NN O O
reveals NN O O
that NN O O
the NN O O
proposed NN O O
method NN O I-Generic
not NN O O
only NN O O
reduces NN O O
an NN O O
extensive NN O O
system NN O O
development NN O O
effort NN O O
but NN O O
also NN O O
improves NN O O
the NN O O
transliteration NN O I-Metric
accuracy NN O I-Metric
significantly NN O O
. NN O O

-DOCSTART- NN O O

A NN O O
bio-inspired NN O I-Method
model NN O I-Method
for NN O O
an NN O O
analog NN O I-Task
programmable NN O I-Task
array NN O I-Task
processor NN O I-Task
-LRB- NN O I-Task
APAP NN O I-Task
-RRB- NN O I-Task
, NN O O
based NN O O
on NN O O
studies NN O O
on NN O O
the NN O O
vertebrate NN O I-OtherScientificTerm
retina NN O I-OtherScientificTerm
, NN O O
has NN O O
permitted NN O O
the NN O O
realization NN O O
of NN O O
complex NN O I-OtherScientificTerm
programmable NN O I-OtherScientificTerm
spatio-temporal NN O I-OtherScientificTerm
dynamics NN O I-OtherScientificTerm
in NN O O
VLSI NN O I-Task
. NN O O

This NN O O
model NN O I-Generic
mimics NN O O
the NN O O
way NN O O
in NN O O
which NN O O
images NN O I-Material
are NN O O
processed NN O O
in NN O O
the NN O O
visual NN O I-Method
pathway NN O I-Method
, NN O O
rendering NN O O
a NN O O
feasible NN O O
alternative NN O O
for NN O O
the NN O O
implementation NN O O
of NN O O
early NN O O
vision NN O I-Task
applications NN O I-Task
in NN O O
standard NN O O
technologies NN O O
. NN O O

A NN O O
prototype NN O I-OtherScientificTerm
chip NN O I-OtherScientificTerm
has NN O O
been NN O O
designed NN O O
and NN O O
fabricated NN O O
in NN O O
a NN O O
0.5 NN O O
µm NN O O
standard NN O O
CMOS NN O I-OtherScientificTerm
process NN O I-OtherScientificTerm
. NN O O

Computing NN O I-Metric
power NN O I-Metric
per NN O I-Metric
area NN O I-Metric
and NN O O
power NN O I-Metric
consumption NN O I-Metric
is NN O O
amongst NN O O
the NN O O
highest NN O O
reported NN O O
for NN O O
a NN O O
single NN O O
chip NN O O
. NN O O

Design NN O O
challenges NN O O
, NN O O
trade-offs NN O O
and NN O O
some NN O O
experimental NN O O
results NN O O
are NN O O
presented NN O O
in NN O O
this NN O O
paper NN O O
. NN O O

-DOCSTART- NN O O

Determiners NN O I-Method
play NN O O
an NN O O
important NN O O
role NN O O
in NN O O
conveying NN O O
the NN O O
meaning NN O O
of NN O O
an NN O O
utterance NN O O
, NN O O
but NN O O
they NN O O
have NN O O
often NN O O
been NN O O
disregarded NN O O
, NN O O
perhaps NN O O
because NN O O
it NN O O
seemed NN O O
more NN O O
important NN O O
to NN O O
devise NN O O
methods NN O O
to NN O O
grasp NN O O
the NN O O
global NN O O
meaning NN O O
of NN O O
a NN O O
sentence NN O O
, NN O O
even NN O O
if NN O O
not NN O O
in NN O O
a NN O O
precise NN O O
way NN O O
. NN O O

Another NN O O
problem NN O O
with NN O O
determiners NN O I-Method
is NN O O
their NN O O
inherent NN O O
ambiguity NN O I-OtherScientificTerm
. NN O O

In NN O O
this NN O O
paper NN O O
we NN O O
propose NN O O
a NN O O
logical NN O I-Method
formalism NN O I-Method
, NN O O
which NN O O
, NN O O
among NN O O
other NN O O
things NN O O
, NN O O
is NN O O
suitable NN O O
for NN O O
representing NN O O
determiners NN O I-Task
without NN O O
forcing NN O O
a NN O O
particular NN O O
interpretation NN O O
when NN O O
their NN O O
meaning NN O O
is NN O O
still NN O O
not NN O O
clear NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
investigate NN O O
the NN O O
verbal NN O I-Method
and NN O I-Method
nonverbal NN O I-Method
means NN O I-Method
for NN O O
grounding NN O I-Task
, NN O O
and NN O O
propose NN O O
a NN O O
design NN O I-Generic
for NN O O
embodied NN O I-Method
conversational NN O I-Method
agents NN O I-Method
that NN O O
relies NN O O
on NN O O
both NN O O
kinds NN O O
of NN O O
signals NN O O
to NN O O
establish NN O O
common NN O I-Task
ground NN O I-Task
in NN O O
human-computer NN O I-Task
interaction NN O I-Task
. NN O O

We NN O O
analyzed NN O O
eye NN O I-OtherScientificTerm
gaze NN O I-OtherScientificTerm
, NN O O
head NN O I-OtherScientificTerm
nods NN O I-OtherScientificTerm
and NN O O
attentional NN O I-OtherScientificTerm
focus NN O I-OtherScientificTerm
in NN O O
the NN O O
context NN O O
of NN O O
a NN O O
direction-giving NN O I-Task
task NN O I-Task
. NN O O

The NN O O
distribution NN O O
of NN O O
nonverbal NN O I-OtherScientificTerm
behaviors NN O I-OtherScientificTerm
differed NN O O
depending NN O O
on NN O O
the NN O O
type NN O O
of NN O O
dialogue NN O I-OtherScientificTerm
move NN O I-OtherScientificTerm
being NN O O
grounded NN O O
, NN O O
and NN O O
the NN O O
overall NN O O
pattern NN O O
reflected NN O O
a NN O O
monitoring NN O O
of NN O O
lack NN O O
of NN O O
negative NN O I-OtherScientificTerm
feedback NN O I-OtherScientificTerm
. NN O O

Based NN O O
on NN O O
these NN O O
results NN O O
, NN O O
we NN O O
present NN O O
an NN O O
ECA NN O I-Method
that NN O O
uses NN O O
verbal NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
nonverbal NN O I-OtherScientificTerm
grounding NN O I-OtherScientificTerm
acts NN O I-OtherScientificTerm
to NN O O
update NN O O
dialogue NN O I-OtherScientificTerm
state NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

Sentence NN O I-Task
boundary NN O I-Task
detection NN O I-Task
in NN O O
speech NN O I-Material
is NN O O
important NN O O
for NN O O
enriching NN O O
speech NN O I-OtherScientificTerm
recognition NN O I-OtherScientificTerm
output NN O I-OtherScientificTerm
, NN O O
making NN O O
it NN O I-Generic
easier NN O O
for NN O O
humans NN O O
to NN O O
read NN O O
and NN O O
downstream NN O O
modules NN O O
to NN O O
process NN O O
. NN O O

In NN O O
previous NN O O
work NN O O
, NN O O
we NN O O
have NN O O
developed NN O O
hidden NN O I-Method
Markov NN O I-Method
model NN O I-Method
-LRB- NN O I-Method
HMM NN O I-Method
-RRB- NN O I-Method
and NN O I-Method
maximum NN O I-Method
entropy NN O I-Method
-LRB- NN O I-Method
Maxent NN O I-Method
-RRB- NN O I-Method
classifiers NN O I-Method
that NN O O
integrate NN O O
textual NN O I-Material
and NN O I-Material
prosodic NN O I-Material
knowledge NN O I-Material
sources NN O I-Material
for NN O O
detecting NN O I-Task
sentence NN O I-Task
boundaries NN O I-Task
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
evaluate NN O O
the NN O O
use NN O O
of NN O O
a NN O O
conditional NN O I-Method
random NN O I-Method
field NN O I-Method
-LRB- NN O I-Method
CRF NN O I-Method
-RRB- NN O I-Method
for NN O O
this NN O O
task NN O I-Generic
and NN O O
relate NN O O
results NN O O
with NN O O
this NN O O
model NN O I-Generic
to NN O O
our NN O O
prior NN O O
work NN O O
. NN O O

We NN O O
evaluate NN O O
across NN O O
two NN O O
corpora NN O I-Generic
-LRB- NN O O
conversational NN O I-Material
telephone NN O I-Material
speech NN O I-Material
and NN O O
broadcast NN O I-Material
news NN O I-Material
speech NN O I-Material
-RRB- NN O O
on NN O O
both NN O O
human NN O I-OtherScientificTerm
transcriptions NN O I-OtherScientificTerm
and NN O O
speech NN O I-OtherScientificTerm
recognition NN O I-OtherScientificTerm
output NN O I-OtherScientificTerm
. NN O O

In NN O O
general NN O O
, NN O O
our NN O O
CRF NN O I-Method
model NN O I-Method
yields NN O O
a NN O O
lower NN O O
error NN O I-Metric
rate NN O I-Metric
than NN O O
the NN O O
HMM NN O I-Method
and NN O I-Method
Max-ent NN O I-Method
models NN O I-Method
on NN O O
the NN O O
NIST NN O I-Material
sentence NN O I-Material
boundary NN O I-Material
detection NN O I-Material
task NN O I-Material
in NN O O
speech NN O I-Material
, NN O O
although NN O O
it NN O O
is NN O O
interesting NN O O
to NN O O
note NN O O
that NN O O
the NN O O
best NN O O
results NN O O
are NN O O
achieved NN O O
by NN O O
three-way NN O I-Method
voting NN O I-Method
among NN O O
the NN O O
classifiers NN O I-Method
. NN O O

This NN O O
probably NN O O
occurs NN O O
because NN O O
each NN O O
model NN O I-Generic
has NN O O
different NN O O
strengths NN O O
and NN O O
weaknesses NN O O
for NN O O
modeling NN O O
the NN O O
knowledge NN O I-Material
sources NN O I-Material
. NN O O

-DOCSTART- NN O O

We NN O O
propose NN O O
a NN O O
novel NN O O
approach NN O I-Generic
to NN O O
associate NN O O
objects NN O O
across NN O O
multiple NN O O
PTZ NN O I-OtherScientificTerm
cameras NN O I-OtherScientificTerm
that NN O O
can NN O O
be NN O O
used NN O O
to NN O O
perform NN O O
camera NN O I-Task
handoff NN O I-Task
in NN O I-Task
wide-area NN O I-Task
surveillance NN O I-Task
scenarios NN O I-Task
. NN O O

While NN O O
previous NN O O
approaches NN O I-Generic
relied NN O O
on NN O O
geometric NN O I-OtherScientificTerm
, NN O I-OtherScientificTerm
appearance NN O I-OtherScientificTerm
, NN O I-OtherScientificTerm
or NN O I-OtherScientificTerm
correlation-based NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
for NN O O
establishing NN O O
correspondences NN O O
between NN O O
static NN O I-OtherScientificTerm
cameras NN O I-OtherScientificTerm
, NN O O
they NN O O
each NN O O
have NN O O
well-known NN O O
limitations NN O O
and NN O O
are NN O O
not NN O O
extendable NN O O
to NN O O
wide-area NN O I-OtherScientificTerm
settings NN O I-OtherScientificTerm
with NN O O
PTZ NN O I-OtherScientificTerm
cameras NN O I-OtherScientificTerm
. NN O O

In NN O O
our NN O O
approach NN O I-Generic
, NN O O
the NN O O
slave NN O I-OtherScientificTerm
camera NN O I-OtherScientificTerm
only NN O O
passively NN O O
follows NN O O
the NN O O
target NN O O
-LRB- NN O O
by NN O O
loose NN O O
registration NN O O
with NN O O
the NN O O
master NN O O
-RRB- NN O O
and NN O O
bootstraps NN O O
itself NN O O
from NN O O
its NN O O
own NN O O
incoming NN O O
imagery NN O O
, NN O O
thus NN O O
effectively NN O O
circumventing NN O O
the NN O O
problems NN O O
faced NN O O
by NN O O
previous NN O O
approaches NN O I-Generic
and NN O O
avoiding NN O O
the NN O O
need NN O O
to NN O O
perform NN O O
any NN O O
model NN O I-OtherScientificTerm
transfer NN O I-OtherScientificTerm
. NN O O

Towards NN O O
this NN O O
goal NN O O
, NN O O
we NN O O
also NN O O
propose NN O O
a NN O O
novel NN O O
Multiple NN O I-Method
Instance NN O I-Method
Learning NN O I-Method
-LRB- NN O I-Method
MIL NN O I-Method
-RRB- NN O I-Method
formulation NN O I-Method
for NN O O
the NN O O
problem NN O O
based NN O O
on NN O O
the NN O O
logistic NN O I-OtherScientificTerm
softmax NN O I-OtherScientificTerm
function NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
covariance-based NN O I-OtherScientificTerm
region NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
within NN O O
a NN O O
MAP NN O I-Method
estimation NN O I-Method
framework NN O I-Method
. NN O O

We NN O O
demonstrate NN O O
our NN O O
approach NN O I-Generic
with NN O O
multiple NN O I-OtherScientificTerm
PTZ NN O I-OtherScientificTerm
camera NN O I-OtherScientificTerm
sequences NN O I-OtherScientificTerm
in NN O O
typical NN O O
outdoor NN O I-OtherScientificTerm
surveillance NN O I-OtherScientificTerm
settings NN O I-OtherScientificTerm
and NN O O
show NN O O
a NN O O
comparison NN O O
with NN O O
state-of-the-art NN O I-Generic
approaches NN O I-Generic
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
solves NN O O
a NN O O
specialized NN O I-Task
regression NN O I-Task
problem NN O I-Task
to NN O O
obtain NN O O
sampling NN O I-OtherScientificTerm
probabilities NN O I-OtherScientificTerm
for NN O O
records NN O I-Material
in NN O O
databases NN O I-Material
. NN O O

The NN O O
goal NN O O
is NN O O
to NN O O
sample NN O O
a NN O O
small NN O O
set NN O O
of NN O O
records NN O I-Material
over NN O O
which NN O O
evaluating NN O O
aggregate NN O I-OtherScientificTerm
queries NN O I-OtherScientificTerm
can NN O O
be NN O O
done NN O O
both NN O O
efficiently NN O O
and NN O O
accurately NN O O
. NN O O

We NN O O
provide NN O O
a NN O O
principled NN O I-Method
and NN O I-Method
provable NN O I-Method
solution NN O I-Method
for NN O O
this NN O O
problem NN O I-Generic
; NN O O
it NN O I-Generic
is NN O O
parameterless NN O O
and NN O O
requires NN O O
no NN O O
data NN O O
insights NN O O
. NN O O

Unlike NN O O
standard NN O O
regression NN O I-Task
problems NN O I-Task
, NN O O
the NN O O
loss NN O I-OtherScientificTerm
is NN O O
inversely NN O O
proportional NN O O
to NN O O
the NN O O
regressed-to NN O I-OtherScientificTerm
values NN O I-OtherScientificTerm
. NN O O

Moreover NN O O
, NN O O
a NN O O
cost NN O I-Method
zero NN O I-Method
solution NN O I-Method
always NN O O
exists NN O O
and NN O O
can NN O O
only NN O O
be NN O O
excluded NN O O
by NN O O
hard NN O I-OtherScientificTerm
budget NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
. NN O O

A NN O O
unique NN O O
form NN O O
of NN O O
reg-ularization NN O I-OtherScientificTerm
is NN O O
also NN O O
needed NN O O
. NN O O

We NN O O
provide NN O O
an NN O O
efficient NN O O
and NN O O
simple NN O O
regularized NN O I-Method
Empirical NN O I-Method
Risk NN O I-Method
Minimization NN O I-Method
-LRB- NN O I-Method
ERM NN O I-Method
-RRB- NN O I-Method
algorithm NN O I-Method
along NN O O
with NN O O
a NN O O
theoretical NN O O
generalization NN O O
result NN O O
. NN O O

Our NN O O
extensive NN O O
experimental NN O O
results NN O O
significantly NN O O
improve NN O O
over NN O O
both NN O O
uniform NN O I-Method
sampling NN O I-Method
and NN O O
standard NN O O
stratified NN O I-Method
sampling NN O I-Method
which NN O O
are NN O O
de-facto NN O O
the NN O O
industry NN O O
standards NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
consider NN O O
the NN O O
problem NN O O
of NN O O
computing NN O O
the NN O O
Kullback-Leibler NN O I-Method
distance NN O I-Method
, NN O O
also NN O O
called NN O O
the NN O O
relative NN O I-Method
entropy NN O I-Method
, NN O O
between NN O O
a NN O O
probabilistic NN O I-Method
context-free NN O I-Method
grammar NN O I-Method
and NN O O
a NN O O
probabilistic NN O I-OtherScientificTerm
finite NN O I-OtherScientificTerm
automaton NN O I-OtherScientificTerm
. NN O O

We NN O O
show NN O O
that NN O O
there NN O O
is NN O O
a NN O O
closed-form NN O I-Method
-LRB- NN O I-Method
analytical NN O I-Method
-RRB- NN O I-Method
solution NN O I-Method
for NN O O
one NN O O
part NN O O
of NN O O
the NN O O
Kullback-Leibler NN O I-Method
distance NN O I-Method
, NN O O
viz NN O O
the NN O O
cross-entropy NN O I-Method
. NN O O

We NN O O
discuss NN O O
several NN O O
applications NN O O
of NN O O
the NN O O
result NN O O
to NN O O
the NN O O
problem NN O O
of NN O O
distributional NN O I-Task
approximation NN O I-Task
of NN O O
probabilistic NN O I-Method
context-free NN O I-Method
grammars NN O I-Method
by NN O O
means NN O O
of NN O O
probabilistic NN O I-Method
finite NN O I-Method
automata NN O I-Method
. NN O O

-DOCSTART- NN O O

In NN O O
spite NN O O
of NN O O
over NN O O
two NN O O
decades NN O O
of NN O O
intense NN O O
research NN O O
, NN O O
illumination NN O I-Task
and NN O O
pose NN O I-Task
invariance NN O I-Task
remain NN O O
prohibitively NN O O
challenging NN O O
aspects NN O O
of NN O O
face NN O I-Task
recognition NN O I-Task
for NN O O
most NN O O
practical NN O O
applications NN O O
. NN O O

The NN O O
objective NN O O
of NN O O
this NN O O
work NN O O
is NN O O
to NN O O
recognize NN O O
faces NN O O
using NN O O
video NN O I-Material
sequences NN O I-Material
both NN O O
for NN O O
training NN O O
and NN O O
recognition NN O O
input NN O O
, NN O O
in NN O O
a NN O O
realistic NN O O
, NN O O
unconstrained NN O O
setup NN O O
in NN O O
which NN O O
lighting NN O I-OtherScientificTerm
, NN O O
pose NN O I-OtherScientificTerm
and NN O O
user NN O I-OtherScientificTerm
motion NN O I-OtherScientificTerm
pattern NN O I-OtherScientificTerm
have NN O O
a NN O O
wide NN O O
variability NN O O
and NN O O
face NN O I-Material
images NN O I-Material
are NN O O
of NN O O
low NN O O
resolution NN O I-Metric
. NN O O

In NN O O
particular NN O O
there NN O O
are NN O O
three NN O O
areas NN O O
of NN O O
novelty NN O O
: NN O O
-LRB- NN O O
i NN O O
-RRB- NN O O
we NN O O
show NN O O
how NN O O
a NN O O
photometric NN O I-Method
model NN O I-Method
of NN O O
image NN O I-Task
formation NN O I-Task
can NN O O
be NN O O
combined NN O O
with NN O O
a NN O O
statistical NN O I-Method
model NN O I-Method
of NN O O
generic NN O I-Task
face NN O I-Task
appearance NN O I-Task
variation NN O I-Task
, NN O O
learnt NN O O
offline NN O O
, NN O O
to NN O O
generalize NN O O
in NN O O
the NN O O
presence NN O O
of NN O O
extreme NN O I-OtherScientificTerm
illumination NN O I-OtherScientificTerm
changes NN O I-OtherScientificTerm
; NN O O
-LRB- NN O O
ii NN O O
-RRB- NN O O
we NN O O
use NN O O
the NN O O
smoothness NN O I-OtherScientificTerm
of NN O O
geodesically NN O I-OtherScientificTerm
local NN O I-OtherScientificTerm
appearance NN O I-OtherScientificTerm
manifold NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
and NN O O
a NN O O
robust NN O I-Method
same-identity NN O I-Method
likelihood NN O I-Method
to NN O O
achieve NN O O
invariance NN O O
to NN O O
unseen NN O I-OtherScientificTerm
head NN O I-OtherScientificTerm
poses NN O I-OtherScientificTerm
; NN O O
and NN O O
-LRB- NN O O
iii NN O O
-RRB- NN O O
we NN O O
introduce NN O O
an NN O O
accurate NN O O
video NN O I-Method
sequence NN O I-Method
'' NN O I-Method
reillumination NN O I-Method
'' NN O I-Method
algorithm NN O I-Method
to NN O O
achieve NN O O
robustness NN O I-Metric
to NN O O
face NN O I-OtherScientificTerm
motion NN O I-OtherScientificTerm
patterns NN O I-OtherScientificTerm
in NN O O
video NN O I-Material
. NN O O

We NN O O
describe NN O O
a NN O O
fully NN O I-Method
automatic NN O I-Method
recognition NN O I-Method
system NN O I-Method
based NN O O
on NN O O
the NN O O
proposed NN O O
method NN O I-Generic
and NN O O
an NN O O
extensive NN O O
evaluation NN O O
on NN O O
171 NN O O
individuals NN O O
and NN O O
over NN O O
1300 NN O O
video NN O I-Material
sequences NN O I-Material
with NN O O
extreme NN O O
illumination NN O I-OtherScientificTerm
, NN O O
pose NN O I-OtherScientificTerm
and NN O O
head NN O I-OtherScientificTerm
motion NN O I-OtherScientificTerm
variation NN O I-OtherScientificTerm
. NN O O

On NN O O
this NN O O
challenging NN O O
data NN O I-Generic
set NN O I-Generic
our NN O O
system NN O I-Generic
consistently NN O O
demonstrated NN O O
a NN O O
nearly NN O O
perfect NN O O
recognition NN O I-Metric
rate NN O I-Metric
-LRB- NN O O
over NN O O
99.7 NN O O
% NN O O
on NN O O
all NN O O
three NN O O
databases NN O I-Material
-RRB- NN O O
, NN O O
significantly NN O O
out-performing NN O O
state-of-the-art NN O O
commercial NN O I-Method
software NN O I-Method
and NN O O
methods NN O I-Generic
from NN O O
the NN O O
literature NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
present NN O O
Minimum NN O I-Method
Bayes-Risk NN O I-Method
-LRB- NN O I-Method
MBR NN O I-Method
-RRB- NN O I-Method
decoding NN O I-Method
for NN O O
statistical NN O I-Task
machine NN O I-Task
translation NN O I-Task
. NN O O

This NN O O
statistical NN O I-Method
approach NN O I-Method
aims NN O O
to NN O O
minimize NN O O
expected NN O I-OtherScientificTerm
loss NN O I-OtherScientificTerm
of NN O O
translation NN O I-OtherScientificTerm
errors NN O I-OtherScientificTerm
under NN O O
loss NN O I-OtherScientificTerm
functions NN O I-OtherScientificTerm
that NN O O
measure NN O O
translation NN O I-Task
performance NN O O
. NN O O

We NN O O
describe NN O O
a NN O O
hierarchy NN O O
of NN O O
loss NN O I-OtherScientificTerm
functions NN O I-OtherScientificTerm
that NN O O
incorporate NN O O
different NN O O
levels NN O O
of NN O O
linguistic NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
from NN O O
word NN O O
strings NN O O
, NN O O
word-to-word NN O I-OtherScientificTerm
alignments NN O I-OtherScientificTerm
from NN O O
an NN O O
MT NN O I-Method
system NN O I-Method
, NN O O
and NN O O
syntactic NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
from NN O O
parse-trees NN O I-OtherScientificTerm
of NN O O
source NN O O
and NN O O
target NN O O
language NN O O
sentences NN O O
. NN O O

We NN O O
report NN O O
the NN O O
performance NN O O
of NN O O
the NN O O
MBR NN O I-Method
decoders NN O I-Method
on NN O O
a NN O O
Chinese-to-English NN O I-Task
translation NN O I-Task
task NN O I-Task
. NN O O

Our NN O O
results NN O O
show NN O O
that NN O O
MBR NN O I-Method
decoding NN O I-Method
can NN O O
be NN O O
used NN O O
to NN O O
tune NN O O
statistical NN O I-Method
MT NN O I-Method
performance NN O O
for NN O O
specific NN O O
loss NN O I-OtherScientificTerm
functions NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
presents NN O O
a NN O O
critical NN O O
discussion NN O O
of NN O O
the NN O O
various NN O O
approaches NN O I-Generic
that NN O O
have NN O O
been NN O O
used NN O O
in NN O O
the NN O O
evaluation NN O I-Task
of NN O I-Task
Natural NN O I-Method
Language NN O I-Method
systems NN O I-Method
. NN O O

We NN O O
conclude NN O O
that NN O O
previous NN O O
approaches NN O I-Generic
have NN O O
neglected NN O O
to NN O O
evaluate NN O O
systems NN O I-Generic
in NN O O
the NN O O
context NN O O
of NN O O
their NN O O
use NN O O
, NN O O
e.g. NN O O
solving NN O O
a NN O O
task NN O I-Generic
requiring NN O O
data NN O I-Task
retrieval NN O I-Task
. NN O O

This NN O O
raises NN O O
questions NN O O
about NN O O
the NN O O
validity NN O O
of NN O O
such NN O O
approaches NN O I-Generic
. NN O O

In NN O O
the NN O O
second NN O O
half NN O O
of NN O O
the NN O O
paper NN O O
, NN O O
we NN O O
report NN O O
a NN O O
laboratory NN O O
study NN O O
using NN O O
the NN O O
Wizard NN O I-Method
of NN O I-Method
Oz NN O I-Method
technique NN O I-Method
to NN O O
identify NN O O
NL NN O I-OtherScientificTerm
requirements NN O I-OtherScientificTerm
for NN O O
carrying NN O O
out NN O O
this NN O O
task NN O I-Generic
. NN O O

We NN O O
evaluate NN O O
the NN O O
demands NN O O
that NN O O
task NN O I-Material
dialogues NN O I-Material
collected NN O O
using NN O O
this NN O O
technique NN O I-Generic
, NN O O
place NN O O
upon NN O O
a NN O O
prototype NN O I-Method
Natural NN O I-Method
Language NN O I-Method
system NN O I-Method
. NN O O

We NN O O
identify NN O O
three NN O O
important NN O O
requirements NN O O
which NN O O
arose NN O O
from NN O O
the NN O O
task NN O I-Generic
that NN O O
we NN O O
gave NN O O
our NN O O
subjects NN O O
: NN O O
operators NN O O
specific NN O O
to NN O O
the NN O O
task NN O I-Generic
of NN O O
database NN O I-Task
access NN O I-Task
, NN O O
complex NN O O
contextual NN O I-OtherScientificTerm
reference NN O I-OtherScientificTerm
and NN O O
reference NN O O
to NN O O
the NN O O
structure NN O O
of NN O O
the NN O O
information NN O O
source NN O O
. NN O O

We NN O O
discuss NN O O
how NN O O
these NN O O
might NN O O
be NN O O
satisfied NN O O
by NN O O
future NN O O
Natural NN O I-Method
Language NN O I-Method
systems NN O I-Method
. NN O O

-DOCSTART- NN O O

We NN O O
present NN O O
results NN O O
on NN O O
addressee NN O I-Task
identification NN O I-Task
in NN O I-Task
four-participants NN O I-Task
face-to-face NN O I-Task
meetings NN O I-Task
using NN O O
Bayesian NN O I-Method
Network NN O I-Method
and NN O O
Naive NN O I-Method
Bayes NN O I-Method
classifiers NN O I-Method
. NN O O

First NN O O
, NN O O
we NN O O
investigate NN O O
how NN O O
well NN O O
the NN O O
addressee NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
a NN O I-OtherScientificTerm
dialogue NN O I-OtherScientificTerm
act NN O I-OtherScientificTerm
can NN O O
be NN O O
predicted NN O O
based NN O O
on NN O O
gaze NN O I-OtherScientificTerm
, NN O O
utterance NN O I-OtherScientificTerm
and NN O O
conversational NN O I-OtherScientificTerm
context NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
. NN O O

Then NN O O
, NN O O
we NN O O
explore NN O O
whether NN O O
information NN O O
about NN O O
meeting NN O I-OtherScientificTerm
context NN O I-OtherScientificTerm
can NN O O
aid NN O O
classifiers NN O I-Method
' NN O O
performances NN O O
. NN O O

Both NN O O
classifiers NN O I-Method
perform NN O O
the NN O O
best NN O O
when NN O O
conversational NN O I-OtherScientificTerm
context NN O I-OtherScientificTerm
and NN O O
utterance NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
are NN O O
combined NN O O
with NN O O
speaker NN O I-OtherScientificTerm
's NN O I-OtherScientificTerm
gaze NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
. NN O O

The NN O O
classifiers NN O I-Method
show NN O O
little NN O O
gain NN O O
from NN O O
information NN O O
about NN O O
meeting NN O I-OtherScientificTerm
context NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

Towards NN O O
deep NN O O
analysis NN O O
of NN O O
compositional NN O I-Task
classes NN O I-Task
of NN O I-Task
paraphrases NN O I-Task
, NN O O
we NN O O
have NN O O
examined NN O O
a NN O O
class-oriented NN O I-Method
framework NN O I-Method
for NN O O
collecting NN O O
paraphrase NN O I-Material
examples NN O I-Material
, NN O O
in NN O O
which NN O O
sentential NN O I-Material
paraphrases NN O I-Material
are NN O O
collected NN O O
for NN O O
each NN O O
paraphrase NN O O
class NN O O
separately NN O O
by NN O O
means NN O O
of NN O O
automatic NN O I-Method
candidate NN O I-Method
generation NN O I-Method
and NN O O
manual NN O I-Method
judgement NN O I-Method
. NN O O

Our NN O O
preliminary NN O O
experiments NN O O
on NN O O
building NN O O
a NN O O
paraphrase NN O I-Material
corpus NN O I-Material
have NN O O
so NN O O
far NN O O
been NN O O
producing NN O O
promising NN O O
results NN O O
, NN O O
which NN O O
we NN O O
have NN O O
evaluated NN O O
according NN O O
to NN O O
cost-efficiency NN O O
, NN O O
exhaustiveness NN O O
, NN O O
and NN O O
reliability NN O O
. NN O O

-DOCSTART- NN O O

The NN O O
purpose NN O O
of NN O O
this NN O O
research NN O O
is NN O O
to NN O O
test NN O O
the NN O O
efficacy NN O O
of NN O O
applying NN O O
automated NN O I-Method
evaluation NN O I-Method
techniques NN O I-Method
, NN O O
originally NN O O
devised NN O O
for NN O O
the NN O O
evaluation NN O I-Task
of NN O I-Task
human NN O I-Task
language NN O I-Task
learners NN O I-Task
, NN O O
to NN O O
the NN O O
output NN O O
of NN O O
machine NN O I-Method
translation NN O I-Method
-LRB- NN O I-Method
MT NN O I-Method
-RRB- NN O I-Method
systems NN O I-Method
. NN O O

We NN O O
believe NN O O
that NN O O
these NN O O
evaluation NN O I-Generic
techniques NN O I-Generic
will NN O O
provide NN O O
information NN O O
about NN O O
both NN O O
the NN O O
human NN O I-Task
language NN O I-Task
learning NN O I-Task
process NN O I-Task
, NN O O
the NN O O
translation NN O I-Task
process NN O I-Task
and NN O O
the NN O O
development NN O O
of NN O O
machine NN O I-Task
translation NN O I-Task
systems NN O I-Task
. NN O O

This NN O O
, NN O O
the NN O O
first NN O O
experiment NN O O
in NN O O
a NN O O
series NN O O
of NN O O
experiments NN O O
, NN O O
looks NN O O
at NN O O
the NN O O
intelligibility NN O O
of NN O O
MT NN O O
output NN O O
. NN O O

A NN O O
language NN O I-Task
learning NN O I-Task
experiment NN O O
showed NN O O
that NN O O
assessors NN O I-Generic
can NN O O
differentiate NN O O
native NN O O
from NN O O
non-native NN O I-Material
language NN O I-Material
essays NN O I-Material
in NN O O
less NN O O
than NN O O
100 NN O O
words NN O O
. NN O O

Even NN O O
more NN O O
illuminating NN O O
was NN O O
the NN O O
factors NN O O
on NN O O
which NN O O
the NN O O
assessors NN O I-Generic
made NN O O
their NN O O
decisions NN O O
. NN O O

We NN O O
tested NN O O
this NN O O
to NN O O
see NN O O
if NN O O
similar NN O O
criteria NN O O
could NN O O
be NN O O
elicited NN O O
from NN O O
duplicating NN O O
the NN O O
experiment NN O O
using NN O O
machine NN O I-OtherScientificTerm
translation NN O I-OtherScientificTerm
output NN O I-OtherScientificTerm
. NN O O

Subjects NN O O
were NN O O
given NN O O
a NN O O
set NN O O
of NN O O
up NN O O
to NN O O
six NN O O
extracts NN O O
of NN O O
translated NN O I-Material
newswire NN O I-Material
text NN O I-Material
. NN O O

Some NN O O
of NN O O
the NN O O
extracts NN O O
were NN O O
expert NN O I-OtherScientificTerm
human NN O I-OtherScientificTerm
translations NN O I-OtherScientificTerm
, NN O O
others NN O O
were NN O O
machine NN O I-OtherScientificTerm
translation NN O I-OtherScientificTerm
outputs NN O I-OtherScientificTerm
. NN O O

The NN O O
subjects NN O O
were NN O O
given NN O O
three NN O O
minutes NN O O
per NN O O
extract NN O O
to NN O O
determine NN O O
whether NN O O
they NN O O
believed NN O O
the NN O O
sample NN O O
output NN O O
to NN O O
be NN O O
an NN O O
expert NN O I-OtherScientificTerm
human NN O I-OtherScientificTerm
translation NN O I-OtherScientificTerm
or NN O O
a NN O O
machine NN O I-OtherScientificTerm
translation NN O I-OtherScientificTerm
. NN O O

Additionally NN O O
, NN O O
they NN O O
were NN O O
asked NN O O
to NN O O
mark NN O O
the NN O O
word NN O O
at NN O O
which NN O O
they NN O O
made NN O O
this NN O O
decision NN O O
. NN O O

The NN O O
results NN O O
of NN O O
this NN O O
experiment NN O O
, NN O O
along NN O O
with NN O O
a NN O O
preliminary NN O O
analysis NN O O
of NN O O
the NN O O
factors NN O O
involved NN O O
in NN O O
the NN O O
decision NN O O
making NN O O
process NN O O
will NN O O
be NN O O
presented NN O O
here NN O O
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
presents NN O O
a NN O O
machine NN O I-Method
learning NN O I-Method
approach NN O I-Method
to NN O O
bare NN O I-Task
slice NN O I-Task
disambiguation NN O I-Task
in NN O O
dialogue NN O I-Material
. NN O O

We NN O O
extract NN O O
a NN O O
set NN O O
of NN O O
heuristic NN O I-Method
principles NN O I-Method
from NN O O
a NN O O
corpus-based NN O I-Material
sample NN O I-Material
and NN O O
formulate NN O O
them NN O O
as NN O O
probabilistic NN O I-OtherScientificTerm
Horn NN O I-OtherScientificTerm
clauses NN O I-OtherScientificTerm
. NN O O

We NN O O
then NN O O
use NN O O
the NN O O
predicates NN O O
of NN O O
such NN O O
clauses NN O I-OtherScientificTerm
to NN O O
create NN O O
a NN O O
set NN O O
of NN O O
domain NN O I-OtherScientificTerm
independent NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
to NN O O
annotate NN O O
an NN O O
input NN O O
dataset NN O O
, NN O O
and NN O O
run NN O O
two NN O O
different NN O O
machine NN O I-Method
learning NN O I-Method
algorithms NN O I-Method
: NN O O
SLIPPER NN O I-Generic
, NN O O
a NN O O
rule-based NN O I-Method
learning NN O I-Method
algorithm NN O I-Method
, NN O O
and NN O O
TiMBL NN O I-Generic
, NN O O
a NN O O
memory-based NN O I-Method
system NN O I-Method
. NN O O

Both NN O O
learners NN O O
perform NN O O
well NN O O
, NN O O
yielding NN O O
similar NN O O
success NN O I-Metric
rates NN O I-Metric
of NN O O
approx NN O O
90 NN O O
% NN O O
. NN O O

The NN O O
results NN O O
show NN O O
that NN O O
the NN O O
features NN O I-OtherScientificTerm
in NN O O
terms NN O O
of NN O O
which NN O O
we NN O O
formulate NN O O
our NN O O
heuristic NN O I-Method
principles NN O I-Method
have NN O O
significant NN O O
predictive NN O O
power NN O O
, NN O O
and NN O O
that NN O O
rules NN O I-OtherScientificTerm
that NN O O
closely NN O O
resemble NN O O
our NN O O
Horn NN O I-OtherScientificTerm
clauses NN O I-OtherScientificTerm
can NN O O
be NN O O
learnt NN O O
automatically NN O O
from NN O O
these NN O O
features NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

We NN O O
suggest NN O O
a NN O O
new NN O O
goal NN O O
and NN O O
evaluation NN O I-Metric
criterion NN O I-Metric
for NN O O
word NN O I-Metric
similarity NN O I-Metric
measures NN O I-Metric
. NN O O

The NN O O
new NN O O
criterion NN O I-Generic
-- NN O O
meaning-entailing NN O I-Metric
substitutability NN O I-Metric
-- NN O O
fits NN O O
the NN O O
needs NN O O
of NN O O
semantic-oriented NN O I-Task
NLP NN O I-Task
applications NN O I-Task
and NN O O
can NN O O
be NN O O
evaluated NN O O
directly NN O O
-LRB- NN O O
independent NN O O
of NN O O
an NN O O
application NN O O
-RRB- NN O O
at NN O O
a NN O O
good NN O O
level NN O O
of NN O O
human NN O I-Metric
agreement NN O I-Metric
. NN O O

Motivated NN O O
by NN O O
this NN O O
semantic NN O I-Metric
criterion NN O I-Metric
we NN O O
analyze NN O O
the NN O O
empirical NN O O
quality NN O O
of NN O O
distributional NN O I-OtherScientificTerm
word NN O I-OtherScientificTerm
feature NN O I-OtherScientificTerm
vectors NN O I-OtherScientificTerm
and NN O O
its NN O O
impact NN O O
on NN O O
word NN O I-Task
similarity NN O I-Task
results NN O O
, NN O O
proposing NN O O
an NN O O
objective NN O O
measure NN O I-Generic
for NN O O
evaluating NN O O
feature NN O I-Metric
vector NN O I-Metric
quality NN O I-Metric
. NN O O

Finally NN O O
, NN O O
a NN O O
novel NN O O
feature NN O I-Method
weighting NN O I-Method
and NN O I-Method
selection NN O I-Method
function NN O I-Method
is NN O O
presented NN O O
, NN O O
which NN O O
yields NN O O
superior NN O O
feature NN O I-OtherScientificTerm
vectors NN O I-OtherScientificTerm
and NN O O
better NN O O
word NN O I-Task
similarity NN O I-Task
performance NN O O
. NN O O

-DOCSTART- NN O O

Reflections NN O I-OtherScientificTerm
in NN O I-OtherScientificTerm
image NN O I-Material
sequences NN O I-Material
consist NN O O
of NN O O
several NN O O
layers NN O O
superimposed NN O O
over NN O O
each NN O O
other NN O O
. NN O O

This NN O O
phenomenon NN O O
causes NN O O
many NN O O
image NN O I-Method
processing NN O I-Method
techniques NN O I-Method
to NN O O
fail NN O O
as NN O O
they NN O O
assume NN O O
the NN O O
presence NN O O
of NN O O
only NN O O
one NN O O
layer NN O O
at NN O O
each NN O O
examined NN O O
site NN O O
e.g. NN O O
motion NN O I-Task
estimation NN O I-Task
and NN O O
object NN O I-Task
recognition NN O I-Task
. NN O O

This NN O O
work NN O O
presents NN O O
an NN O O
automated NN O O
technique NN O I-Generic
for NN O O
detecting NN O I-Task
reflections NN O I-Task
in NN O I-Task
image NN O I-Task
sequences NN O I-Task
by NN O O
analyzing NN O O
motion NN O I-OtherScientificTerm
trajectories NN O I-OtherScientificTerm
of NN O O
feature NN O I-OtherScientificTerm
points NN O I-OtherScientificTerm
. NN O O

It NN O I-Generic
models NN O O
reflection NN O I-OtherScientificTerm
as NN O O
regions NN O O
containing NN O O
two NN O O
different NN O O
layers NN O O
moving NN O O
over NN O O
each NN O O
other NN O O
. NN O O

We NN O O
present NN O O
a NN O O
strong NN O O
detector NN O I-Generic
based NN O O
on NN O O
combining NN O O
a NN O O
set NN O O
of NN O O
weak NN O O
detectors NN O I-Generic
. NN O O

We NN O O
use NN O O
novel NN O O
priors NN O I-OtherScientificTerm
, NN O O
generate NN O O
sparse NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
dense NN O I-OtherScientificTerm
detection NN O I-OtherScientificTerm
maps NN O I-OtherScientificTerm
and NN O O
our NN O O
results NN O O
show NN O O
high NN O O
detection NN O I-Metric
rate NN O I-Metric
with NN O O
rejection NN O O
to NN O O
pathological NN O I-OtherScientificTerm
motion NN O I-OtherScientificTerm
and NN O O
occlusion NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
considers NN O O
the NN O O
problem NN O O
of NN O O
reconstructing NN O I-Task
the NN O I-Task
motion NN O I-Task
of NN O I-Task
a NN O I-Task
3D NN O I-Task
articulated NN O I-Task
tree NN O I-Task
from NN O O
2D NN O I-OtherScientificTerm
point NN O I-OtherScientificTerm
correspondences NN O I-OtherScientificTerm
subject NN O O
to NN O O
some NN O O
temporal NN O I-OtherScientificTerm
prior NN O I-OtherScientificTerm
. NN O O

Hitherto NN O O
, NN O O
smooth NN O I-OtherScientificTerm
motion NN O I-OtherScientificTerm
has NN O O
been NN O O
encouraged NN O O
using NN O O
a NN O O
trajectory NN O I-OtherScientificTerm
basis NN O I-OtherScientificTerm
, NN O O
yielding NN O O
a NN O O
hard NN O I-Task
combinatorial NN O I-Task
problem NN O I-Task
with NN O O
time NN O I-Metric
complexity NN O I-Metric
growing NN O O
exponentially NN O O
in NN O O
the NN O O
number NN O O
of NN O O
frames NN O O
. NN O O

Branch NN O I-Method
and NN O I-Method
bound NN O I-Method
strategies NN O I-Method
have NN O O
previously NN O O
attempted NN O O
to NN O O
curb NN O O
this NN O O
complexity NN O I-Generic
whilst NN O O
maintaining NN O O
global NN O I-OtherScientificTerm
optimality NN O I-OtherScientificTerm
. NN O O

However NN O O
, NN O O
they NN O I-Generic
provide NN O O
no NN O O
guarantee NN O O
of NN O O
being NN O O
more NN O O
efficient NN O O
than NN O O
exhaustive NN O I-Method
search NN O I-Method
. NN O O

Inspired NN O O
by NN O O
recent NN O O
work NN O O
which NN O O
reconstructs NN O O
general NN O O
trajectories NN O O
using NN O O
compact NN O I-Method
high-pass NN O I-Method
filters NN O I-Method
, NN O O
we NN O O
develop NN O O
a NN O O
dynamic NN O I-Method
programming NN O I-Method
approach NN O I-Method
which NN O O
scales NN O O
linearly NN O O
in NN O O
the NN O O
number NN O O
of NN O O
frames NN O O
, NN O O
leveraging NN O O
the NN O O
intrinsically NN O O
local NN O O
nature NN O O
of NN O O
filter NN O I-OtherScientificTerm
interactions NN O I-OtherScientificTerm
. NN O O

Extension NN O O
to NN O O
affine NN O I-Method
projection NN O I-Method
enables NN O O
reconstruction NN O I-Task
without NN O O
estimating NN O I-Method
cameras NN O I-Method
. NN O O

-DOCSTART- NN O O

Topical NN O I-Task
blog NN O I-Task
post NN O I-Task
retrieval NN O I-Task
is NN O O
the NN O O
task NN O O
of NN O O
ranking NN O I-Task
blog NN O I-Material
posts NN O I-Material
with NN O O
respect NN O O
to NN O O
their NN O O
relevance NN O I-Metric
for NN O O
a NN O O
given NN O O
topic NN O O
. NN O O

To NN O O
improve NN O O
topical NN O I-Task
blog NN O I-Task
post NN O I-Task
retrieval NN O I-Task
we NN O O
incorporate NN O O
textual NN O I-OtherScientificTerm
credibility NN O I-OtherScientificTerm
indicators NN O I-OtherScientificTerm
in NN O O
the NN O O
retrieval NN O I-Method
process NN O I-Method
. NN O O

We NN O O
consider NN O O
two NN O O
groups NN O O
of NN O O
indicators NN O I-Generic
: NN O O
post NN O O
level NN O O
-LRB- NN O O
determined NN O O
using NN O O
information NN O O
about NN O O
individual NN O O
blog NN O I-Material
posts NN O I-Material
only NN O O
-RRB- NN O O
and NN O O
blog NN O O
level NN O O
-LRB- NN O O
determined NN O O
using NN O O
information NN O O
from NN O O
the NN O O
underlying NN O O
blogs NN O I-Material
-RRB- NN O O
. NN O O

We NN O O
describe NN O O
how NN O O
to NN O O
estimate NN O O
these NN O O
indicators NN O I-Generic
and NN O O
how NN O O
to NN O O
integrate NN O O
them NN O I-Generic
into NN O O
a NN O O
retrieval NN O I-Method
approach NN O I-Method
based NN O O
on NN O O
language NN O I-Method
models NN O I-Method
. NN O O

Experiments NN O O
on NN O O
the NN O O
TREC NN O I-Material
Blog NN O I-Material
track NN O I-Material
test NN O I-Material
set NN O I-Material
show NN O O
that NN O O
both NN O O
groups NN O O
of NN O O
credibility NN O I-OtherScientificTerm
indicators NN O I-OtherScientificTerm
significantly NN O O
improve NN O O
retrieval NN O I-Metric
effectiveness NN O I-Metric
; NN O O
the NN O O
best NN O O
performance NN O O
is NN O O
achieved NN O O
when NN O O
combining NN O O
them NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
investigate NN O O
the NN O O
problem NN O O
of NN O O
learning NN O O
to NN O O
predict NN O O
moves NN O O
in NN O O
the NN O O
board NN O I-Task
game NN O I-Task
of NN O I-Task
Go NN O I-Task
from NN O O
game NN O I-Material
records NN O I-Material
of NN O I-Material
expert NN O I-Material
players NN O I-Material
. NN O O

In NN O O
particular NN O O
, NN O O
we NN O O
obtain NN O O
a NN O O
probability NN O I-OtherScientificTerm
distribution NN O I-OtherScientificTerm
over NN O O
legal NN O O
moves NN O O
for NN O O
professional NN O O
play NN O O
in NN O O
a NN O O
given NN O O
position NN O O
. NN O O

This NN O O
distribution NN O I-Generic
has NN O O
numerous NN O O
applications NN O O
in NN O O
computer NN O I-OtherScientificTerm
Go NN O I-OtherScientificTerm
, NN O O
including NN O O
serving NN O O
as NN O O
an NN O O
efficient NN O O
stand-alone NN O I-OtherScientificTerm
Go NN O I-OtherScientificTerm
player NN O I-OtherScientificTerm
. NN O O

It NN O I-Generic
would NN O O
also NN O O
be NN O O
effective NN O O
as NN O O
a NN O O
move NN O I-Method
selector NN O I-Method
and NN O O
move NN O I-Method
sorter NN O I-Method
for NN O O
game NN O I-Method
tree NN O I-Method
search NN O I-Method
and NN O O
as NN O O
a NN O O
training NN O I-Task
tool NN O I-Task
for NN O O
Go NN O I-OtherScientificTerm
players NN O I-OtherScientificTerm
. NN O O

Our NN O O
method NN O I-Generic
has NN O O
two NN O O
major NN O O
components NN O O
: NN O O
a NN O O
-RRB- NN O O
a NN O O
pattern NN O I-Method
extraction NN O I-Method
scheme NN O I-Method
for NN O O
efficiently NN O O
harvesting NN O O
patterns NN O O
of NN O O
given NN O O
size NN O O
and NN O O
shape NN O O
from NN O O
expert NN O O
game NN O O
records NN O O
and NN O O
b NN O O
-RRB- NN O O
a NN O O
Bayesian NN O I-Method
learning NN O I-Method
algorithm NN O I-Method
-LRB- NN O O
in NN O O
two NN O O
variants NN O O
-RRB- NN O O
that NN O O
learns NN O O
a NN O O
distribution NN O O
over NN O O
the NN O O
values NN O O
of NN O O
a NN O O
move NN O O
given NN O O
a NN O O
board NN O O
position NN O O
based NN O O
on NN O O
the NN O O
local NN O I-OtherScientificTerm
pattern NN O I-OtherScientificTerm
context NN O I-OtherScientificTerm
. NN O O

The NN O O
system NN O I-Generic
is NN O O
trained NN O O
on NN O O
181,000 NN O O
expert NN O I-Material
games NN O I-Material
and NN O O
shows NN O O
excellent NN O O
prediction NN O O
performance NN O O
as NN O O
indicated NN O O
by NN O O
its NN O O
ability NN O O
to NN O O
perfectly NN O O
predict NN O O
the NN O O
moves NN O O
made NN O O
by NN O O
professional NN O O
Go NN O O
players NN O O
in NN O O
34 NN O O
% NN O O
of NN O O
test NN O O
positions NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
present NN O O
a NN O O
novel NN O O
approach NN O I-Generic
for NN O O
automatically NN O I-Task
acquiring NN O I-Task
English NN O I-Task
topic NN O I-Task
signatures NN O I-Task
. NN O O

Given NN O O
a NN O O
particular NN O O
concept NN O I-OtherScientificTerm
, NN O O
or NN O O
word NN O I-OtherScientificTerm
sense NN O I-OtherScientificTerm
, NN O O
a NN O O
topic NN O I-OtherScientificTerm
signature NN O I-OtherScientificTerm
is NN O O
a NN O O
set NN O O
of NN O O
words NN O O
that NN O O
tend NN O O
to NN O O
co-occur NN O O
with NN O O
it NN O O
. NN O O

Topic NN O I-OtherScientificTerm
signatures NN O I-OtherScientificTerm
can NN O O
be NN O O
useful NN O O
in NN O O
a NN O O
number NN O O
of NN O O
Natural NN O I-Task
Language NN O I-Task
Processing NN O I-Task
-LRB- NN O I-Task
NLP NN O I-Task
-RRB- NN O I-Task
applications NN O I-Task
, NN O O
such NN O O
as NN O O
Word NN O I-Task
Sense NN O I-Task
Disambiguation NN O I-Task
-LRB- NN O I-Task
WSD NN O I-Task
-RRB- NN O I-Task
and NN O O
Text NN O I-Task
Summarisation NN O I-Task
. NN O O

Our NN O O
method NN O I-Generic
takes NN O O
advantage NN O O
of NN O O
the NN O O
different NN O O
way NN O O
in NN O O
which NN O O
word NN O I-OtherScientificTerm
senses NN O I-OtherScientificTerm
are NN O O
lexicalised NN O O
in NN O O
English NN O I-Material
and NN O O
Chinese NN O I-Material
, NN O O
and NN O O
also NN O O
exploits NN O O
the NN O O
large NN O O
amount NN O O
of NN O O
Chinese NN O I-Material
text NN O I-Material
available NN O O
in NN O O
corpora NN O I-Generic
and NN O O
on NN O O
the NN O O
Web NN O I-Material
. NN O O

We NN O O
evaluated NN O O
the NN O O
topic NN O I-OtherScientificTerm
signatures NN O I-OtherScientificTerm
on NN O O
a NN O O
WSD NN O I-Task
task NN O I-Task
, NN O O
where NN O O
we NN O O
trained NN O O
a NN O O
second-order NN O I-Method
vector NN O I-Method
cooccurrence NN O I-Method
algorithm NN O I-Method
on NN O O
standard NN O O
WSD NN O I-Material
datasets NN O I-Material
, NN O O
with NN O O
promising NN O O
results NN O O
. NN O O

-DOCSTART- NN O O

Joint NN O I-Task
matrix NN O I-Task
triangularization NN O I-Task
is NN O O
often NN O O
used NN O O
for NN O O
estimating NN O O
the NN O O
joint NN O I-OtherScientificTerm
eigenstructure NN O I-OtherScientificTerm
of NN O O
a NN O O
set NN O O
M NN O O
of NN O O
matrices NN O O
, NN O O
with NN O O
applications NN O O
in NN O O
signal NN O I-Task
processing NN O I-Task
and NN O O
machine NN O I-Task
learning NN O I-Task
. NN O O

We NN O O
consider NN O O
the NN O O
problem NN O O
of NN O O
approximate NN O I-Task
joint NN O I-Task
matrix NN O I-Task
triangularization NN O I-Task
when NN O O
the NN O O
matrices NN O O
in NN O O
M NN O O
are NN O O
jointly NN O O
diagonalizable NN O O
and NN O O
real NN O O
, NN O O
but NN O O
we NN O O
only NN O O
observe NN O O
a NN O O
set NN O O
M NN O O
' NN O O
of NN O O
noise NN O O
perturbed NN O O
versions NN O O
of NN O O
the NN O O
matrices NN O O
in NN O O
M NN O O
. NN O O

Our NN O O
main NN O O
result NN O O
is NN O O
a NN O O
first-order NN O I-OtherScientificTerm
upper NN O I-OtherScientificTerm
bound NN O I-OtherScientificTerm
on NN O O
the NN O O
distance NN O O
between NN O O
any NN O O
approximate NN O I-Method
joint NN O I-Method
triangularizer NN O I-Method
of NN O O
the NN O O
matrices NN O O
in NN O O
M NN O O
' NN O O
and NN O O
any NN O O
exact NN O I-Method
joint NN O I-Method
triangularizer NN O I-Method
of NN O O
the NN O O
matrices NN O O
in NN O O
M NN O O
. NN O O

The NN O O
bound NN O I-Generic
depends NN O O
only NN O O
on NN O O
the NN O O
observable NN O O
matrices NN O O
in NN O O
M NN O O
' NN O O
and NN O O
the NN O O
noise NN O O
level NN O O
. NN O O

In NN O O
particular NN O O
, NN O O
it NN O O
does NN O O
not NN O O
depend NN O O
on NN O O
optimization NN O O
specific NN O O
properties NN O O
of NN O O
the NN O O
triangularizer NN O I-Method
, NN O O
such NN O O
as NN O O
its NN O O
proximity NN O O
to NN O O
critical NN O O
points NN O O
, NN O O
that NN O O
are NN O O
typical NN O O
of NN O O
existing NN O O
bounds NN O O
in NN O O
the NN O O
literature NN O O
. NN O O

To NN O O
our NN O O
knowledge NN O O
, NN O O
this NN O O
is NN O O
the NN O O
first NN O O
a NN O O
posteriori NN O I-OtherScientificTerm
bound NN O I-OtherScientificTerm
for NN O O
joint NN O I-Task
matrix NN O I-Task
decomposition NN O I-Task
. NN O O

We NN O O
demonstrate NN O O
the NN O O
bound NN O O
on NN O O
synthetic NN O I-Material
data NN O I-Material
for NN O O
which NN O O
the NN O O
ground NN O O
truth NN O O
is NN O O
known NN O O
. NN O O

-DOCSTART- NN O O

The NN O O
psycholinguistic NN O I-OtherScientificTerm
literature NN O I-OtherScientificTerm
provides NN O O
evidence NN O O
for NN O O
syntactic NN O I-OtherScientificTerm
priming NN O I-OtherScientificTerm
, NN O O
i.e. NN O O
, NN O O
the NN O O
tendency NN O O
to NN O O
repeat NN O O
structures NN O O
. NN O O

This NN O O
paper NN O O
describes NN O O
a NN O O
method NN O I-Generic
for NN O O
incorporating NN O O
priming NN O I-OtherScientificTerm
into NN O O
an NN O O
incremental NN O I-Method
probabilistic NN O I-Method
parser NN O I-Method
. NN O O

Three NN O O
models NN O O
are NN O O
compared NN O O
, NN O O
which NN O O
involve NN O O
priming NN O I-OtherScientificTerm
of NN O O
rules NN O I-OtherScientificTerm
between NN O O
sentences NN O O
, NN O O
within NN O O
sentences NN O O
, NN O O
and NN O O
within NN O O
coordinate NN O I-OtherScientificTerm
structures NN O I-OtherScientificTerm
. NN O O

These NN O O
models NN O O
simulate NN O O
the NN O O
reading NN O O
time NN O O
advantage NN O O
for NN O O
parallel NN O I-OtherScientificTerm
structures NN O I-OtherScientificTerm
found NN O O
in NN O O
human NN O I-Material
data NN O I-Material
, NN O O
and NN O O
also NN O O
yield NN O O
a NN O O
small NN O O
increase NN O O
in NN O O
overall NN O O
parsing NN O I-Metric
accuracy NN O I-Metric
. NN O O

-DOCSTART- NN O O

Learned NN O I-Method
confidence NN O I-Method
measures NN O I-Method
gain NN O O
increasing NN O O
importance NN O O
for NN O O
outlier NN O I-Task
removal NN O I-Task
and NN O O
quality NN O I-Task
improvement NN O I-Task
in NN O O
stereo NN O I-Task
vision NN O I-Task
. NN O O

However NN O O
, NN O O
acquiring NN O O
the NN O O
necessary NN O O
training NN O O
data NN O O
is NN O O
typically NN O O
a NN O O
tedious NN O O
and NN O O
time NN O O
consuming NN O O
task NN O I-Generic
that NN O O
involves NN O O
manual NN O I-OtherScientificTerm
interaction NN O I-OtherScientificTerm
, NN O O
active NN O I-OtherScientificTerm
sensing NN O I-OtherScientificTerm
devices NN O I-OtherScientificTerm
and/or NN O O
synthetic NN O I-OtherScientificTerm
scenes NN O I-OtherScientificTerm
. NN O O

To NN O O
overcome NN O O
this NN O O
problem NN O I-Generic
, NN O O
we NN O O
propose NN O O
a NN O O
new NN O O
, NN O O
flexible NN O O
, NN O O
and NN O O
scalable NN O O
way NN O O
for NN O O
generating NN O O
training NN O O
data NN O O
that NN O O
only NN O O
requires NN O O
a NN O O
set NN O O
of NN O O
stereo NN O I-Material
images NN O I-Material
as NN O O
input NN O O
. NN O O

The NN O O
key NN O O
idea NN O O
of NN O O
our NN O O
approach NN O I-Generic
is NN O O
to NN O O
use NN O O
different NN O O
view NN O I-OtherScientificTerm
points NN O I-OtherScientificTerm
for NN O O
reasoning NN O O
about NN O O
contradictions NN O O
and NN O O
consistencies NN O O
between NN O O
multiple NN O I-OtherScientificTerm
depth NN O I-OtherScientificTerm
maps NN O I-OtherScientificTerm
generated NN O O
with NN O O
the NN O O
same NN O O
stereo NN O I-Method
algorithm NN O I-Method
. NN O O

This NN O O
enables NN O O
us NN O O
to NN O O
generate NN O O
a NN O O
huge NN O O
amount NN O O
of NN O O
training NN O O
data NN O O
in NN O O
a NN O O
fully NN O O
automated NN O O
manner NN O O
. NN O O

Among NN O O
other NN O O
experiments NN O O
, NN O O
we NN O O
demonstrate NN O O
the NN O O
potential NN O O
of NN O O
our NN O O
approach NN O I-Generic
by NN O O
boosting NN O O
the NN O O
performance NN O O
of NN O O
three NN O O
learned NN O I-Method
confidence NN O I-Method
measures NN O I-Method
on NN O O
the NN O O
KITTI2012 NN O I-Material
dataset NN O I-Material
by NN O O
simply NN O O
training NN O O
them NN O I-Generic
on NN O O
a NN O O
vast NN O O
amount NN O O
of NN O O
automatically NN O I-Material
generated NN O I-Material
training NN O I-Material
data NN O I-Material
rather NN O O
than NN O O
a NN O O
limited NN O O
amount NN O O
of NN O O
laser NN O I-Material
ground NN O I-Material
truth NN O I-Material
data NN O I-Material
. NN O O

-DOCSTART- NN O O

An NN O O
important NN O O
area NN O O
of NN O O
learning NN O I-Task
in NN O I-Task
autonomous NN O I-Task
agents NN O I-Task
is NN O O
the NN O O
ability NN O O
to NN O O
learn NN O O
domain-speciic NN O I-Method
models NN O I-Method
of NN O I-Method
actions NN O I-Method
to NN O O
be NN O O
used NN O O
by NN O O
planning NN O I-Task
systems NN O I-Task
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
present NN O O
methods NN O I-Generic
by NN O O
which NN O O
an NN O O
agent NN O O
learns NN O O
action NN O I-Method
models NN O I-Method
from NN O O
its NN O O
own NN O O
experience NN O O
and NN O O
from NN O O
its NN O O
observation NN O O
of NN O O
a NN O O
domain NN O I-OtherScientificTerm
expert NN O I-OtherScientificTerm
. NN O O

These NN O O
methods NN O I-Metric
diier NN O O
from NN O O
previous NN O O
work NN O O
in NN O O
the NN O O
area NN O O
in NN O O
two NN O O
ways NN O O
: NN O O
the NN O O
use NN O O
of NN O O
an NN O O
action NN O I-Method
model NN O I-Method
formalism NN O I-Method
which NN O O
is NN O O
better NN O O
suited NN O O
to NN O O
the NN O O
needs NN O O
of NN O O
a NN O O
re-active NN O I-Method
agent NN O I-Method
, NN O O
and NN O O
successful NN O O
implementation NN O O
of NN O O
noise-handling NN O I-Method
mechanisms NN O I-Method
. NN O O

Training NN O O
instances NN O O
are NN O O
generated NN O O
from NN O O
experience NN O O
and NN O O
observation NN O O
, NN O O
and NN O O
a NN O O
variant NN O O
of NN O O
GOLEM NN O I-Method
is NN O O
used NN O O
to NN O O
learn NN O O
action NN O I-Method
models NN O I-Method
from NN O O
these NN O O
instances NN O O
. NN O O

The NN O O
integrated NN O I-Method
learning NN O I-Method
system NN O I-Method
has NN O O
been NN O O
experimentally NN O O
validated NN O O
in NN O O
simulated NN O I-Task
construction NN O I-Task
and NN O O
ooce NN O I-Task
domains NN O I-Task
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
describes NN O O
FERRET NN O I-Method
, NN O O
an NN O O
interactive NN O I-Method
question-answering NN O I-Method
-LRB- NN O I-Method
Q/A NN O I-Method
-RRB- NN O I-Method
system NN O I-Method
designed NN O O
to NN O O
address NN O O
the NN O O
challenges NN O O
of NN O O
integrating NN O I-Task
automatic NN O I-Task
Q/A NN O I-Task
applications NN O I-Task
into NN O I-Task
real-world NN O I-Task
environments NN O I-Task
. NN O O

FERRET NN O I-Method
utilizes NN O O
a NN O O
novel NN O O
approach NN O I-Generic
to NN O O
Q/A NN O I-Method
known NN O O
as NN O O
predictive NN O I-Method
questioning NN O I-Method
which NN O O
attempts NN O O
to NN O O
identify NN O O
the NN O O
questions NN O O
-LRB- NN O O
and NN O O
answers NN O O
-RRB- NN O O
that NN O O
users NN O O
need NN O O
by NN O O
analyzing NN O O
how NN O O
a NN O O
user NN O O
interacts NN O O
with NN O O
a NN O O
system NN O O
while NN O O
gathering NN O O
information NN O O
related NN O O
to NN O O
a NN O O
particular NN O O
scenario NN O O
. NN O O

-DOCSTART- NN O O

In NN O O
order NN O O
to NN O O
build NN O O
robust NN O O
automatic NN O I-Task
abstracting NN O I-Task
systems NN O I-Task
, NN O O
there NN O O
is NN O O
a NN O O
need NN O O
for NN O O
better NN O O
training NN O I-Material
resources NN O I-Material
than NN O O
are NN O O
currently NN O O
available NN O O
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
introduce NN O O
an NN O O
annotation NN O I-Method
scheme NN O I-Method
for NN O O
scientific NN O I-Material
articles NN O I-Material
which NN O O
can NN O O
be NN O O
used NN O O
to NN O O
build NN O O
such NN O O
a NN O O
resource NN O I-Generic
in NN O O
a NN O O
consistent NN O O
way NN O O
. NN O O

The NN O O
seven NN O O
categories NN O O
of NN O O
the NN O O
scheme NN O I-Generic
are NN O O
based NN O O
on NN O O
rhetorical NN O I-Method
moves NN O I-Method
of NN O I-Method
argumentation NN O I-Method
. NN O O

Our NN O O
experimental NN O O
results NN O O
show NN O O
that NN O O
the NN O O
scheme NN O I-Generic
is NN O O
stable NN O O
, NN O O
reproducible NN O O
and NN O O
intuitive NN O O
to NN O O
use NN O O
. NN O O

-DOCSTART- NN O O

The NN O O
automated NN O I-Task
segmentation NN O I-Task
of NN O O
images NN O I-Material
into NN O O
semantically NN O O
meaningful NN O O
parts NN O O
requires NN O O
shape NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
since NN O O
low-level NN O I-Method
feature NN O I-Method
analysis NN O I-Method
alone NN O O
often NN O O
fails NN O O
to NN O O
reach NN O O
this NN O O
goal NN O O
. NN O O

We NN O O
introduce NN O O
a NN O O
novel NN O O
method NN O I-Method
of NN O O
shape NN O I-Task
constrained NN O I-Task
image NN O I-Task
segmentation NN O I-Task
which NN O O
is NN O O
based NN O O
on NN O O
mixtures NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
feature NN O I-OtherScientificTerm
distributions NN O I-OtherScientificTerm
for NN O O
color NN O I-OtherScientificTerm
and NN O O
texture NN O I-OtherScientificTerm
as NN O O
well NN O O
as NN O O
probabilistic NN O I-OtherScientificTerm
shape NN O I-OtherScientificTerm
knowledge NN O I-OtherScientificTerm
. NN O O

The NN O O
combined NN O O
approach NN O I-Generic
is NN O O
formulated NN O O
in NN O O
the NN O O
framework NN O O
of NN O O
Bayesian NN O I-Method
statistics NN O I-Method
to NN O O
account NN O O
for NN O O
the NN O O
robust-ness NN O I-Task
requirement NN O I-Task
in NN O I-Task
image NN O I-Task
understanding NN O I-Task
. NN O O

Experimental NN O O
evidence NN O O
shows NN O O
that NN O O
semantically NN O O
meaningful NN O O
segments NN O O
are NN O O
inferred NN O O
, NN O O
even NN O O
when NN O O
image NN O I-Material
data NN O I-Material
alone NN O O
gives NN O O
rise NN O O
to NN O O
ambiguous NN O O
segmentations NN O O
. NN O O

-DOCSTART- NN O O

The NN O O
goal NN O O
of NN O O
this NN O O
work NN O O
is NN O O
the NN O O
enrichment NN O O
of NN O O
human-machine NN O I-Task
interactions NN O I-Task
in NN O O
a NN O O
natural NN O I-OtherScientificTerm
language NN O I-OtherScientificTerm
environment NN O I-OtherScientificTerm
. NN O O

Because NN O O
a NN O O
speaker NN O O
and NN O O
listener NN O O
can NN O O
not NN O O
be NN O O
assured NN O O
to NN O O
have NN O O
the NN O O
same NN O O
beliefs NN O O
, NN O O
contexts NN O O
, NN O O
perceptions NN O O
, NN O O
backgrounds NN O O
, NN O O
or NN O O
goals NN O O
, NN O O
at NN O O
each NN O O
point NN O O
in NN O O
a NN O O
conversation NN O O
, NN O O
difficulties NN O O
and NN O O
mistakes NN O O
arise NN O O
when NN O O
a NN O O
listener NN O O
interprets NN O O
a NN O O
speaker NN O O
's NN O O
utterance NN O O
. NN O O

These NN O O
mistakes NN O O
can NN O O
lead NN O O
to NN O O
various NN O O
kinds NN O O
of NN O O
misunderstandings NN O O
between NN O O
speaker NN O O
and NN O O
listener NN O O
, NN O O
including NN O O
reference NN O I-OtherScientificTerm
failures NN O I-OtherScientificTerm
or NN O O
failure NN O O
to NN O O
understand NN O O
the NN O O
speaker NN O I-OtherScientificTerm
's NN O I-OtherScientificTerm
intention NN O I-OtherScientificTerm
. NN O O

We NN O O
call NN O O
these NN O O
misunderstandings NN O O
miscommunication NN O I-Task
. NN O O

Such NN O O
mistakes NN O O
can NN O O
slow NN O O
, NN O O
and NN O O
possibly NN O O
break NN O O
down NN O O
, NN O O
communication NN O O
. NN O O

Our NN O O
goal NN O O
is NN O O
to NN O O
recognize NN O O
and NN O O
isolate NN O O
such NN O O
miscommunications NN O I-Task
and NN O O
circumvent NN O O
them NN O I-Generic
. NN O O

This NN O O
paper NN O O
highlights NN O O
a NN O O
particular NN O O
class NN O O
of NN O O
miscommunication NN O I-Task
-- NN O O
reference NN O I-Task
problems NN O I-Task
-- NN O O
by NN O O
describing NN O O
a NN O O
case NN O O
study NN O O
and NN O O
techniques NN O I-Generic
for NN O O
avoiding NN O O
failures NN O I-Task
of NN O I-Task
reference NN O I-Task
. NN O O

We NN O O
want NN O O
to NN O O
illustrate NN O O
a NN O O
framework NN O O
less NN O O
restrictive NN O O
than NN O O
earlier NN O O
ones NN O O
by NN O O
allowing NN O O
a NN O O
speaker NN O O
leeway NN O O
in NN O O
forming NN O O
an NN O O
utterance NN O O
about NN O O
a NN O O
task NN O O
and NN O O
in NN O O
determining NN O O
the NN O O
conversational NN O O
vehicle NN O O
to NN O O
deliver NN O O
it NN O O
. NN O O

The NN O O
paper NN O O
also NN O O
promotes NN O O
a NN O O
new NN O O
view NN O O
for NN O O
extensional NN O I-Task
reference NN O I-Task
. NN O O

-DOCSTART- NN O O

Combination NN O I-Method
methods NN O I-Method
are NN O O
an NN O O
effective NN O O
way NN O O
of NN O O
improving NN O O
system NN O O
performance NN O O
. NN O O

This NN O O
paper NN O O
examines NN O O
the NN O O
benefits NN O O
of NN O O
system NN O I-Method
combination NN O I-Method
for NN O O
unsupervised NN O I-Task
WSD NN O I-Task
. NN O O

We NN O O
investigate NN O O
several NN O O
voting NN O I-Method
- NN O I-Method
and NN O I-Method
arbiter-based NN O I-Method
combination NN O I-Method
strategies NN O I-Method
over NN O O
a NN O O
diverse NN O O
pool NN O O
of NN O O
unsupervised NN O I-Method
WSD NN O I-Method
systems NN O I-Method
. NN O O

Our NN O O
combination NN O I-Method
methods NN O I-Method
rely NN O O
on NN O O
predominant NN O I-OtherScientificTerm
senses NN O I-OtherScientificTerm
which NN O O
are NN O O
derived NN O O
automatically NN O O
from NN O O
raw NN O I-Material
text NN O I-Material
. NN O O

Experiments NN O O
using NN O O
the NN O O
SemCor NN O I-Material
and NN O I-Material
Senseval-3 NN O I-Material
data NN O I-Material
sets NN O I-Material
demonstrate NN O O
that NN O O
our NN O O
ensembles NN O I-Generic
yield NN O O
significantly NN O O
better NN O O
results NN O O
when NN O O
compared NN O O
with NN O O
state-of-the-art NN O I-Generic
. NN O O

-DOCSTART- NN O O

The NN O O
applicability NN O O
of NN O O
many NN O O
current NN O O
information NN O I-Method
extraction NN O I-Method
techniques NN O I-Method
is NN O O
severely NN O O
limited NN O O
by NN O O
the NN O O
need NN O O
for NN O O
supervised NN O I-Material
training NN O I-Material
data NN O I-Material
. NN O O

We NN O O
demonstrate NN O O
that NN O O
for NN O O
certain NN O O
field NN O I-Task
structured NN O I-Task
extraction NN O I-Task
tasks NN O I-Task
, NN O O
such NN O O
as NN O O
classified NN O I-Material
advertisements NN O I-Material
and NN O O
bibliographic NN O I-Material
citations NN O I-Material
, NN O O
small NN O O
amounts NN O O
of NN O O
prior NN O I-OtherScientificTerm
knowledge NN O I-OtherScientificTerm
can NN O O
be NN O O
used NN O O
to NN O O
learn NN O O
effective NN O O
models NN O O
in NN O O
a NN O O
primarily NN O O
unsupervised NN O O
fashion NN O O
. NN O O

Although NN O O
hidden NN O I-Method
Markov NN O I-Method
models NN O I-Method
-LRB- NN O I-Method
HMMs NN O I-Method
-RRB- NN O I-Method
provide NN O O
a NN O O
suitable NN O O
generative NN O I-Method
model NN O I-Method
for NN O O
field NN O I-Material
structured NN O I-Material
text NN O I-Material
, NN O O
general NN O O
unsupervised NN O I-Method
HMM NN O I-Method
learning NN O I-Method
fails NN O O
to NN O O
learn NN O O
useful NN O O
structure NN O O
in NN O O
either NN O O
of NN O O
our NN O O
domains NN O O
. NN O O

However NN O O
, NN O O
one NN O O
can NN O O
dramatically NN O O
improve NN O O
the NN O O
quality NN O O
of NN O O
the NN O O
learned NN O O
structure NN O O
by NN O O
exploiting NN O O
simple NN O O
prior NN O I-OtherScientificTerm
knowledge NN O I-OtherScientificTerm
of NN O O
the NN O O
desired NN O O
solutions NN O O
. NN O O

In NN O O
both NN O O
domains NN O O
, NN O O
we NN O O
found NN O O
that NN O O
unsupervised NN O I-Method
methods NN O I-Method
can NN O O
attain NN O O
accuracies NN O I-Metric
with NN O O
400 NN O O
unlabeled NN O I-Material
examples NN O I-Material
comparable NN O O
to NN O O
those NN O O
attained NN O O
by NN O O
supervised NN O I-Method
methods NN O I-Method
on NN O O
50 NN O O
labeled NN O I-Material
examples NN O I-Material
, NN O O
and NN O O
that NN O O
semi-supervised NN O I-Method
methods NN O I-Method
can NN O O
make NN O O
good NN O O
use NN O O
of NN O O
small NN O O
amounts NN O O
of NN O O
labeled NN O I-Material
data NN O I-Material
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
gives NN O O
an NN O O
overall NN O O
account NN O O
of NN O O
a NN O O
prototype NN O O
natural NN O I-Method
language NN O I-Method
question NN O I-Method
answering NN O I-Method
system NN O I-Method
, NN O O
called NN O O
Chat-80 NN O I-Method
. NN O O

Chat-80 NN O I-Method
has NN O O
been NN O O
designed NN O O
to NN O O
be NN O O
both NN O O
efficient NN O O
and NN O O
easily NN O O
adaptable NN O O
to NN O O
a NN O O
variety NN O O
of NN O O
applications NN O O
. NN O O

The NN O O
system NN O I-Generic
is NN O O
implemented NN O O
entirely NN O O
in NN O O
Prolog NN O I-OtherScientificTerm
, NN O O
a NN O O
programming NN O I-OtherScientificTerm
language NN O I-OtherScientificTerm
based NN O O
on NN O O
logic NN O I-OtherScientificTerm
. NN O O

With NN O O
the NN O O
aid NN O O
of NN O O
a NN O O
logic-based NN O I-Method
grammar NN O I-Method
formalism NN O I-Method
called NN O O
extraposition NN O I-Method
grammars NN O I-Method
, NN O O
Chat-80 NN O I-Method
translates NN O O
English NN O O
questions NN O O
into NN O O
the NN O O
Prolog NN O I-OtherScientificTerm
subset NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
logic NN O I-OtherScientificTerm
. NN O O

The NN O O
resulting NN O O
logical NN O I-OtherScientificTerm
expression NN O I-OtherScientificTerm
is NN O O
then NN O O
transformed NN O O
by NN O O
a NN O O
planning NN O I-Method
algorithm NN O I-Method
into NN O O
efficient NN O O
Prolog NN O I-OtherScientificTerm
, NN O O
cf. NN O O
query NN O I-Method
optimisation NN O I-Method
in NN O O
a NN O O
relational NN O I-Material
database NN O I-Material
. NN O O

Finally NN O O
, NN O O
the NN O O
Prolog NN O I-OtherScientificTerm
form NN O I-OtherScientificTerm
is NN O O
executed NN O O
to NN O O
yield NN O O
the NN O O
answer NN O O
. NN O O

-DOCSTART- NN O O

Human NN O I-Task
action NN O I-Task
recognition NN O I-Task
from NN O O
well-segmented NN O I-Material
3D NN O I-Material
skeleton NN O I-Material
data NN O I-Material
has NN O O
been NN O O
intensively NN O O
studied NN O O
and NN O O
attracting NN O O
an NN O O
increasing NN O O
attention NN O O
. NN O O

Online NN O I-Task
action NN O I-Task
detection NN O I-Task
goes NN O O
one NN O O
step NN O O
further NN O O
and NN O O
is NN O O
more NN O O
challenging NN O O
, NN O O
which NN O O
identifies NN O O
the NN O O
action NN O I-OtherScientificTerm
type NN O I-OtherScientificTerm
and NN O O
localizes NN O O
the NN O O
action NN O I-OtherScientificTerm
positions NN O I-OtherScientificTerm
on NN O O
the NN O O
fly NN O O
from NN O O
the NN O O
untrimmed NN O I-Material
stream NN O I-Material
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
study NN O O
the NN O O
problem NN O O
of NN O O
online NN O I-Task
action NN O I-Task
detection NN O I-Task
from NN O O
the NN O O
streaming NN O I-Material
skeleton NN O I-Material
data NN O I-Material
. NN O O

We NN O O
propose NN O O
a NN O O
multi-task NN O I-Method
end-to-end NN O I-Method
Joint NN O I-Method
Classification-Regression NN O I-Method
Recurrent NN O I-Method
Neural NN O I-Method
Network NN O I-Method
to NN O O
better NN O O
explore NN O O
the NN O O
action NN O I-OtherScientificTerm
type NN O I-OtherScientificTerm
and NN O O
temporal NN O I-OtherScientificTerm
localiza-tion NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
. NN O O

By NN O O
employing NN O O
a NN O O
joint NN O I-OtherScientificTerm
classification NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
regression NN O I-OtherScientificTerm
optimization NN O I-OtherScientificTerm
objective NN O I-OtherScientificTerm
, NN O O
this NN O O
network NN O I-Generic
is NN O O
capable NN O O
of NN O O
automatically NN O O
localizing NN O O
the NN O O
start NN O O
and NN O O
end NN O O
points NN O O
of NN O O
actions NN O O
more NN O O
accurately NN O O
. NN O O

Specifically NN O O
, NN O O
by NN O O
leveraging NN O O
the NN O O
merits NN O O
of NN O O
the NN O O
deep NN O I-Method
Long NN O I-Method
Short-Term NN O I-Method
Memory NN O I-Method
-LRB- NN O I-Method
LSTM NN O I-Method
-RRB- NN O I-Method
subnetwork NN O I-Method
, NN O O
the NN O O
proposed NN O O
model NN O I-Generic
automatically NN O O
captures NN O O
the NN O O
complex NN O O
long-range NN O I-OtherScientificTerm
temporal NN O I-OtherScientificTerm
dynamics NN O I-OtherScientificTerm
, NN O O
which NN O O
naturally NN O O
avoids NN O O
the NN O O
typical NN O O
sliding NN O I-Method
window NN O I-Method
design NN O I-Method
and NN O O
thus NN O O
ensures NN O O
high NN O O
computational NN O I-OtherScientificTerm
efficiency NN O I-OtherScientificTerm
. NN O O

Furthermore NN O O
, NN O O
the NN O O
subtask NN O O
of NN O O
regression NN O I-Task
optimization NN O I-Task
provides NN O O
the NN O O
ability NN O O
to NN O O
forecast NN O O
the NN O O
action NN O O
prior NN O O
to NN O O
its NN O O
occurrence NN O O
. NN O O

To NN O O
evaluate NN O O
our NN O O
proposed NN O O
model NN O I-Generic
, NN O O
we NN O O
build NN O O
a NN O O
large NN O O
streaming NN O I-Material
video NN O I-Material
dataset NN O I-Material
with NN O O
annotations NN O O
. NN O O

Experimental NN O O
results NN O O
on NN O O
our NN O O
dataset NN O I-Generic
and NN O O
the NN O O
public NN O O
G3D NN O I-Material
dataset NN O I-Material
both NN O O
demonstrate NN O O
very NN O O
promising NN O O
performance NN O O
of NN O O
our NN O O
scheme NN O O
. NN O O

-DOCSTART- NN O O

The NN O O
task NN O O
of NN O O
machine NN O I-Task
translation NN O I-Task
-LRB- NN O I-Task
MT NN O I-Task
-RRB- NN O I-Task
evaluation NN O I-Task
is NN O O
closely NN O O
related NN O O
to NN O O
the NN O O
task NN O O
of NN O O
sentence-level NN O I-Task
semantic NN O I-Task
equivalence NN O I-Task
classification NN O I-Task
. NN O O

This NN O O
paper NN O O
investigates NN O O
the NN O O
utility NN O O
of NN O O
applying NN O O
standard NN O O
MT NN O I-Metric
evaluation NN O I-Metric
methods NN O I-Metric
-LRB- NN O O
BLEU NN O I-Metric
, NN O O
NIST NN O I-Metric
, NN O O
WER NN O I-Metric
and NN O O
PER NN O I-Metric
-RRB- NN O O
to NN O O
building NN O O
classifiers NN O I-Method
to NN O O
predict NN O O
semantic NN O I-Task
equivalence NN O I-Task
and NN O O
entailment NN O I-Task
. NN O O

We NN O O
also NN O O
introduce NN O O
a NN O O
novel NN O O
classification NN O I-Method
method NN O I-Method
based NN O O
on NN O O
PER NN O I-Metric
which NN O O
leverages NN O O
part NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
speech NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
of NN O O
the NN O O
words NN O O
contributing NN O O
to NN O O
the NN O O
word NN O I-Task
matches NN O I-Task
and NN O I-Task
non-matches NN O I-Task
in NN O O
the NN O O
sentence NN O O
. NN O O

Our NN O O
results NN O O
show NN O O
that NN O O
MT NN O I-Method
evaluation NN O I-Method
techniques NN O I-Method
are NN O O
able NN O O
to NN O O
produce NN O O
useful NN O O
features NN O I-OtherScientificTerm
for NN O O
paraphrase NN O I-Task
classification NN O I-Task
and NN O O
to NN O O
a NN O O
lesser NN O O
extent NN O O
entailment NN O I-Task
. NN O O

Our NN O O
technique NN O I-Generic
gives NN O O
a NN O O
substantial NN O O
improvement NN O O
in NN O O
paraphrase NN O I-Metric
classification NN O I-Metric
accuracy NN O I-Metric
over NN O O
all NN O O
of NN O O
the NN O O
other NN O O
models NN O I-Generic
used NN O O
in NN O O
the NN O O
experiments NN O O
. NN O O

-DOCSTART- NN O O

Given NN O O
an NN O O
object NN O I-Method
model NN O I-Method
and NN O O
a NN O O
black-box NN O I-Metric
measure NN O I-Metric
of NN O I-Metric
similarity NN O I-Metric
between NN O O
the NN O O
model NN O I-Generic
and NN O O
candidate NN O O
targets NN O O
, NN O O
we NN O O
consider NN O O
visual NN O I-Task
object NN O I-Task
tracking NN O I-Task
as NN O O
a NN O O
numerical NN O I-Task
optimization NN O I-Task
problem NN O I-Task
. NN O O

During NN O O
normal NN O O
tracking NN O O
conditions NN O O
when NN O O
the NN O O
object NN O O
is NN O O
visible NN O O
from NN O O
frame NN O O
to NN O O
frame NN O O
, NN O O
local NN O I-Method
optimization NN O I-Method
is NN O O
used NN O O
to NN O O
track NN O O
the NN O O
local NN O I-OtherScientificTerm
mode NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
the NN O I-OtherScientificTerm
similarity NN O I-OtherScientificTerm
measure NN O I-OtherScientificTerm
in NN O O
a NN O O
parameter NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
translation NN O I-OtherScientificTerm
, NN O I-OtherScientificTerm
rotation NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
scale NN O I-OtherScientificTerm
. NN O O

However NN O O
, NN O O
when NN O O
the NN O O
object NN O O
becomes NN O O
partially NN O O
or NN O O
totally NN O O
occluded NN O O
, NN O O
such NN O O
local NN O I-Method
tracking NN O I-Method
is NN O O
prone NN O O
to NN O O
failure NN O O
, NN O O
especially NN O O
when NN O O
common NN O O
prediction NN O I-Method
techniques NN O I-Method
like NN O O
the NN O O
Kalman NN O I-Method
filter NN O I-Method
do NN O O
not NN O O
provide NN O O
a NN O O
good NN O O
estimate NN O O
of NN O O
object NN O O
parameters NN O O
in NN O O
future NN O O
frames NN O O
. NN O O

To NN O O
recover NN O O
from NN O O
these NN O O
inevitable NN O O
tracking NN O O
failures NN O O
, NN O O
we NN O O
consider NN O O
object NN O I-Task
detection NN O I-Task
as NN O O
a NN O O
global NN O I-Task
optimization NN O I-Task
problem NN O I-Task
and NN O O
solve NN O O
it NN O I-Generic
via NN O O
Adaptive NN O I-Method
Simulated NN O I-Method
Annealing NN O I-Method
-LRB- NN O I-Method
ASA NN O I-Method
-RRB- NN O I-Method
, NN O O
a NN O O
method NN O I-Generic
that NN O O
avoids NN O O
becoming NN O O
trapped NN O O
at NN O O
local NN O O
modes NN O O
and NN O O
is NN O O
much NN O O
faster NN O O
than NN O O
exhaustive NN O I-Method
search NN O I-Method
. NN O O

As NN O O
a NN O O
Monte NN O I-Method
Carlo NN O I-Method
approach NN O I-Method
, NN O O
ASA NN O I-Method
stochastically NN O O
samples NN O O
the NN O O
parameter NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
, NN O O
in NN O O
contrast NN O O
to NN O O
local NN O I-Method
deterministic NN O I-Method
search NN O I-Method
. NN O O

We NN O O
apply NN O O
cluster NN O I-Method
analysis NN O I-Method
on NN O O
the NN O O
sampled NN O I-OtherScientificTerm
parameter NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
to NN O O
redetect NN O O
the NN O O
object NN O O
and NN O O
renew NN O O
the NN O O
local NN O I-Method
tracker NN O I-Method
. NN O O

Our NN O O
numerical NN O I-Method
hybrid NN O I-Method
local NN O I-Method
and NN O I-Method
global NN O I-Method
mode-seeking NN O I-Method
tracker NN O I-Method
is NN O O
validated NN O O
on NN O O
challenging NN O O
airborne NN O I-Material
videos NN O I-Material
with NN O O
heavy NN O I-OtherScientificTerm
occlusion NN O I-OtherScientificTerm
and NN O O
large NN O O
camera NN O I-OtherScientificTerm
motions NN O I-OtherScientificTerm
. NN O O

Our NN O O
approach NN O I-Generic
outperforms NN O O
state-of-the-art NN O I-Generic
trackers NN O I-Generic
on NN O O
the NN O O
VIVID NN O I-Material
benchmark NN O I-Material
datasets NN O I-Material
. NN O O

-DOCSTART- NN O O

Techniques NN O I-Generic
for NN O O
automatically NN O I-Method
training NN O I-Method
modules NN O I-Method
of NN O O
a NN O O
natural NN O I-Method
language NN O I-Method
generator NN O I-Method
have NN O O
recently NN O O
been NN O O
proposed NN O O
, NN O O
but NN O O
a NN O O
fundamental NN O O
concern NN O O
is NN O O
whether NN O O
the NN O O
quality NN O I-Metric
of NN O I-Metric
utterances NN O I-OtherScientificTerm
produced NN O O
with NN O O
trainable NN O I-Method
components NN O I-Method
can NN O O
compete NN O O
with NN O O
hand-crafted NN O I-Method
template-based NN O I-Method
or NN O I-Method
rule-based NN O I-Method
approaches NN O I-Method
. NN O O

In NN O O
this NN O O
paper NN O O
We NN O O
experimentally NN O O
evaluate NN O O
a NN O O
trainable NN O I-Method
sentence NN O I-Method
planner NN O I-Method
for NN O O
a NN O O
spoken NN O I-Method
dialogue NN O I-Method
system NN O I-Method
by NN O O
eliciting NN O O
subjective NN O I-OtherScientificTerm
human NN O I-OtherScientificTerm
judgments NN O I-OtherScientificTerm
. NN O O

In NN O O
order NN O O
to NN O O
perform NN O O
an NN O O
exhaustive NN O O
comparison NN O O
, NN O O
we NN O O
also NN O O
evaluate NN O O
a NN O O
hand-crafted NN O I-Method
template-based NN O I-Method
generation NN O I-Method
component NN O I-Method
, NN O O
two NN O O
rule-based NN O I-Method
sentence NN O I-Method
planners NN O I-Method
, NN O O
and NN O O
two NN O O
baseline NN O I-Method
sentence NN O I-Method
planners NN O I-Method
. NN O O

We NN O O
show NN O O
that NN O O
the NN O O
trainable NN O I-Method
sentence NN O I-Method
planner NN O I-Method
performs NN O O
better NN O O
than NN O O
the NN O O
rule-based NN O I-Method
systems NN O I-Method
and NN O O
the NN O O
baselines NN O I-Generic
, NN O O
and NN O O
as NN O O
well NN O O
as NN O O
the NN O O
hand-crafted NN O I-Method
system NN O I-Method
. NN O O

-DOCSTART- NN O O

A NN O O
new NN O O
algorithm NN O I-Generic
is NN O O
proposed NN O O
for NN O O
novel NN O I-Task
view NN O I-Task
generation NN O I-Task
in NN O O
one-to-one NN O I-Task
teleconferencing NN O I-Task
applications NN O I-Task
. NN O O

Given NN O O
the NN O O
video NN O I-Material
streams NN O I-Material
acquired NN O O
by NN O O
two NN O O
cameras NN O I-OtherScientificTerm
placed NN O O
on NN O O
either NN O O
side NN O O
of NN O O
a NN O O
computer NN O I-OtherScientificTerm
monitor NN O I-OtherScientificTerm
, NN O O
the NN O O
proposed NN O O
algorithm NN O I-Generic
synthesises NN O O
images NN O I-Material
from NN O O
a NN O O
virtual NN O I-OtherScientificTerm
camera NN O I-OtherScientificTerm
in NN O O
arbitrary NN O I-OtherScientificTerm
position NN O I-OtherScientificTerm
-LRB- NN O O
typically NN O O
located NN O O
within NN O O
the NN O O
monitor NN O O
-RRB- NN O O
to NN O O
facilitate NN O O
eye NN O I-OtherScientificTerm
contact NN O I-OtherScientificTerm
. NN O O

Our NN O O
technique NN O I-Generic
is NN O O
based NN O O
on NN O O
an NN O O
improved NN O O
, NN O O
dynamic-programming NN O I-Method
, NN O I-Method
stereo NN O I-Method
algorithm NN O I-Method
for NN O O
efficient NN O O
novel-view NN O I-Task
generation NN O I-Task
. NN O O

The NN O O
two NN O O
main NN O O
contributions NN O O
of NN O O
this NN O O
paper NN O O
are NN O O
: NN O O
i NN O O
-RRB- NN O O
a NN O O
new NN O O
type NN O O
of NN O O
three-plane NN O I-Method
graph NN O I-Method
for NN O O
dense-stereo NN O I-Method
dynamic-programming NN O I-Method
, NN O O
that NN O O
encourages NN O O
correct NN O O
occlusion NN O I-Task
labeling NN O I-Task
; NN O O
ii NN O O
-RRB- NN O O
a NN O O
compact NN O I-OtherScientificTerm
geometric NN O I-OtherScientificTerm
derivation NN O I-OtherScientificTerm
for NN O O
novel-view NN O I-Task
synthesis NN O I-Task
by NN O O
direct NN O I-Method
projection NN O I-Method
of NN O I-Method
the NN O I-Method
minimum-cost NN O I-Method
surface NN O I-Method
. NN O O

Furthermore NN O O
, NN O O
this NN O O
paper NN O O
presents NN O O
a NN O O
novel NN O O
algorithm NN O I-Generic
for NN O O
the NN O O
temporal NN O I-Task
maintenance NN O I-Task
of NN O I-Task
a NN O I-Task
background NN O I-Task
model NN O I-Task
to NN O O
enhance NN O O
the NN O O
rendering NN O I-Task
of NN O I-Task
occlusions NN O I-Task
and NN O O
reduce NN O O
temporal NN O I-OtherScientificTerm
artefacts NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
flicker NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
; NN O O
and NN O O
a NN O O
cost NN O I-Method
aggregation NN O I-Method
algorithm NN O I-Method
that NN O O
acts NN O O
directly NN O O
on NN O O
our NN O O
three-dimensional NN O I-OtherScientificTerm
matching NN O I-OtherScientificTerm
cost NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
. NN O O

Examples NN O O
are NN O O
given NN O O
that NN O O
demonstrate NN O O
the NN O O
robustness NN O I-Metric
of NN O O
the NN O O
new NN O O
algorithm NN O I-Generic
to NN O O
spatial NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
temporal NN O I-OtherScientificTerm
artefacts NN O I-OtherScientificTerm
for NN O O
long NN O I-Material
stereo NN O I-Material
video NN O I-Material
streams NN O I-Material
. NN O O

These NN O O
include NN O O
demonstrations NN O O
of NN O O
synthesis NN O I-Task
of NN O I-Task
cyclopean NN O I-OtherScientificTerm
views NN O I-OtherScientificTerm
of NN O I-Task
extended NN O I-Material
conversational NN O I-Material
sequences NN O I-Material
. NN O O

We NN O O
further NN O O
demonstrate NN O O
synthesis NN O I-Task
from NN O O
a NN O O
freely NN O O
translating NN O I-OtherScientificTerm
virtual NN O I-OtherScientificTerm
camera NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

Manual NN O I-Task
acquisition NN O I-Task
of NN O I-Task
semantic NN O I-Task
constraints NN O I-Task
in NN O O
broad NN O O
domains NN O O
is NN O O
very NN O O
expensive NN O O
. NN O O

This NN O O
paper NN O O
presents NN O O
an NN O O
automatic NN O O
scheme NN O O
for NN O O
collecting NN O O
statistics NN O O
on NN O O
cooccurrence NN O I-OtherScientificTerm
patterns NN O I-OtherScientificTerm
in NN O O
a NN O O
large NN O O
corpus NN O O
. NN O O

To NN O O
a NN O O
large NN O O
extent NN O O
, NN O O
these NN O O
statistics NN O O
reflect NN O O
semantic NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
and NN O O
thus NN O O
are NN O O
used NN O O
to NN O O
disambiguate NN O O
anaphora NN O I-OtherScientificTerm
references NN O I-OtherScientificTerm
and NN O O
syntactic NN O I-OtherScientificTerm
ambiguities NN O I-OtherScientificTerm
. NN O O

The NN O O
scheme NN O O
was NN O O
implemented NN O O
by NN O O
gathering NN O O
statistics NN O O
on NN O O
the NN O O
output NN O O
of NN O O
other NN O O
linguistic NN O I-Method
tools NN O I-Method
. NN O O

An NN O O
experiment NN O O
was NN O O
performed NN O O
to NN O O
resolve NN O O
references NN O O
of NN O O
the NN O O
pronoun NN O O
`` NN O O
it NN O O
'' NN O O
in NN O O
sentences NN O O
that NN O O
were NN O O
randomly NN O O
selected NN O O
from NN O O
the NN O O
corpus NN O O
. NN O O

The NN O O
results NN O O
of NN O O
the NN O O
experiment NN O O
show NN O O
that NN O O
in NN O O
most NN O O
of NN O O
the NN O O
cases NN O O
the NN O O
cooccurrence NN O I-OtherScientificTerm
statistics NN O I-OtherScientificTerm
indeed NN O O
reflect NN O O
the NN O O
semantic NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
and NN O O
thus NN O O
provide NN O O
a NN O O
basis NN O O
for NN O O
a NN O O
useful NN O O
disambiguation NN O I-Method
tool NN O I-Method
. NN O O

-DOCSTART- NN O O

We NN O O
present NN O O
a NN O O
novel NN O O
method NN O I-Generic
for NN O O
discovering NN O I-Task
parallel NN O I-Task
sentences NN O I-Task
in NN O O
comparable NN O I-Material
, NN O I-Material
non-parallel NN O I-Material
corpora NN O I-Material
. NN O O

We NN O O
train NN O O
a NN O O
maximum NN O I-Method
entropy NN O I-Method
classifier NN O I-Method
that NN O O
, NN O O
given NN O O
a NN O O
pair NN O O
of NN O O
sentences NN O O
, NN O O
can NN O O
reliably NN O O
determine NN O O
whether NN O O
or NN O O
not NN O O
they NN O O
are NN O O
translations NN O O
of NN O O
each NN O O
other NN O O
. NN O O

Using NN O O
this NN O O
approach NN O I-Generic
, NN O O
we NN O O
extract NN O O
parallel NN O I-Material
data NN O I-Material
from NN O O
large NN O O
Chinese NN O I-Material
, NN O I-Material
Arabic NN O I-Material
, NN O I-Material
and NN O I-Material
English NN O I-Material
non-parallel NN O I-Material
newspaper NN O I-Material
corpora NN O I-Material
. NN O O

We NN O O
evaluate NN O O
the NN O O
quality NN O O
of NN O O
the NN O O
extracted NN O O
data NN O O
by NN O O
showing NN O O
that NN O O
it NN O I-Generic
improves NN O O
the NN O O
performance NN O O
of NN O O
a NN O O
state-of-the-art NN O O
statistical NN O I-Method
machine NN O I-Method
translation NN O I-Method
system NN O I-Method
. NN O O

We NN O O
also NN O O
show NN O O
that NN O O
a NN O O
good-quality NN O O
MT NN O I-Method
system NN O I-Method
can NN O O
be NN O O
built NN O O
from NN O O
scratch NN O O
by NN O O
starting NN O O
with NN O O
a NN O O
very NN O O
small NN O O
parallel NN O I-Material
corpus NN O I-Material
-LRB- NN O O
100,000 NN O O
words NN O O
-RRB- NN O O
and NN O O
exploiting NN O O
a NN O O
large NN O O
non-parallel NN O I-Material
corpus NN O I-Material
. NN O O

Thus NN O O
, NN O O
our NN O O
method NN O I-Generic
can NN O O
be NN O O
applied NN O O
with NN O O
great NN O O
benefit NN O O
to NN O O
language NN O O
pairs NN O O
for NN O O
which NN O O
only NN O O
scarce NN O I-Material
resources NN O I-Material
are NN O O
available NN O O
. NN O O

-DOCSTART- NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
describe NN O O
a NN O O
search NN O I-Generic
procedure NN O I-Generic
for NN O O
statistical NN O I-Task
machine NN O I-Task
translation NN O I-Task
-LRB- NN O I-Task
MT NN O I-Task
-RRB- NN O I-Task
based NN O O
on NN O O
dynamic NN O I-Method
programming NN O I-Method
-LRB- NN O I-Method
DP NN O I-Method
-RRB- NN O I-Method
. NN O O

Starting NN O O
from NN O O
a NN O O
DP-based NN O I-Method
solution NN O I-Method
to NN O O
the NN O O
traveling NN O I-Method
salesman NN O I-Method
problem NN O I-Method
, NN O O
we NN O O
present NN O O
a NN O O
novel NN O O
technique NN O I-Generic
to NN O O
restrict NN O O
the NN O O
possible NN O O
word NN O I-OtherScientificTerm
reordering NN O I-OtherScientificTerm
between NN O O
source NN O O
and NN O O
target NN O O
language NN O O
in NN O O
order NN O O
to NN O O
achieve NN O O
an NN O O
efficient NN O O
search NN O I-Method
algorithm NN O I-Method
. NN O O

A NN O O
search NN O O
restriction NN O O
especially NN O O
useful NN O O
for NN O O
the NN O O
translation NN O O
direction NN O O
from NN O O
German NN O O
to NN O O
English NN O O
is NN O O
presented NN O O
. NN O O

The NN O O
experimental NN O O
tests NN O O
are NN O O
carried NN O O
out NN O O
on NN O O
the NN O O
Verbmobil NN O I-Material
task NN O I-Material
-LRB- NN O O
German-English NN O I-Material
, NN O O
8000-word NN O O
vocabulary NN O O
-RRB- NN O O
, NN O O
which NN O O
is NN O O
a NN O O
limited-domain NN O I-Task
spoken-language NN O I-Task
task NN O I-Task
. NN O O

-DOCSTART- NN O O

A NN O O
purely NN O O
functional NN O O
implementation NN O O
of NN O O
LR-parsers NN O I-Method
is NN O O
given NN O O
, NN O O
together NN O O
with NN O O
a NN O O
simple NN O O
correctness NN O I-Method
proof NN O I-Method
. NN O O

It NN O I-Generic
is NN O O
presented NN O O
as NN O O
a NN O O
generalization NN O O
of NN O O
the NN O O
recursive NN O I-Method
descent NN O I-Method
parser NN O I-Method
. NN O O

For NN O O
non-LR NN O I-Method
grammars NN O I-Method
the NN O O
time-complexity NN O I-Metric
of NN O O
our NN O O
parser NN O I-Method
is NN O O
cubic NN O O
if NN O O
the NN O O
functions NN O O
that NN O O
constitute NN O O
the NN O O
parser NN O I-Method
are NN O O
implemented NN O O
as NN O O
memo-functions NN O I-Method
, NN O O
i.e. NN O O
functions NN O O
that NN O O
memorize NN O O
the NN O O
results NN O O
of NN O O
previous NN O O
invocations NN O O
. NN O O

Memo-functions NN O I-Method
also NN O O
facilitate NN O O
a NN O O
simple NN O O
way NN O O
to NN O O
construct NN O O
a NN O O
very NN O O
compact NN O O
representation NN O O
of NN O O
the NN O O
parse NN O I-OtherScientificTerm
forest NN O I-OtherScientificTerm
. NN O O

For NN O O
LR NN O I-Method
-LRB- NN O I-Method
0 NN O I-Method
-RRB- NN O I-Method
grammars NN O I-Method
, NN O O
our NN O O
algorithm NN O I-Generic
is NN O O
closely NN O O
related NN O O
to NN O O
the NN O O
recursive NN O I-OtherScientificTerm
ascent NN O I-OtherScientificTerm
parsers NN O I-OtherScientificTerm
recently NN O O
discovered NN O O
by NN O O
Kruse-man NN O O
Aretz NN O O
-LSB- NN O O
1 NN O O
-RSB- NN O O
and NN O O
Roberts NN O O
-LSB- NN O O
2 NN O O
-RSB- NN O O
. NN O O

Extended NN O I-Method
CF NN O I-Method
grammars NN O I-Method
-LRB- NN O O
grammars NN O I-Method
with NN O O
regular NN O I-OtherScientificTerm
expressions NN O I-OtherScientificTerm
at NN O O
the NN O O
right NN O O
hand NN O O
side NN O O
-RRB- NN O O
can NN O O
be NN O O
parsed NN O O
with NN O O
a NN O O
simple NN O O
modification NN O O
of NN O O
the NN O O
LR-parser NN O I-Method
for NN O O
normal NN O O
CF NN O I-Method
grammars NN O I-Method
. NN O O

-DOCSTART- NN O O

In NN O O
this NN O O
paper NN O O
we NN O O
explore NN O O
a NN O O
new NN O O
theory NN O I-Method
of NN O I-Method
discourse NN O I-Method
structure NN O I-Method
that NN O O
stresses NN O O
the NN O O
role NN O O
of NN O O
purpose NN O O
and NN O O
processing NN O O
in NN O O
discourse NN O I-OtherScientificTerm
. NN O O

In NN O O
this NN O O
theory NN O I-Generic
, NN O O
discourse NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
is NN O O
composed NN O O
of NN O O
three NN O O
separate NN O O
but NN O O
interrelated NN O O
components NN O I-Generic
: NN O O
the NN O O
structure NN O O
of NN O O
the NN O O
sequence NN O O
of NN O O
utterances NN O O
-LRB- NN O O
called NN O O
the NN O O
linguistic NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
-RRB- NN O O
, NN O O
a NN O O
structure NN O O
of NN O O
purposes NN O O
-LRB- NN O O
called NN O O
the NN O O
intentional NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
-RRB- NN O O
, NN O O
and NN O O
the NN O O
state NN O O
of NN O O
focus NN O O
of NN O O
attention NN O O
-LRB- NN O O
called NN O O
the NN O O
attentional NN O I-OtherScientificTerm
state NN O I-OtherScientificTerm
-RRB- NN O O
. NN O O

The NN O O
linguistic NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
consists NN O O
of NN O O
segments NN O O
of NN O O
the NN O O
discourse NN O I-OtherScientificTerm
into NN O O
which NN O O
the NN O O
utterances NN O O
naturally NN O O
aggregate NN O O
. NN O O

The NN O O
intentional NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
captures NN O O
the NN O O
discourse-relevant NN O I-OtherScientificTerm
purposes NN O I-OtherScientificTerm
, NN O O
expressed NN O O
in NN O O
each NN O O
of NN O O
the NN O O
linguistic NN O O
segments NN O O
as NN O O
well NN O O
as NN O O
relationships NN O O
among NN O O
them NN O O
. NN O O

The NN O O
attentional NN O I-OtherScientificTerm
state NN O I-OtherScientificTerm
is NN O O
an NN O O
abstraction NN O O
of NN O O
the NN O O
focus NN O O
of NN O O
attention NN O O
of NN O O
the NN O O
participants NN O O
as NN O O
the NN O O
discourse NN O I-OtherScientificTerm
unfolds NN O O
. NN O O

The NN O O
attentional NN O I-OtherScientificTerm
state NN O I-OtherScientificTerm
, NN O O
being NN O O
dynamic NN O O
, NN O O
records NN O O
the NN O O
objects NN O O
, NN O O
properties NN O O
, NN O O
and NN O O
relations NN O O
that NN O O
are NN O O
salient NN O O
at NN O O
each NN O O
point NN O O
of NN O O
the NN O O
discourse NN O I-OtherScientificTerm
. NN O O

The NN O O
distinction NN O O
among NN O O
these NN O O
components NN O O
is NN O O
essential NN O O
to NN O O
provide NN O O
an NN O O
adequate NN O O
explanation NN O O
of NN O O
such NN O O
discourse NN O I-OtherScientificTerm
phenomena NN O I-OtherScientificTerm
as NN O O
cue NN O I-OtherScientificTerm
phrases NN O I-OtherScientificTerm
, NN O O
referring NN O I-OtherScientificTerm
expressions NN O I-OtherScientificTerm
, NN O O
and NN O O
interruptions NN O I-OtherScientificTerm
. NN O O

The NN O O
theory NN O I-Method
of NN O I-Method
attention NN O I-Method
, NN O I-Method
intention NN O I-Method
, NN O I-Method
and NN O I-Method
aggregation NN O I-Method
of NN O I-Method
utterances NN O I-Method
is NN O O
illustrated NN O O
in NN O O
the NN O O
paper NN O O
with NN O O
a NN O O
number NN O O
of NN O O
example NN O O
discourses NN O I-OtherScientificTerm
. NN O O

Various NN O O
properties NN O O
of NN O O
discourse NN O I-OtherScientificTerm
are NN O O
described NN O O
, NN O O
and NN O O
explanations NN O O
for NN O O
the NN O O
behaviour NN O O
of NN O O
cue NN O I-OtherScientificTerm
phrases NN O I-OtherScientificTerm
, NN O O
referring NN O I-OtherScientificTerm
expressions NN O I-OtherScientificTerm
, NN O O
and NN O O
interruptions NN O I-OtherScientificTerm
are NN O O
explored NN O O
. NN O O

This NN O O
theory NN O I-Generic
provides NN O O
a NN O O
framework NN O O
for NN O O
describing NN O O
the NN O O
processing NN O O
of NN O O
utterances NN O O
in NN O O
a NN O O
discourse NN O I-OtherScientificTerm
. NN O O

Discourse NN O I-Method
processing NN O I-Method
requires NN O O
recognizing NN O O
how NN O O
the NN O O
utterances NN O O
of NN O O
the NN O O
discourse NN O I-OtherScientificTerm
aggregate NN O O
into NN O O
segments NN O O
, NN O O
recognizing NN O O
the NN O O
intentions NN O O
expressed NN O O
in NN O O
the NN O O
discourse NN O I-OtherScientificTerm
and NN O O
the NN O O
relationships NN O O
among NN O O
intentions NN O O
, NN O O
and NN O O
tracking NN O O
the NN O O
discourse NN O I-OtherScientificTerm
through NN O O
the NN O O
operation NN O O
of NN O O
the NN O O
mechanisms NN O O
associated NN O O
with NN O O
attentional NN O I-OtherScientificTerm
state NN O I-OtherScientificTerm
. NN O O

This NN O O
processing NN O I-Generic
description NN O O
specifies NN O O
in NN O O
these NN O O
recognition NN O I-Task
tasks NN O I-Task
the NN O O
role NN O O
of NN O O
information NN O O
from NN O O
the NN O O
discourse NN O I-OtherScientificTerm
and NN O O
from NN O O
the NN O O
participants NN O O
' NN O O
knowledge NN O O
of NN O O
the NN O O
domain NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
examine NN O O
the NN O O
relationship NN O O
between NN O O
the NN O O
two NN O O
grammatical NN O I-Method
formalisms NN O I-Method
: NN O O
Tree NN O I-Method
Adjoining NN O I-Method
Grammars NN O I-Method
and NN O O
Head NN O I-Method
Grammars NN O I-Method
. NN O O

We NN O O
briefly NN O O
investigate NN O O
the NN O O
weak NN O O
equivalence NN O O
of NN O O
the NN O O
two NN O O
formalisms NN O I-Generic
. NN O O

We NN O O
then NN O O
turn NN O O
to NN O O
a NN O O
discussion NN O O
comparing NN O O
the NN O O
linguistic NN O I-OtherScientificTerm
expressiveness NN O I-OtherScientificTerm
of NN O O
the NN O O
two NN O O
formalisms NN O I-Generic
. NN O O

-DOCSTART- NN O O

We NN O O
provide NN O O
a NN O O
unified NN O O
account NN O O
of NN O O
sentence-level NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
text-level NN O I-OtherScientificTerm
anaphora NN O I-OtherScientificTerm
within NN O O
the NN O O
framework NN O O
of NN O O
a NN O O
dependency-based NN O I-Method
grammar NN O I-Method
model NN O I-Method
. NN O O

Criteria NN O I-Generic
for NN O O
anaphora NN O I-Task
resolution NN O I-Task
within NN O I-Task
sentence NN O I-Task
boundaries NN O I-Task
rephrase NN O O
major NN O O
concepts NN O O
from NN O O
GB NN O I-Method
's NN O I-Method
binding NN O I-Method
theory NN O I-Method
, NN O O
while NN O O
those NN O I-Generic
for NN O O
text-level NN O I-OtherScientificTerm
anaphora NN O I-OtherScientificTerm
incorporate NN O O
an NN O O
adapted NN O O
version NN O O
of NN O O
a NN O O
Grosz-Sidner-style NN O I-Method
focus NN O I-Method
model NN O I-Method
. NN O O

-DOCSTART- NN O O

Coedition NN O I-Task
of NN O O
a NN O O
natural NN O I-Material
language NN O I-Material
text NN O I-Material
and NN O O
its NN O O
representation NN O O
in NN O O
some NN O O
interlingual NN O O
form NN O O
seems NN O O
the NN O O
best NN O O
and NN O O
simplest NN O O
way NN O O
to NN O O
share NN O O
text NN O I-OtherScientificTerm
revision NN O I-OtherScientificTerm
across NN O O
languages NN O I-Material
. NN O O

For NN O O
various NN O O
reasons NN O O
, NN O O
UNL NN O I-OtherScientificTerm
graphs NN O I-OtherScientificTerm
are NN O O
the NN O O
best NN O O
candidates NN O O
in NN O O
this NN O O
context NN O O
. NN O O

We NN O O
are NN O O
developing NN O O
a NN O O
prototype NN O O
where NN O O
, NN O O
in NN O O
the NN O O
simplest NN O O
sharing NN O O
scenario NN O O
, NN O O
naive NN O O
users NN O O
interact NN O O
directly NN O O
with NN O O
the NN O O
text NN O O
in NN O O
their NN O O
language NN O O
-LRB- NN O O
L0 NN O O
-RRB- NN O O
, NN O O
and NN O O
indirectly NN O O
with NN O O
the NN O O
associated NN O O
graph NN O I-OtherScientificTerm
. NN O O

The NN O O
modified NN O O
graph NN O I-OtherScientificTerm
is NN O O
then NN O O
sent NN O O
to NN O O
the NN O O
UNL-L0 NN O I-Method
deconverter NN O I-Method
and NN O O
the NN O O
result NN O O
shown NN O O
. NN O O

If NN O O
is NN O O
is NN O O
satisfactory NN O O
, NN O O
the NN O O
errors NN O O
were NN O O
probably NN O O
due NN O O
to NN O O
the NN O O
graph NN O I-OtherScientificTerm
, NN O O
not NN O O
to NN O O
the NN O O
deconverter NN O I-Method
, NN O O
and NN O O
the NN O O
graph NN O I-OtherScientificTerm
is NN O O
sent NN O O
to NN O O
deconverters NN O I-Method
in NN O O
other NN O O
languages NN O O
. NN O O

Versions NN O O
in NN O O
some NN O O
other NN O O
languages NN O O
known NN O O
by NN O O
the NN O O
user NN O O
may NN O O
be NN O O
displayed NN O O
, NN O O
so NN O O
that NN O O
improvement NN O O
sharing NN O O
is NN O O
visible NN O O
and NN O O
encouraging NN O O
. NN O O

As NN O O
new NN O O
versions NN O O
are NN O O
added NN O O
with NN O O
appropriate NN O O
tags NN O O
and NN O O
attributes NN O O
in NN O O
the NN O O
original NN O I-Material
multilingual NN O I-Material
document NN O I-Material
, NN O O
nothing NN O O
is NN O O
ever NN O O
lost NN O O
, NN O O
and NN O O
cooperative NN O O
working NN O O
on NN O O
a NN O O
document NN O O
is NN O O
rendered NN O O
feasible NN O O
. NN O O

On NN O O
the NN O O
internal NN O O
side NN O O
, NN O O
liaisons NN O I-OtherScientificTerm
are NN O O
established NN O O
between NN O O
elements NN O O
of NN O O
the NN O O
text NN O O
and NN O O
the NN O O
graph NN O O
by NN O O
using NN O O
broadly NN O O
available NN O O
resources NN O I-Generic
such NN O O
as NN O O
a NN O O
LO-English NN O I-Material
or NN O I-Material
better NN O I-Material
a NN O I-Material
L0-UNL NN O I-Material
dictionary NN O I-Material
, NN O O
a NN O O
morphosyntactic NN O I-Method
parser NN O I-Method
of NN O I-Method
L0 NN O I-Method
, NN O O
and NN O O
a NN O O
canonical NN O I-OtherScientificTerm
graph2tree NN O I-OtherScientificTerm
transformation NN O I-OtherScientificTerm
. NN O O

Establishing NN O O
a NN O O
`` NN O O
best NN O O
'' NN O O
correspondence NN O O
between NN O O
the NN O O
'' NN O O
UNL-tree NN O I-OtherScientificTerm
+ NN O I-OtherScientificTerm
L0 NN O I-OtherScientificTerm
'' NN O O
and NN O O
the NN O O
'' NN O O
MS-L0 NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
'' NN O O
, NN O O
a NN O O
lattice NN O I-OtherScientificTerm
, NN O O
may NN O O
be NN O O
done NN O O
using NN O O
the NN O O
dictionary NN O I-Material
and NN O O
trying NN O O
to NN O O
align NN O O
the NN O O
tree NN O O
and NN O O
the NN O O
selected NN O O
trajectory NN O O
with NN O O
as NN O O
few NN O O
crossing NN O I-OtherScientificTerm
liaisons NN O I-OtherScientificTerm
as NN O O
possible NN O O
. NN O O

A NN O O
central NN O O
goal NN O O
of NN O O
this NN O O
research NN O O
is NN O O
to NN O O
merge NN O O
approaches NN O O
from NN O O
pivot NN O I-Task
MT NN O I-Task
, NN O O
interactive NN O I-Task
MT NN O I-Task
, NN O O
and NN O O
multilingual NN O I-Task
text NN O I-Task
authoring NN O I-Task
. NN O O

-DOCSTART- NN O O

The NN O O
reality NN O O
of NN O O
analogies NN O I-Task
between NN O I-Task
words NN O I-Task
is NN O O
refuted NN O O
by NN O O
noone NN O O
-LRB- NN O O
e.g. NN O O
, NN O O
I NN O O
walked NN O O
is NN O O
to NN O O
to NN O O
walk NN O O
as NN O O
I NN O O
laughed NN O O
is NN O O
to NN O O
to NN O O
laugh NN O O
, NN O O
noted NN O O
I NN O O
walked NN O O
: NN O O
to NN O O
walk NN O O
: NN O O
: NN O O
I NN O O
laughed NN O O
: NN O O
to NN O O
laugh NN O O
-RRB- NN O O
. NN O O

But NN O O
computational NN O O
linguists NN O O
seem NN O O
to NN O O
be NN O O
quite NN O O
dubious NN O O
about NN O O
analogies NN O I-Task
between NN O I-Task
sentences NN O I-Task
: NN O O
they NN O O
would NN O O
not NN O O
be NN O O
enough NN O O
numerous NN O O
to NN O O
be NN O O
of NN O O
any NN O O
use NN O O
. NN O O

We NN O O
report NN O O
experiments NN O O
conducted NN O O
on NN O O
a NN O O
multilingual NN O I-Material
corpus NN O I-Material
to NN O O
estimate NN O O
the NN O O
number NN O O
of NN O O
analogies NN O I-OtherScientificTerm
among NN O O
the NN O O
sentences NN O O
that NN O O
it NN O O
contains NN O O
. NN O O

We NN O O
give NN O O
two NN O O
estimates NN O O
, NN O O
a NN O O
lower NN O O
one NN O O
and NN O O
a NN O O
higher NN O O
one NN O O
. NN O O

As NN O O
an NN O O
analogy NN O I-OtherScientificTerm
must NN O O
be NN O O
valid NN O O
on NN O O
the NN O O
level NN O O
of NN O O
form NN O O
as NN O O
well NN O O
as NN O O
on NN O O
the NN O O
level NN O O
of NN O O
meaning NN O O
, NN O O
we NN O O
relied NN O O
on NN O O
the NN O O
idea NN O O
that NN O O
translation NN O O
should NN O O
preserve NN O O
meaning NN O O
to NN O O
test NN O O
for NN O O
similar NN O O
meanings NN O O
. NN O O

-DOCSTART- NN O O

Our NN O O
goal NN O O
is NN O O
to NN O O
learn NN O O
a NN O O
Mahalanobis NN O I-Task
distance NN O I-Task
by NN O O
minimizing NN O O
a NN O O
loss NN O I-OtherScientificTerm
defined NN O O
on NN O O
the NN O O
weighted NN O I-OtherScientificTerm
sum NN O I-OtherScientificTerm
of NN O O
the NN O O
precision NN O I-Metric
at NN O O
different NN O O
ranks NN O I-OtherScientificTerm
. NN O O

Our NN O O
core NN O O
motivation NN O O
is NN O O
that NN O O
minimizing NN O O
a NN O O
weighted NN O I-OtherScientificTerm
rank NN O I-OtherScientificTerm
loss NN O I-OtherScientificTerm
is NN O O
a NN O O
natural NN O O
criterion NN O O
for NN O O
many NN O O
problems NN O O
in NN O O
computer NN O I-Task
vision NN O I-Task
such NN O O
as NN O O
person NN O I-Task
re-identification NN O I-Task
. NN O O

We NN O O
propose NN O O
a NN O O
novel NN O O
metric NN O I-Method
learning NN O I-Method
formulation NN O I-Method
called NN O O
Weighted NN O I-Method
Approximate NN O I-Method
Rank NN O I-Method
Component NN O I-Method
Analysis NN O I-Method
-LRB- NN O I-Method
WARCA NN O I-Method
-RRB- NN O I-Method
. NN O O

We NN O O
then NN O O
derive NN O O
a NN O O
scalable NN O O
stochastic NN O I-Method
gradient NN O I-Method
descent NN O I-Method
algorithm NN O I-Method
for NN O O
the NN O O
resulting NN O O
learning NN O I-Generic
problem NN O I-Generic
. NN O O

We NN O O
also NN O O
derive NN O O
an NN O O
efficient NN O O
non-linear NN O I-Method
extension NN O I-Method
of NN O I-Method
WARCA NN O B-Method
by NN O O
using NN O O
the NN O O
kernel NN O I-Method
trick NN O I-Method
. NN O O

Kernel NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
embedding NN O I-OtherScientificTerm
decouples NN O O
the NN O O
training NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
prediction NN O I-OtherScientificTerm
costs NN O I-OtherScientificTerm
from NN O O
the NN O O
data NN O I-OtherScientificTerm
dimension NN O I-OtherScientificTerm
and NN O O
enables NN O O
us NN O O
to NN O O
plug NN O O
inarbitrary NN O I-Method
distance NN O I-Method
measures NN O I-Method
which NN O O
are NN O O
more NN O O
natural NN O O
for NN O O
the NN O O
features NN O I-OtherScientificTerm
. NN O O

We NN O O
also NN O O
address NN O O
a NN O O
more NN O O
general NN O O
problem NN O O
of NN O O
matrix NN O I-OtherScientificTerm
rank NN O I-OtherScientificTerm
degeneration NN O I-OtherScientificTerm
& NN O O
non-isolated NN O I-OtherScientificTerm
minima NN O I-OtherScientificTerm
in NN O O
the NN O O
low-rank NN O I-Task
matrix NN O I-Task
optimization NN O I-Task
by NN O O
using NN O O
new NN O O
type NN O O
of NN O O
regularizer NN O I-Method
which NN O O
approximately NN O O
enforces NN O O
the NN O O
or-thonormality NN O I-OtherScientificTerm
of NN O O
the NN O O
learned NN O I-OtherScientificTerm
matrix NN O I-OtherScientificTerm
very NN O O
efficiently NN O O
. NN O O

We NN O O
validate NN O O
this NN O O
new NN O O
method NN O I-Generic
on NN O O
nine NN O O
standard NN O O
person NN O I-Material
re-identification NN O I-Material
datasets NN O I-Material
including NN O O
two NN O O
large NN O O
scale NN O I-Material
Market-1501 NN O I-Material
and NN O O
CUHK03 NN O I-Material
datasets NN O I-Material
and NN O O
show NN O O
that NN O O
we NN O O
improve NN O O
upon NN O O
the NN O O
current NN O O
state-of-the-art NN O O
methods NN O O
on NN O O
all NN O O
of NN O O
them NN O I-Generic
. NN O O

-DOCSTART- NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
discuss NN O O
language NN O I-Method
model NN O I-Method
adaptation NN O I-Method
methods NN O I-Method
given NN O O
a NN O O
word NN O I-OtherScientificTerm
list NN O I-OtherScientificTerm
and NN O O
a NN O O
raw NN O I-Material
corpus NN O I-Material
. NN O O

In NN O O
this NN O O
situation NN O O
, NN O O
the NN O O
general NN O O
method NN O I-Generic
is NN O O
to NN O O
segment NN O O
the NN O O
raw NN O I-Material
corpus NN O I-Material
automatically NN O O
using NN O O
a NN O O
word NN O I-OtherScientificTerm
list NN O I-OtherScientificTerm
, NN O O
correct NN O O
the NN O O
output NN O O
sentences NN O O
by NN O O
hand NN O O
, NN O O
and NN O O
build NN O O
a NN O O
model NN O I-Generic
from NN O O
the NN O O
segmented NN O I-Material
corpus NN O I-Material
. NN O O

In NN O O
this NN O O
sentence-by-sentence NN O I-Method
error NN O I-Method
correction NN O I-Method
method NN O I-Method
, NN O O
however NN O O
, NN O O
the NN O O
annotator NN O O
encounters NN O O
grammatically NN O O
complicated NN O O
positions NN O O
and NN O O
this NN O O
results NN O O
in NN O O
a NN O O
decrease NN O O
of NN O O
productivity NN O O
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
propose NN O O
to NN O O
concentrate NN O O
on NN O O
correcting NN O O
the NN O O
positions NN O O
in NN O O
which NN O O
the NN O O
words NN O O
in NN O O
the NN O O
list NN O O
appear NN O O
by NN O O
taking NN O O
a NN O O
word NN O O
as NN O O
a NN O O
correction NN O O
unit NN O O
. NN O O

This NN O O
method NN O I-Generic
allows NN O O
us NN O O
to NN O O
avoid NN O O
these NN O O
problems NN O O
and NN O O
go NN O O
directly NN O O
to NN O O
capturing NN O O
the NN O O
statistical NN O O
behavior NN O O
of NN O O
specific NN O O
words NN O O
in NN O O
the NN O O
application NN O O
. NN O O

In NN O O
the NN O O
experiments NN O O
, NN O O
we NN O O
used NN O O
a NN O O
variety NN O O
of NN O O
methods NN O I-Generic
for NN O O
preparing NN O I-Task
a NN O I-Task
segmented NN O I-Task
corpus NN O I-Task
and NN O O
compared NN O O
the NN O O
language NN O I-Method
models NN O I-Method
by NN O O
their NN O O
speech NN O I-Metric
recognition NN O I-Metric
accuracies NN O I-Metric
. NN O O

The NN O O
results NN O O
showed NN O O
the NN O O
advantages NN O O
of NN O O
our NN O O
method NN O I-Generic
. NN O O

-DOCSTART- NN O O

Many NN O O
practical NN O O
modeling NN O I-Task
problems NN O I-Task
involve NN O O
discrete NN O I-Material
data NN O I-Material
that NN O O
are NN O O
best NN O O
represented NN O O
as NN O O
draws NN O O
from NN O O
multinomial NN O I-Method
or NN O I-Method
categorical NN O I-Method
distributions NN O I-Method
. NN O O

For NN O O
example NN O O
, NN O O
nucleotides NN O I-Material
in NN O I-Material
a NN O I-Material
DNA NN O I-Material
sequence NN O I-Material
, NN O O
children NN O O
's NN O O
names NN O O
in NN O O
a NN O O
given NN O O
state NN O O
and NN O O
year NN O O
, NN O O
and NN O O
text NN O I-Material
documents NN O I-Material
are NN O O
all NN O O
commonly NN O O
modeled NN O O
with NN O O
multinomial NN O I-Method
distributions NN O I-Method
. NN O O

In NN O O
all NN O O
of NN O O
these NN O O
cases NN O O
, NN O O
we NN O O
expect NN O O
some NN O O
form NN O O
of NN O O
dependency NN O O
between NN O O
the NN O O
draws NN O O
: NN O O
the NN O O
nucleotide NN O I-OtherScientificTerm
at NN O O
one NN O O
position NN O O
in NN O O
the NN O O
DNA NN O I-OtherScientificTerm
strand NN O I-OtherScientificTerm
may NN O O
depend NN O O
on NN O O
the NN O O
preceding NN O I-OtherScientificTerm
nucleotides NN O I-OtherScientificTerm
, NN O O
children NN O O
's NN O O
names NN O O
are NN O O
highly NN O O
correlated NN O O
from NN O O
year NN O O
to NN O O
year NN O O
, NN O O
and NN O O
topics NN O O
in NN O O
text NN O O
may NN O O
be NN O O
correlated NN O O
and NN O O
dynamic NN O O
. NN O O

These NN O O
dependencies NN O O
are NN O O
not NN O O
naturally NN O O
captured NN O O
by NN O O
the NN O O
typical NN O O
Dirichlet-multinomial NN O I-Method
formulation NN O I-Method
. NN O O

Here NN O O
, NN O O
we NN O O
leverage NN O O
a NN O O
logistic NN O I-Method
stick-breaking NN O I-Method
representation NN O I-Method
and NN O O
recent NN O O
innovations NN O O
in NN O O
Pólya-gamma NN O I-Task
augmentation NN O I-Task
to NN O O
reformu-late NN O O
the NN O O
multinomial NN O I-Method
distribution NN O I-Method
in NN O O
terms NN O O
of NN O O
latent NN O I-OtherScientificTerm
variables NN O I-OtherScientificTerm
with NN O O
jointly NN O I-OtherScientificTerm
Gaussian NN O I-OtherScientificTerm
likelihoods NN O I-OtherScientificTerm
, NN O O
enabling NN O O
us NN O O
to NN O O
take NN O O
advantage NN O O
of NN O O
a NN O O
host NN O O
of NN O O
Bayesian NN O I-Method
inference NN O I-Method
techniques NN O I-Method
for NN O O
Gaussian NN O I-Method
models NN O I-Method
with NN O O
minimal NN O I-OtherScientificTerm
overhead NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

MINPRAN NN O I-Method
, NN O O
a NN O O
new NN O O
robust NN O I-Method
operator NN O I-Method
, NN O O
nds NN O O
good NN O O
ts NN O O
in NN O O
data NN O O
sets NN O O
where NN O O
more NN O O
than NN O O
50 NN O O
% NN O O
of NN O O
the NN O O
points NN O O
are NN O O
outliers NN O O
. NN O O

Unlike NN O O
other NN O O
techniques NN O I-Generic
that NN O O
handle NN O O
large NN O I-Metric
outlier NN O I-Metric
percentages NN O I-Metric
, NN O O
MINPRAN NN O I-Method
does NN O O
not NN O O
rely NN O O
on NN O O
a NN O O
known NN O O
error NN O I-OtherScientificTerm
bound NN O I-OtherScientificTerm
for NN O O
the NN O O
good NN O O
data NN O O
. NN O O

Instead NN O O
it NN O I-Generic
assumes NN O O
that NN O O
the NN O O
bad NN O O
data NN O O
are NN O O
randomly NN O O
-LRB- NN O O
uniformly NN O O
-RRB- NN O O
distributed NN O O
within NN O O
the NN O O
dynamic NN O I-OtherScientificTerm
range NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
the NN O I-OtherScientificTerm
sensor NN O I-OtherScientificTerm
. NN O O

Based NN O O
on NN O O
this NN O O
, NN O O
MINPRAN NN O I-Method
uses NN O O
random NN O I-Method
sampling NN O I-Method
to NN O O
search NN O O
for NN O O
the NN O O
t NN O O
and NN O O
the NN O O
number NN O O
of NN O O
inliers NN O O
to NN O O
the NN O O
t NN O O
that NN O O
are NN O O
least NN O O
likely NN O O
to NN O O
have NN O O
occurred NN O O
randomly NN O O
. NN O O

It NN O I-Generic
runs NN O O
in NN O O
time NN O O
O NN O O
-LRB- NN O O
N NN O O
2 NN O O
+ NN O O
SN NN O O
log NN O O
N NN O O
-RRB- NN O O
, NN O O
where NN O O
S NN O O
is NN O O
the NN O O
number NN O O
of NN O O
random NN O O
samples NN O O
and NN O O
N NN O O
is NN O O
the NN O O
number NN O O
of NN O O
data NN O O
points NN O O
. NN O O

We NN O O
demonstrate NN O O
analytically NN O O
that NN O O
MINPRAN NN O I-Method
distinguishes NN O O
good NN O O
ts NN O O
from NN O O
ts NN O O
to NN O O
random NN O O
data NN O O
, NN O O
and NN O O
that NN O O
MINPRAN NN O I-Method
nds NN O O
accurate NN O O
ts NN O O
and NN O O
nearly NN O O
the NN O O
correct NN O O
number NN O O
of NN O O
inliers NN O O
, NN O O
regardless NN O O
of NN O O
the NN O O
percentage NN O I-Metric
of NN O I-Metric
true NN O I-Metric
inliers NN O I-Metric
. NN O O

MINPRAN NN O I-Method
's NN O O
properties NN O O
are NN O O
connrmed NN O O
experimentally NN O O
on NN O O
synthetic NN O I-Material
data NN O I-Material
and NN O O
compare NN O O
favorably NN O O
to NN O O
least NN O I-Method
median NN O I-Method
of NN O I-Method
squares NN O I-Method
. NN O O

Related NN O O
work NN O O
applies NN O O
MINPRAN NN O I-Method
to NN O O
complex NN O I-OtherScientificTerm
range NN O I-OtherScientificTerm
and NN O O
intensity NN O I-Material
data NN O I-Material
23 NN O O
-RSB- NN O O
. NN O O

-DOCSTART- NN O O

Metagrammatical NN O I-Method
formalisms NN O I-Method
that NN O O
combine NN O O
context-free NN O I-OtherScientificTerm
phrase NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
rules NN O I-OtherScientificTerm
and NN O O
metarules NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
MPS NN O I-OtherScientificTerm
grammars NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
allow NN O O
concise NN O O
statement NN O O
of NN O O
generalizations NN O O
about NN O O
the NN O O
syntax NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
natural NN O I-OtherScientificTerm
languages NN O I-OtherScientificTerm
. NN O O

Unconstrained NN O I-Method
MPS NN O I-Method
grammars NN O I-Method
, NN O O
unfortunately NN O O
, NN O O
are NN O O
not NN O O
computationally NN O O
safe NN O O
. NN O O

We NN O O
evaluate NN O O
several NN O O
proposals NN O O
for NN O O
constraining NN O O
them NN O I-Generic
, NN O O
basing NN O O
our NN O O
assessment NN O O
on NN O O
computational NN O I-Metric
tractability NN O I-Metric
and NN O I-Metric
explanatory NN O I-Metric
adequacy NN O I-Metric
. NN O O

We NN O O
show NN O O
that NN O O
none NN O O
of NN O O
them NN O I-Generic
satisfies NN O O
both NN O O
criteria NN O I-Generic
, NN O O
and NN O O
suggest NN O O
new NN O O
directions NN O O
for NN O O
research NN O O
on NN O O
alternative NN O O
metagrammatical NN O I-Method
formalisms NN O I-Method
. NN O O

-DOCSTART- NN O O

The NN O O
unique NN O O
properties NN O O
of NN O O
tree-adjoining NN O I-Method
grammars NN O I-Method
-LRB- NN O I-Method
TAG NN O I-Method
-RRB- NN O I-Method
present NN O O
a NN O O
challenge NN O O
for NN O O
the NN O O
application NN O O
of NN O O
TAGs NN O I-Method
beyond NN O O
the NN O O
limited NN O O
confines NN O O
of NN O O
syntax NN O I-OtherScientificTerm
, NN O O
for NN O O
instance NN O O
, NN O O
to NN O O
the NN O O
task NN O O
of NN O O
semantic NN O I-Task
interpretation NN O I-Task
or NN O O
automatic NN O I-Task
translation NN O I-Task
of NN O I-Task
natural NN O I-Task
language NN O I-Task
. NN O O

We NN O O
present NN O O
a NN O O
variant NN O I-Method
of NN O I-Method
TAGs NN O B-Method
, NN O O
called NN O O
synchronous NN O I-Method
TAGs NN O I-Method
, NN O O
which NN O O
characterize NN O O
correspondences NN O O
between NN O O
languages NN O O
. NN O O

The NN O O
formalism NN O O
's NN O O
intended NN O O
usage NN O O
is NN O O
to NN O O
relate NN O O
expressions NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
natural NN O I-OtherScientificTerm
languages NN O I-OtherScientificTerm
to NN O O
their NN O O
associated NN O O
semantics NN O I-OtherScientificTerm
represented NN O O
in NN O O
a NN O O
logical NN O I-OtherScientificTerm
form NN O I-OtherScientificTerm
language NN O I-OtherScientificTerm
, NN O O
or NN O O
to NN O O
their NN O O
translates NN O O
in NN O O
another NN O O
natural NN O I-OtherScientificTerm
language NN O I-OtherScientificTerm
; NN O O
in NN O O
summary NN O O
, NN O O
we NN O O
intend NN O O
it NN O O
to NN O O
allow NN O O
TAGs NN O I-Method
to NN O O
be NN O O
used NN O O
beyond NN O O
their NN O O
role NN O O
in NN O O
syntax NN O I-OtherScientificTerm
proper NN O I-OtherScientificTerm
. NN O O

We NN O O
discuss NN O O
the NN O O
application NN O O
of NN O O
synchronous NN O I-Method
TAGs NN O I-Method
to NN O O
concrete NN O O
examples NN O O
, NN O O
mentioning NN O O
primarily NN O O
in NN O O
passing NN O O
some NN O O
computational NN O O
issues NN O O
that NN O O
arise NN O O
in NN O O
its NN O O
interpretation NN O O
. NN O O

-DOCSTART- NN O O

A NN O O
model-based NN O I-Method
approach NN O I-Method
to NN O O
on-line NN O I-Task
cursive NN O I-Task
handwriting NN O I-Task
analysis NN O I-Task
and NN O I-Task
recognition NN O I-Task
is NN O O
presented NN O O
and NN O O
evaluated NN O O
. NN O O

In NN O O
this NN O O
model NN O I-Generic
, NN O O
on-line NN O I-Task
handwriting NN O I-Task
is NN O O
considered NN O O
as NN O O
a NN O O
modulation NN O O
of NN O O
a NN O O
simple NN O O
cycloidal NN O I-Task
pen NN O I-Task
motion NN O I-Task
, NN O O
described NN O O
by NN O O
two NN O O
coupled NN O O
oscillations NN O O
with NN O O
a NN O O
constant NN O I-OtherScientificTerm
linear NN O I-OtherScientificTerm
drift NN O I-OtherScientificTerm
along NN O O
the NN O O
line NN O O
of NN O O
the NN O O
writing NN O O
. NN O O

By NN O O
slow NN O O
modulations NN O O
of NN O O
the NN O O
amplitudes NN O O
and NN O O
phase NN O O
lags NN O O
of NN O O
the NN O O
two NN O O
oscillators NN O O
, NN O O
a NN O O
general NN O O
pen NN O I-OtherScientificTerm
trajectory NN O I-OtherScientificTerm
can NN O O
be NN O O
efficiently NN O O
encoded NN O O
. NN O O

These NN O O
parameters NN O O
are NN O O
then NN O O
quantized NN O O
into NN O O
a NN O O
small NN O O
number NN O O
of NN O O
values NN O O
without NN O O
altering NN O O
the NN O O
writing NN O I-OtherScientificTerm
intelligibility NN O I-OtherScientificTerm
. NN O O

A NN O O
general NN O O
procedure NN O O
for NN O O
the NN O O
estimation NN O O
and NN O O
quantization NN O O
of NN O O
these NN O O
cycloidal NN O I-OtherScientificTerm
motion NN O I-OtherScientificTerm
parameters NN O I-OtherScientificTerm
for NN O O
arbitrary NN O I-Material
handwriting NN O I-Material
is NN O O
presented NN O O
. NN O O

The NN O O
result NN O O
is NN O O
a NN O O
discrete NN O I-Method
motor NN O I-Method
control NN O I-Method
representation NN O I-Method
of NN O O
the NN O O
continuous NN O I-OtherScientificTerm
pen NN O I-OtherScientificTerm
motion NN O I-OtherScientificTerm
, NN O O
via NN O O
the NN O O
quantized NN O O
levels NN O O
of NN O O
the NN O O
model NN O O
parameters NN O O
. NN O O

This NN O O
motor NN O I-Method
control NN O I-Method
representation NN O I-Method
enables NN O O
successful NN O O
word NN O I-Task
spotting NN O I-Task
and NN O O
matching NN O I-Task
of NN O I-Task
cursive NN O I-Task
scripts NN O I-Task
. NN O O

Our NN O O
experiments NN O O
clearly NN O O
indicate NN O O
the NN O O
potential NN O O
of NN O O
this NN O O
dynamic NN O I-Method
representation NN O I-Method
for NN O O
complete NN O O
cursive NN O I-Task
handwriting NN O I-Task
recognition NN O I-Task
. NN O O

-DOCSTART- NN O O

In NN O O
the NN O O
Object NN O I-Task
Recognition NN O I-Task
task NN O I-Task
, NN O O
there NN O O
exists NN O O
a NN O O
di-chotomy NN O O
between NN O O
the NN O O
categorization NN O I-Task
of NN O I-Task
objects NN O I-Task
and NN O O
estimating NN O I-Task
object NN O I-Task
pose NN O I-Task
, NN O O
where NN O O
the NN O O
former NN O I-Generic
necessitates NN O O
a NN O O
view-invariant NN O I-Method
representation NN O I-Method
, NN O O
while NN O O
the NN O O
latter NN O I-Generic
requires NN O O
a NN O O
representation NN O I-Generic
capable NN O O
of NN O O
capturing NN O O
pose NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
over NN O O
different NN O O
categories NN O O
of NN O O
objects NN O O
. NN O O

With NN O O
the NN O O
rise NN O O
of NN O O
deep NN O I-Method
archi-tectures NN O I-Method
, NN O O
the NN O O
prime NN O O
focus NN O O
has NN O O
been NN O O
on NN O O
object NN O I-Task
category NN O I-Task
recognition NN O I-Task
. NN O O

Deep NN O I-Method
learning NN O I-Method
methods NN O I-Method
have NN O O
achieved NN O O
wide NN O O
success NN O O
in NN O O
this NN O O
task NN O I-Generic
. NN O O

In NN O O
contrast NN O O
, NN O O
object NN O I-Task
pose NN O I-Task
estimation NN O I-Task
using NN O O
these NN O O
approaches NN O I-Generic
has NN O O
received NN O O
relatively NN O O
less NN O O
attention NN O O
. NN O O

In NN O O
this NN O O
work NN O O
, NN O O
we NN O O
study NN O O
how NN O O
Convolutional NN O I-Method
Neural NN O I-Method
Networks NN O I-Method
-LRB- NN O I-Method
CNN NN O I-Method
-RRB- NN O I-Method
architectures NN O I-Method
can NN O O
be NN O O
adapted NN O O
to NN O O
the NN O O
task NN O O
of NN O O
simultaneous NN O O
object NN O I-Task
recognition NN O I-Task
and NN O O
pose NN O I-Task
estimation NN O I-Task
. NN O O

We NN O O
investigate NN O O
and NN O O
analyze NN O O
the NN O O
layers NN O I-OtherScientificTerm
of NN O O
various NN O O
CNN NN O I-Method
models NN O I-Method
and NN O O
extensively NN O O
compare NN O O
between NN O O
them NN O I-Generic
with NN O O
the NN O O
goal NN O O
of NN O O
discovering NN O O
how NN O O
the NN O O
layers NN O I-Method
of NN O I-Method
distributed NN O I-Method
representations NN O I-Method
within NN O O
CNNs NN O I-Method
represent NN O O
object NN O I-OtherScientificTerm
pose NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
and NN O O
how NN O O
this NN O I-Generic
contradicts NN O O
with NN O O
object NN O I-Method
category NN O I-Method
representations NN O I-Method
. NN O O

We NN O O
extensively NN O O
experiment NN O O
on NN O O
two NN O O
recent NN O O
large NN O O
and NN O O
challenging NN O O
multi-view NN O I-Material
datasets NN O I-Material
and NN O O
we NN O O
achieve NN O O
better NN O O
than NN O O
the NN O O
state-of-the-art NN O O
. NN O O

-DOCSTART- NN O O

In NN O O
this NN O O
paper NN O O
we NN O O
present NN O O
our NN O O
recent NN O O
work NN O O
on NN O O
harvesting NN O O
English-Chinese NN O I-Material
bitexts NN O I-Material
of NN O O
the NN O O
laws NN O O
of NN O O
Hong NN O O
Kong NN O O
from NN O O
the NN O O
Web NN O O
and NN O O
aligning NN O O
them NN O I-Generic
to NN O O
the NN O O
subparagraph NN O O
level NN O O
via NN O O
utilizing NN O O
the NN O O
numbering NN O I-Method
system NN O I-Method
in NN O O
the NN O O
legal NN O I-OtherScientificTerm
text NN O I-OtherScientificTerm
hierarchy NN O I-OtherScientificTerm
. NN O O

Basic NN O O
methodology NN O O
and NN O O
practical NN O O
techniques NN O O
are NN O O
reported NN O O
in NN O O
detail NN O O
. NN O O

The NN O O
resultant NN O O
bilingual NN O I-Material
corpus NN O I-Material
, NN O O
10.4 NN O O
M NN O O
English NN O O
words NN O O
and NN O O
18.3 NN O O
M NN O O
Chinese NN O O
characters NN O O
, NN O O
is NN O O
an NN O O
authoritative NN O O
and NN O O
comprehensive NN O O
text NN O O
collection NN O O
covering NN O O
the NN O O
specific NN O O
and NN O O
special NN O O
domain NN O O
of NN O O
HK NN O O
laws NN O O
. NN O O

It NN O I-Generic
is NN O O
particularly NN O O
valuable NN O O
to NN O O
empirical NN O I-Task
MT NN O I-Task
research NN O I-Task
. NN O O

This NN O O
piece NN O O
of NN O O
work NN O O
has NN O O
also NN O O
laid NN O O
a NN O O
foundation NN O O
for NN O O
exploring NN O O
and NN O O
harvesting NN O O
English-Chinese NN O I-Material
bitexts NN O I-Material
in NN O O
a NN O O
larger NN O O
volume NN O O
from NN O O
the NN O O
Web NN O O
. NN O O

-DOCSTART- NN O O

Light NN O I-OtherScientificTerm
fields NN O I-OtherScientificTerm
are NN O O
image-based NN O I-Method
representations NN O I-Method
that NN O O
use NN O O
densely NN O I-OtherScientificTerm
sampled NN O I-OtherScientificTerm
rays NN O I-OtherScientificTerm
as NN O O
a NN O O
scene NN O I-OtherScientificTerm
description NN O I-OtherScientificTerm
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
explore NN O O
geometric NN O I-Task
structures NN O I-Task
of NN O I-Task
3D NN O I-Task
lines NN O I-Task
in NN O O
ray NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
for NN O O
improving NN O O
light NN O I-Task
field NN O I-Task
triangulation NN O I-Task
and NN O O
stereo NN O I-Task
matching NN O I-Task
. NN O O

The NN O O
triangulation NN O I-Task
problem NN O I-Task
aims NN O O
to NN O O
fill NN O O
in NN O O
the NN O O
ray NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
with NN O O
continuous NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
non-overlapping NN O I-OtherScientificTerm
simplices NN O I-OtherScientificTerm
anchored NN O O
at NN O O
sampled NN O O
points NN O O
-LRB- NN O O
rays NN O O
-RRB- NN O O
. NN O O

Such NN O O
a NN O O
triangulation NN O I-OtherScientificTerm
provides NN O O
a NN O O
piecewise-linear NN O I-OtherScientificTerm
interpolant NN O I-OtherScientificTerm
useful NN O O
for NN O O
light NN O I-Task
field NN O I-Task
super-resolution NN O I-Task
. NN O O

We NN O O
show NN O O
that NN O O
the NN O O
light NN O I-OtherScientificTerm
field NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
is NN O O
largely NN O O
bi-linear NN O O
due NN O O
to NN O O
3D NN O I-OtherScientificTerm
line NN O I-OtherScientificTerm
segments NN O I-OtherScientificTerm
in NN O O
the NN O O
scene NN O O
, NN O O
and NN O O
direct NN O O
tri-angulation NN O O
of NN O O
these NN O O
bilinear NN O I-OtherScientificTerm
subspaces NN O I-OtherScientificTerm
leads NN O O
to NN O O
large NN O O
errors NN O O
. NN O O

We NN O O
instead NN O O
present NN O O
a NN O O
simple NN O O
but NN O O
effective NN O O
algorithm NN O O
to NN O O
first NN O O
map NN O O
bilinear NN O I-OtherScientificTerm
subspaces NN O I-OtherScientificTerm
to NN O O
line NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
and NN O O
then NN O O
apply NN O O
Constrained NN O I-Method
Delaunay NN O I-Method
Triangulation NN O I-Method
-LRB- NN O I-Method
CDT NN O I-Method
-RRB- NN O I-Method
. NN O O

Based NN O O
on NN O O
our NN O O
analysis NN O O
, NN O O
we NN O O
further NN O O
develop NN O O
a NN O O
novel NN O O
line-assisted NN O I-Method
graph-cut NN O I-Method
-LRB- NN O I-Method
LAGC NN O I-Method
-RRB- NN O I-Method
algorithm NN O I-Method
that NN O O
effectively NN O O
encodes NN O O
3D NN O I-OtherScientificTerm
line NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
into NN O O
light NN O I-Task
field NN O I-Task
stereo NN O I-Task
matching NN O I-Task
. NN O O

Experiments NN O O
on NN O O
synthetic NN O I-Material
and NN O I-Material
real NN O I-Material
data NN O I-Material
show NN O O
that NN O O
both NN O O
our NN O O
triangulation NN O I-Method
and NN O I-Method
LAGC NN O I-Method
algorithms NN O I-Method
outperform NN O O
state-of-the-art NN O I-Generic
solutions NN O I-Generic
in NN O O
accuracy NN O I-Metric
and NN O O
visual NN O I-Metric
quality NN O I-Metric
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
presents NN O O
a NN O O
phrase-based NN O I-Method
statistical NN O I-Method
machine NN O I-Method
translation NN O I-Method
method NN O I-Method
, NN O O
based NN O O
on NN O O
non-contiguous NN O I-Material
phrases NN O I-Material
, NN O O
i.e. NN O O
phrases NN O O
with NN O O
gaps NN O O
. NN O O

A NN O O
method NN O I-Generic
for NN O O
producing NN O O
such NN O O
phrases NN O I-Generic
from NN O O
a NN O O
word-aligned NN O I-Material
corpora NN O I-Material
is NN O O
proposed NN O O
. NN O O

A NN O O
statistical NN O I-Method
translation NN O I-Method
model NN O I-Method
is NN O O
also NN O O
presented NN O O
that NN O O
deals NN O O
such NN O O
phrases NN O I-Generic
, NN O O
as NN O O
well NN O O
as NN O O
a NN O O
training NN O I-Method
method NN O I-Method
based NN O O
on NN O O
the NN O O
maximization NN O I-Metric
of NN O I-Metric
translation NN O I-Metric
accuracy NN O I-Metric
, NN O O
as NN O O
measured NN O O
with NN O O
the NN O O
NIST NN O I-Metric
evaluation NN O I-Metric
metric NN O I-Metric
. NN O O

Translations NN O I-OtherScientificTerm
are NN O O
produced NN O O
by NN O O
means NN O O
of NN O O
a NN O O
beam-search NN O I-Method
decoder NN O I-Method
. NN O O

Experimental NN O O
results NN O O
are NN O O
presented NN O O
, NN O O
that NN O O
demonstrate NN O O
how NN O O
the NN O O
proposed NN O O
method NN O I-Generic
allows NN O O
to NN O O
better NN O O
generalize NN O O
from NN O O
the NN O O
training NN O O
data NN O O
. NN O O

-DOCSTART- NN O O

GLOSSER NN O I-Method
is NN O O
designed NN O O
to NN O O
support NN O O
reading NN O I-Task
and NN O I-Task
learning NN O I-Task
to NN O O
read NN O O
in NN O O
a NN O O
foreign NN O O
language NN O O
. NN O O

There NN O O
are NN O O
four NN O O
language NN O I-Generic
pairs NN O I-Generic
currently NN O O
supported NN O O
by NN O O
GLOSSER NN O I-Method
: NN O O
English-Bulgarian NN O I-Material
, NN O O
English-Estonian NN O I-Material
, NN O O
English-Hungarian NN O I-Material
and NN O O
French-Dutch NN O I-Material
. NN O O

The NN O O
program NN O I-Generic
is NN O O
operational NN O O
on NN O O
UNIX NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
Windows NN O I-OtherScientificTerm
'95 NN O I-OtherScientificTerm
platforms NN O I-OtherScientificTerm
, NN O O
and NN O O
has NN O O
undergone NN O O
a NN O O
pilot NN O O
user-study NN O I-OtherScientificTerm
. NN O O

A NN O O
demonstration NN O O
-LRB- NN O O
in NN O O
UNIX NN O O
-RRB- NN O O
for NN O O
Applied NN O I-Task
Natural NN O I-Task
Language NN O I-Task
Processing NN O I-Task
emphasizes NN O O
components NN O I-Generic
put NN O O
to NN O O
novel NN O O
technical NN O O
uses NN O O
in NN O O
intelligent NN O I-Task
computer-assisted NN O I-Task
morphological NN O I-Task
analysis NN O I-Task
-LRB- NN O I-Task
ICALL NN O I-Task
-RRB- NN O I-Task
, NN O O
including NN O O
disambiguated NN O I-Method
morphological NN O I-Method
analysis NN O I-Method
and NN O O
lemmatized NN O I-Method
indexing NN O I-Method
for NN O O
an NN O O
aligned NN O I-Material
bilingual NN O I-Material
corpus NN O I-Material
of NN O O
word NN O O
examples NN O O
. NN O O

-DOCSTART- NN O O

We NN O O
present NN O O
a NN O O
new NN O O
part-of-speech NN O I-Method
tagger NN O I-Method
that NN O O
demonstrates NN O O
the NN O O
following NN O O
ideas NN O O
: NN O O
-LRB- NN O O
i NN O O
-RRB- NN O O
explicit NN O O
use NN O O
of NN O O
both NN O O
preceding NN O O
and NN O O
following NN O O
tag NN O I-OtherScientificTerm
contexts NN O I-OtherScientificTerm
via NN O O
a NN O O
dependency NN O I-Method
network NN O I-Method
representation NN O I-Method
, NN O O
-LRB- NN O O
ii NN O O
-RRB- NN O O
broad NN O O
use NN O O
of NN O O
lexical NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
, NN O O
including NN O O
jointly NN O O
conditioning NN O O
on NN O O
multiple NN O I-OtherScientificTerm
consecutive NN O I-OtherScientificTerm
words NN O I-OtherScientificTerm
, NN O O
-LRB- NN O O
iii NN O O
-RRB- NN O O
effective NN O O
use NN O O
of NN O O
priors NN O I-OtherScientificTerm
in NN O I-OtherScientificTerm
conditional NN O I-OtherScientificTerm
loglinear NN O I-OtherScientificTerm
models NN O I-OtherScientificTerm
, NN O O
and NN O O
-LRB- NN O O
iv NN O O
-RRB- NN O O
fine-grained NN O I-Method
modeling NN O I-Method
of NN O I-Method
unknown NN O I-Method
word NN O I-Method
features NN O I-Method
. NN O O

Using NN O O
these NN O O
ideas NN O O
together NN O O
, NN O O
the NN O O
resulting NN O O
tagger NN O I-Method
gives NN O O
a NN O O
97.24 NN O O
% NN O O
accuracy NN O I-Metric
on NN O O
the NN O O
Penn NN O I-Material
Treebank NN O I-Material
WSJ NN O I-Material
, NN O O
an NN O O
error NN O I-Metric
reduction NN O O
of NN O O
4.4 NN O O
% NN O O
on NN O O
the NN O O
best NN O O
previous NN O O
single NN O O
automatically NN O O
learned NN O O
tagging NN O I-Task
result NN O O
. NN O O

-DOCSTART- NN O O

Person NN O I-Task
re-identification NN O I-Task
is NN O O
challenging NN O O
due NN O O
to NN O O
the NN O O
large NN O O
variations NN O O
of NN O O
pose NN O I-OtherScientificTerm
, NN O O
illumination NN O I-OtherScientificTerm
, NN O O
occlusion NN O I-OtherScientificTerm
and NN O O
camera NN O I-OtherScientificTerm
view NN O I-OtherScientificTerm
. NN O O

Owing NN O O
to NN O O
these NN O O
variations NN O O
, NN O O
the NN O O
pedestrian NN O I-Material
data NN O I-Material
is NN O O
distributed NN O O
as NN O O
highly-curved NN O I-Method
manifolds NN O I-Method
in NN O O
the NN O O
feature NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
, NN O O
despite NN O O
the NN O O
current NN O O
convolutional NN O I-Method
neural NN O I-Method
networks NN O I-Method
-LRB- NN O I-Method
CNN NN O I-Method
-RRB- NN O I-Method
's NN O O
capability NN O O
of NN O O
feature NN O I-OtherScientificTerm
extraction NN O I-OtherScientificTerm
. NN O O

However NN O O
, NN O O
the NN O O
distribution NN O O
is NN O O
unknown NN O O
, NN O O
so NN O O
it NN O O
is NN O O
difficult NN O O
to NN O O
use NN O O
the NN O O
geodesic NN O I-OtherScientificTerm
distance NN O I-OtherScientificTerm
when NN O O
comparing NN O O
two NN O O
samples NN O O
. NN O O

In NN O O
practice NN O O
, NN O O
the NN O O
current NN O O
deep NN O I-Method
embedding NN O I-Method
methods NN O I-Method
use NN O O
the NN O O
Euclidean NN O I-OtherScientificTerm
distance NN O I-OtherScientificTerm
for NN O O
the NN O O
training NN O O
and NN O O
test NN O O
. NN O O

On NN O O
the NN O O
other NN O O
hand NN O O
, NN O O
the NN O O
manifold NN O I-Method
learning NN O I-Method
methods NN O I-Method
suggest NN O O
to NN O O
use NN O O
the NN O O
Euclidean NN O I-OtherScientificTerm
distance NN O I-OtherScientificTerm
in NN O O
the NN O O
local NN O I-OtherScientificTerm
range NN O I-OtherScientificTerm
, NN O O
combining NN O O
with NN O O
the NN O O
graphical NN O I-OtherScientificTerm
relationship NN O I-OtherScientificTerm
between NN O O
samples NN O O
, NN O O
for NN O O
approximating NN O O
the NN O O
geodesic NN O I-OtherScientificTerm
distance NN O I-OtherScientificTerm
. NN O O

From NN O O
this NN O O
point NN O O
of NN O O
view NN O O
, NN O O
selecting NN O O
suitable NN O O
positive NN O O
-LRB- NN O O
i.e. NN O O
intra-class NN O O
-RRB- NN O O
training NN O O
samples NN O O
within NN O O
a NN O O
local NN O I-OtherScientificTerm
range NN O I-OtherScientificTerm
is NN O O
critical NN O O
for NN O O
training NN O O
the NN O O
CNN NN O I-OtherScientificTerm
embedding NN O I-OtherScientificTerm
, NN O O
especially NN O O
when NN O O
the NN O O
data NN O I-Generic
has NN O O
large NN O O
intra-class NN O I-OtherScientificTerm
variations NN O I-OtherScientificTerm
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
propose NN O O
a NN O O
novel NN O O
moderate NN O I-Method
positive NN O I-Method
sample NN O I-Method
mining NN O I-Method
method NN O I-Method
to NN O O
train NN O O
robust NN O I-Method
CNN NN O I-Method
for NN O O
person NN O I-Task
re-identification NN O I-Task
, NN O O
dealing NN O O
with NN O O
the NN O O
problem NN O O
of NN O O
large NN O O
variation NN O O
. NN O O

In NN O O
addition NN O O
, NN O O
we NN O O
improve NN O O
the NN O O
learning NN O I-Generic
by NN O O
a NN O O
metric NN O I-OtherScientificTerm
weight NN O I-OtherScientificTerm
constraint NN O I-OtherScientificTerm
, NN O O
so NN O O
that NN O O
the NN O O
learned NN O I-OtherScientificTerm
metric NN O I-OtherScientificTerm
has NN O O
a NN O O
better NN O O
generalization NN O I-OtherScientificTerm
ability NN O I-OtherScientificTerm
. NN O O

Experiments NN O O
show NN O O
that NN O O
these NN O O
two NN O O
strategies NN O O
are NN O O
effective NN O O
in NN O O
learning NN O O
robust NN O I-OtherScientificTerm
deep NN O I-OtherScientificTerm
metrics NN O I-OtherScientificTerm
for NN O O
person NN O I-Task
re-identification NN O I-Task
, NN O O
and NN O O
accordingly NN O O
our NN O O
deep NN O I-Method
model NN O I-Method
significantly NN O O
outperforms NN O O
the NN O O
state-of-the-art NN O I-Generic
methods NN O I-Generic
on NN O O
several NN O O
benchmarks NN O O
of NN O O
person NN O I-Task
re-identification NN O I-Task
. NN O O

Therefore NN O O
, NN O O
the NN O O
study NN O O
presented NN O O
in NN O O
this NN O O
paper NN O O
may NN O O
be NN O O
useful NN O O
in NN O O
inspiring NN O O
new NN O O
designs NN O O
of NN O O
deep NN O I-Method
models NN O I-Method
for NN O O
person NN O I-Task
re-identification NN O I-Task
. NN O O

-DOCSTART- NN O O

Utterance NN O I-Method
Verification NN O I-Method
-LRB- NN O I-Method
UV NN O I-Method
-RRB- NN O I-Method
is NN O O
a NN O O
critical NN O O
function NN O O
of NN O O
an NN O O
Automatic NN O I-Method
Speech NN O I-Method
Recognition NN O I-Method
-LRB- NN O I-Method
ASR NN O I-Method
-RRB- NN O I-Method
System NN O I-Method
working NN O O
on NN O O
real NN O O
applications NN O O
where NN O O
spontaneous NN O I-OtherScientificTerm
speech NN O I-OtherScientificTerm
, NN O O
out-of-vocabulary NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
OOV NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
words NN O I-OtherScientificTerm
and NN O O
acoustic NN O I-OtherScientificTerm
noises NN O I-OtherScientificTerm
are NN O O
present NN O O
. NN O O

In NN O O
this NN O O
paper NN O O
we NN O O
present NN O O
a NN O O
new NN O O
UV NN O I-Method
procedure NN O I-Method
with NN O O
two NN O O
major NN O O
features NN O O
: NN O O
a NN O O
-RRB- NN O O
Confidence NN O I-Method
tests NN O I-Method
are NN O O
applied NN O O
to NN O O
decoded NN O I-OtherScientificTerm
string NN O I-OtherScientificTerm
hypotheses NN O I-OtherScientificTerm
obtained NN O O
from NN O O
using NN O O
word NN O O
and NN O O
garbage NN O O
models NN O O
that NN O O
represent NN O O
OOV NN O I-OtherScientificTerm
words NN O I-OtherScientificTerm
and NN O O
noises NN O I-OtherScientificTerm
. NN O O

Thus NN O O
the NN O O
ASR NN O I-Method
system NN O I-Method
is NN O O
designed NN O O
to NN O O
deal NN O O
with NN O O
what NN O O
we NN O O
refer NN O O
to NN O O
as NN O O
Word NN O I-Task
Spotting NN O I-Task
and NN O O
Noise NN O I-Task
Spotting NN O I-Task
capabilities NN O I-Task
. NN O O

b NN O O
-RRB- NN O O
The NN O O
UV NN O I-Method
procedure NN O I-Method
is NN O O
based NN O O
on NN O O
three NN O O
different NN O O
confidence NN O I-Method
tests NN O I-Method
, NN O O
two NN O I-Generic
based NN O O
on NN O O
acoustic NN O I-Metric
measures NN O I-Metric
and NN O O
one NN O I-Generic
founded NN O O
on NN O O
linguistic NN O I-OtherScientificTerm
information NN O I-OtherScientificTerm
, NN O O
applied NN O O
in NN O O
a NN O O
hierarchical NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
. NN O O

Experimental NN O O
results NN O O
from NN O O
a NN O O
real NN O O
telephone NN O I-Task
application NN O I-Task
on NN O O
a NN O O
natural NN O I-Task
number NN O I-Task
recognition NN O I-Task
task NN O I-Task
show NN O O
an NN O O
50 NN O O
% NN O O
reduction NN O O
in NN O O
recognition NN O I-Metric
errors NN O I-Metric
with NN O O
a NN O O
moderate NN O O
12 NN O O
% NN O O
rejection NN O I-Metric
rate NN O I-Metric
of NN O O
correct NN O O
utterances NN O O
and NN O O
a NN O O
low NN O O
1.5 NN O O
% NN O O
rate NN O O
of NN O O
false NN O I-Metric
acceptance NN O I-Metric
. NN O O

-DOCSTART- NN O O

A NN O O
critical NN O O
step NN O O
in NN O O
encoding NN O I-Task
sound NN O I-Task
for NN O O
neuronal NN O I-Task
processing NN O I-Task
occurs NN O O
when NN O O
the NN O O
analog NN O I-OtherScientificTerm
pressure NN O I-OtherScientificTerm
wave NN O I-OtherScientificTerm
is NN O O
coded NN O O
into NN O O
discrete NN O I-OtherScientificTerm
nerve-action NN O I-OtherScientificTerm
potentials NN O I-OtherScientificTerm
. NN O O

Recent NN O O
pool NN O I-Method
models NN O I-Method
of NN O O
the NN O O
inner NN O I-OtherScientificTerm
hair NN O I-OtherScientificTerm
cell NN O I-OtherScientificTerm
synapse NN O I-OtherScientificTerm
do NN O O
not NN O O
reproduce NN O O
the NN O O
dead NN O O
time NN O O
period NN O O
after NN O O
an NN O O
intense NN O O
stimulus NN O O
, NN O O
so NN O O
we NN O O
used NN O O
visual NN O I-Method
inspection NN O I-Method
and NN O O
automatic NN O I-Method
speech NN O I-Method
recognition NN O I-Method
-LRB- NN O I-Method
ASR NN O I-Method
-RRB- NN O I-Method
to NN O O
investigate NN O O
an NN O O
offset NN O I-Method
adaptation NN O I-Method
-LRB- NN O I-Method
OA NN O I-Method
-RRB- NN O I-Method
model NN O I-Method
proposed NN O O
by NN O O
Zhang NN O O
et NN O O
al. NN O O
-LSB- NN O O
1 NN O O
-RSB- NN O O
. NN O O

OA NN O I-Method
improved NN O O
phase NN O I-Task
locking NN O I-Task
in NN O I-Task
the NN O I-Task
auditory NN O I-Task
nerve NN O I-Task
-LRB- NN O I-Task
AN NN O I-Task
-RRB- NN O I-Task
and NN O O
raised NN O O
ASR NN O I-Metric
accuracy NN O I-Metric
for NN O O
features NN O I-OtherScientificTerm
derived NN O O
from NN O O
AN NN O I-OtherScientificTerm
fibers NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
ANFs NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
. NN O O

We NN O O
also NN O O
found NN O O
that NN O O
OA NN O I-Method
is NN O O
crucial NN O O
for NN O O
auditory NN O I-Task
processing NN O I-Task
by NN O O
onset NN O I-OtherScientificTerm
neurons NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
ONs NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
in NN O O
the NN O O
next NN O O
neuronal NN O O
stage NN O O
, NN O O
the NN O O
auditory NN O I-OtherScientificTerm
brainstem NN O I-OtherScientificTerm
. NN O O

Multi-layer NN O I-Method
perceptrons NN O I-Method
-LRB- NN O I-Method
MLPs NN O I-Method
-RRB- NN O I-Method
performed NN O O
much NN O O
better NN O O
than NN O O
standard NN O O
Gaussian NN O I-Method
mixture NN O I-Method
models NN O I-Method
-LRB- NN O I-Method
GMMs NN O I-Method
-RRB- NN O I-Method
for NN O O
both NN O O
our NN O O
ANF-based NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
ON-based NN O I-OtherScientificTerm
auditory NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
. NN O O

Similar NN O O
results NN O O
were NN O O
previously NN O O
obtained NN O O
with NN O O
MSG NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
Modulation-filtered NN O I-OtherScientificTerm
Spec-troGram NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
auditory NN O I-OtherScientificTerm
features NN O I-OtherScientificTerm
-LSB- NN O O
2 NN O O
-RSB- NN O O
. NN O O

Thus NN O O
we NN O O
believe NN O O
researchers NN O O
working NN O O
with NN O O
novel NN O O
features NN O I-OtherScientificTerm
should NN O O
consider NN O O
trying NN O O
MLPs NN O I-Method
. NN O O

-DOCSTART- NN O O

Recent NN O O
progress NN O O
in NN O O
computer NN O I-Task
vision NN O I-Task
has NN O O
been NN O O
driven NN O O
by NN O O
high-capacity NN O I-Method
models NN O I-Method
trained NN O O
on NN O O
large NN O I-Material
datasets NN O I-Material
. NN O O

Unfortunately NN O O
, NN O O
creating NN O O
large NN O I-Material
datasets NN O I-Material
with NN O O
pixel-level NN O I-OtherScientificTerm
labels NN O I-OtherScientificTerm
has NN O O
been NN O O
extremely NN O O
costly NN O O
due NN O O
to NN O O
the NN O O
amount NN O O
of NN O O
human NN O O
effort NN O O
required NN O O
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
present NN O O
an NN O O
approach NN O I-Generic
to NN O O
rapidly NN O O
creating NN O O
pixel-accurate NN O I-OtherScientificTerm
semantic NN O I-OtherScientificTerm
label NN O I-OtherScientificTerm
maps NN O I-OtherScientificTerm
for NN O O
images NN O I-Material
extracted NN O O
from NN O O
modern NN O I-OtherScientificTerm
computer NN O I-OtherScientificTerm
games NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

We NN O O
propose NN O O
a NN O O
novel NN O O
step NN O O
toward NN O O
the NN O O
unsupervised NN O I-Task
seg-mentation NN O I-Task
of NN O I-Task
whole NN O I-Task
objects NN O I-Task
by NN O O
combining NN O O
'' NN O O
hints NN O O
'' NN O O
of NN O O
partial NN O I-Task
scene NN O I-Task
segmentation NN O I-Task
offered NN O O
by NN O O
multiple NN O O
soft NN O I-OtherScientificTerm
, NN O I-OtherScientificTerm
binary NN O I-OtherScientificTerm
mattes NN O I-OtherScientificTerm
. NN O O

These NN O O
mattes NN O I-Generic
are NN O O
implied NN O O
by NN O O
a NN O O
set NN O O
of NN O O
hypothesized NN O I-OtherScientificTerm
object NN O I-OtherScientificTerm
boundary NN O I-OtherScientificTerm
fragments NN O I-OtherScientificTerm
in NN O O
the NN O O
scene NN O O
. NN O O

Rather NN O O
than NN O O
trying NN O O
to NN O O
find NN O O
or NN O O
define NN O O
a NN O O
single NN O O
'' NN O O
best NN O O
'' NN O O
segmentation NN O O
, NN O O
we NN O O
generate NN O O
multiple NN O O
segmentations NN O O
of NN O O
an NN O O
image NN O O
. NN O O

This NN O O
reflects NN O O
contemporary NN O I-Generic
methods NN O I-Generic
for NN O O
unsupervised NN O I-Task
object NN O I-Task
discovery NN O I-Task
from NN O O
groups NN O O
of NN O O
images NN O O
, NN O O
and NN O O
it NN O O
allows NN O O
us NN O O
to NN O O
define NN O O
intuitive NN O O
evaluation NN O O
met-rics NN O O
for NN O O
our NN O O
sets NN O O
of NN O O
segmentations NN O O
based NN O O
on NN O O
the NN O O
accurate NN O O
and NN O O
parsimonious NN O O
delineation NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
scene NN O I-OtherScientificTerm
objects NN O I-OtherScientificTerm
. NN O O

Our NN O O
proposed NN O O
approach NN O I-Generic
builds NN O O
on NN O O
recent NN O O
advances NN O O
in NN O O
spectral NN O I-Method
clustering NN O I-Method
, NN O O
image NN O I-Method
matting NN O I-Method
, NN O O
and NN O O
boundary NN O I-Method
detection NN O I-Method
. NN O O

It NN O I-Generic
is NN O O
demonstrated NN O O
qualitatively NN O O
and NN O O
quantitatively NN O O
on NN O O
a NN O O
dataset NN O I-Material
of NN O I-Material
scenes NN O I-Material
and NN O O
is NN O O
suitable NN O O
for NN O O
current NN O O
work NN O O
in NN O O
unsupervised NN O I-Task
object NN O I-Task
discovery NN O I-Task
without NN O O
top-down NN O O
knowledge NN O O
. NN O O

-DOCSTART- NN O O

Language NN O I-Metric
resource NN O I-Metric
quality NN O I-Metric
is NN O O
crucial NN O O
in NN O O
NLP NN O I-Task
. NN O O

Many NN O O
of NN O O
the NN O O
resources NN O O
used NN O O
are NN O O
derived NN O O
from NN O O
data NN O O
created NN O O
by NN O O
human NN O O
beings NN O O
out NN O O
of NN O O
an NN O O
NLP NN O I-Task
context NN O O
, NN O O
especially NN O O
regarding NN O O
MT NN O I-Task
and NN O O
reference NN O I-Task
translations NN O I-Task
. NN O O

Indeed NN O O
, NN O O
automatic NN O I-Task
evaluations NN O I-Task
need NN O O
high-quality NN O I-Material
data NN O I-Material
that NN O O
allow NN O O
the NN O O
comparison NN O O
of NN O O
both NN O O
automatic NN O O
and NN O O
human NN O O
translations NN O O
. NN O O

The NN O O
validation NN O O
of NN O O
these NN O O
resources NN O I-Generic
is NN O O
widely NN O O
recommended NN O O
before NN O O
being NN O O
used NN O O
. NN O O

This NN O O
paper NN O O
describes NN O O
the NN O O
impact NN O O
of NN O O
using NN O O
different-quality NN O I-OtherScientificTerm
references NN O I-OtherScientificTerm
on NN O O
evaluation NN O I-Generic
. NN O O

Surprisingly NN O O
enough NN O O
, NN O O
similar NN O O
scores NN O O
are NN O O
obtained NN O O
in NN O O
many NN O O
cases NN O O
regardless NN O O
of NN O O
the NN O O
quality NN O O
. NN O O

Thus NN O O
, NN O O
the NN O O
limitations NN O O
of NN O O
the NN O O
automatic NN O I-Metric
metrics NN O I-Metric
used NN O O
within NN O O
MT NN O I-Task
are NN O O
also NN O O
discussed NN O O
in NN O O
this NN O O
regard NN O O
. NN O O

-DOCSTART- NN O O

This NN O O
poster NN O O
paper NN O O
describes NN O O
a NN O O
full NN O I-Task
scale NN O I-Task
two-level NN O I-Task
morphological NN O I-Task
description NN O I-Task
-LRB- NN O O
Karttunen NN O O
, NN O O
1983 NN O O
; NN O O
Koskenniemi NN O O
, NN O O
1983 NN O O
-RRB- NN O O
of NN O O
Turkish NN O I-Material
word NN O I-Material
structures NN O I-Material
. NN O O

The NN O O
description NN O I-Generic
has NN O O
been NN O O
implemented NN O O
using NN O O
the NN O O
PC-KIMMO NN O I-Method
environment NN O I-Method
-LRB- NN O O
Antworth NN O O
, NN O O
1990 NN O O
-RRB- NN O O
and NN O O
is NN O O
based NN O O
on NN O O
a NN O O
root NN O I-Material
word NN O I-Material
lexicon NN O I-Material
of NN O O
about NN O O
23,000 NN O O
roots NN O O
words NN O O
. NN O O

Almost NN O O
all NN O O
the NN O O
special NN O O
cases NN O O
of NN O O
and NN O O
exceptions NN O O
to NN O O
phonological NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
morphological NN O I-OtherScientificTerm
rules NN O I-OtherScientificTerm
have NN O O
been NN O O
implemented NN O O
. NN O O

Turkish NN O I-Material
is NN O O
an NN O O
agglutinative NN O I-Material
language NN O I-Material
with NN O O
word NN O I-OtherScientificTerm
structures NN O I-OtherScientificTerm
formed NN O O
by NN O O
productive NN O I-OtherScientificTerm
affixations NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
derivational NN O I-OtherScientificTerm
and NN O I-OtherScientificTerm
inflectional NN O I-OtherScientificTerm
suffixes NN O I-OtherScientificTerm
to NN O O
root NN O O
words NN O O
. NN O O

Turkish NN O I-Material
has NN O O
finite-state NN O I-OtherScientificTerm
but NN O O
nevertheless NN O O
rather NN O O
complex NN O O
morphotactics NN O O
. NN O O

Morphemes NN O I-OtherScientificTerm
added NN O O
to NN O O
a NN O O
root NN O O
word NN O O
or NN O O
a NN O O
stem NN O O
can NN O O
convert NN O O
the NN O O
word NN O O
from NN O O
a NN O O
nominal NN O O
to NN O O
a NN O O
verbal NN O I-OtherScientificTerm
structure NN O I-OtherScientificTerm
or NN O O
vice-versa NN O O
, NN O O
or NN O O
can NN O O
create NN O O
adverbial NN O I-OtherScientificTerm
constructs NN O I-OtherScientificTerm
. NN O O

The NN O O
surface NN O I-Task
realizations NN O I-Task
of NN O I-Task
morphological NN O I-Task
constructions NN O I-Task
are NN O O
constrained NN O O
and NN O O
modified NN O O
by NN O O
a NN O O
number NN O O
of NN O O
phonetic NN O I-OtherScientificTerm
rules NN O I-OtherScientificTerm
such NN O O
as NN O O
vowel NN O I-OtherScientificTerm
harmony NN O I-OtherScientificTerm
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
deals NN O O
with NN O O
the NN O O
problem NN O O
of NN O O
generating NN O O
the NN O O
fundamental NN O I-OtherScientificTerm
frequency NN O I-OtherScientificTerm
-LRB- NN O I-OtherScientificTerm
F0 NN O I-OtherScientificTerm
-RRB- NN O I-OtherScientificTerm
contour NN O I-OtherScientificTerm
of NN O I-OtherScientificTerm
speech NN O I-OtherScientificTerm
from NN O O
a NN O O
text NN O I-Material
input NN O I-Material
for NN O O
text-to-speech NN O I-Task
synthesis NN O I-Task
. NN O O

We NN O O
have NN O O
previously NN O O
introduced NN O O
a NN O O
statistical NN O I-Method
model NN O I-Method
describing NN O O
the NN O O
generating NN O O
process NN O O
of NN O O
speech NN O I-OtherScientificTerm
F0 NN O I-OtherScientificTerm
contours NN O I-OtherScientificTerm
, NN O O
based NN O O
on NN O O
the NN O O
discrete-time NN O O
version NN O O
of NN O O
the NN O O
Fujisaki NN O I-Method
model NN O I-Method
. NN O O

One NN O O
remarkable NN O I-OtherScientificTerm
feature NN O I-OtherScientificTerm
of NN O O
this NN O O
model NN O I-Generic
is NN O O
that NN O O
it NN O I-Generic
has NN O O
allowed NN O O
us NN O O
to NN O O
derive NN O O
an NN O O
efficient NN O O
algorithm NN O I-Generic
based NN O O
on NN O O
powerful NN O O
statistical NN O I-Method
methods NN O I-Method
for NN O O
estimating NN O O
the NN O O
Fujisaki-model NN O I-OtherScientificTerm
parameters NN O I-OtherScientificTerm
from NN O O
raw NN O I-OtherScientificTerm
F0 NN O I-OtherScientificTerm
contours NN O I-OtherScientificTerm
. NN O O

To NN O O
associate NN O O
a NN O O
sequence NN O O
of NN O O
the NN O O
Fujisaki-model NN O I-OtherScientificTerm
parameters NN O I-OtherScientificTerm
with NN O O
a NN O O
text NN O I-Material
input NN O I-Material
based NN O O
on NN O O
statistical NN O I-Method
learning NN O I-Method
, NN O O
this NN O O
paper NN O O
proposes NN O O
extending NN O O
this NN O O
model NN O I-Generic
to NN O O
a NN O O
context-dependent NN O O
one NN O O
. NN O O

We NN O O
further NN O O
propose NN O O
a NN O O
parameter NN O I-Method
training NN O I-Method
algorithm NN O I-Method
for NN O O
the NN O O
present NN O O
model NN O I-Generic
based NN O O
on NN O O
a NN O O
decision NN O I-Method
tree-based NN O I-Method
context NN O I-Method
clustering NN O I-Method
. NN O O

-DOCSTART- NN O O

We NN O O
introduce NN O O
a NN O O
method NN O I-Generic
to NN O O
accelerate NN O O
the NN O O
evaluation NN O I-Task
of NN O I-Task
object NN O I-Task
detection NN O I-Task
cascades NN O I-Task
with NN O O
the NN O O
help NN O O
of NN O O
a NN O O
divide-and-conquer NN O I-Method
procedure NN O I-Method
in NN O O
the NN O O
space NN O I-Method
of NN O I-Method
candidate NN O I-Method
regions NN O I-Method
. NN O O

Compared NN O O
to NN O O
the NN O O
exhaustive NN O I-Method
procedure NN O I-Method
that NN O O
thus NN O O
far NN O O
is NN O O
the NN O O
state-of-the-art NN O O
for NN O O
cascade NN O I-Task
evaluation NN O I-Task
, NN O O
the NN O O
proposed NN O O
method NN O I-Generic
requires NN O O
fewer NN O O
evaluations NN O O
of NN O O
the NN O O
classifier NN O I-OtherScientificTerm
functions NN O I-OtherScientificTerm
, NN O O
thereby NN O O
speeding NN O O
up NN O O
the NN O O
search NN O I-Task
. NN O O

Furthermore NN O O
, NN O O
we NN O O
show NN O O
how NN O O
the NN O O
recently NN O O
developed NN O O
efficient NN O O
subwindow NN O I-Method
search NN O I-Method
-LRB- NN O I-Method
ESS NN O I-Method
-RRB- NN O I-Method
procedure NN O I-Method
-LSB- NN O O
11 NN O O
-RSB- NN O O
can NN O O
be NN O O
integrated NN O O
into NN O O
the NN O O
last NN O O
stage NN O O
of NN O O
our NN O O
method NN O I-Generic
. NN O O

This NN O O
allows NN O O
us NN O O
to NN O O
use NN O O
our NN O O
method NN O I-Generic
to NN O O
act NN O O
not NN O O
only NN O O
as NN O O
a NN O O
faster NN O O
procedure NN O O
for NN O O
cascade NN O I-Task
evaluation NN O I-Task
, NN O O
but NN O O
also NN O O
as NN O O
a NN O O
tool NN O O
to NN O O
perform NN O O
efficient NN O O
branch-and-bound NN O I-Task
object NN O I-Task
detection NN O I-Task
with NN O O
nonlinear NN O I-OtherScientificTerm
quality NN O I-OtherScientificTerm
functions NN O I-OtherScientificTerm
, NN O O
in NN O O
particular NN O O
kernel-ized NN O I-Method
support NN O I-Method
vector NN O I-Method
machines NN O I-Method
. NN O O

Experiments NN O O
on NN O O
the NN O O
PASCAL NN O I-Material
VOC NN O I-Material
2006 NN O I-Material
dataset NN O I-Material
show NN O O
an NN O O
acceleration NN O O
of NN O O
more NN O O
than NN O O
50 NN O O
% NN O O
by NN O O
our NN O O
method NN O I-Generic
compared NN O O
to NN O O
standard NN O O
cascade NN O I-Method
evaluation NN O I-Method
. NN O O

-DOCSTART- NN O O

Background NN O I-Task
modeling NN O I-Task
is NN O O
an NN O O
important NN O O
component NN O O
of NN O O
many NN O O
vision NN O I-Task
systems NN O I-Task
. NN O O

Existing NN O O
work NN O O
in NN O O
the NN O O
area NN O O
has NN O O
mostly NN O O
addressed NN O O
scenes NN O O
that NN O O
consist NN O O
of NN O O
static NN O I-OtherScientificTerm
or NN O I-OtherScientificTerm
quasi-static NN O I-OtherScientificTerm
structures NN O I-OtherScientificTerm
. NN O O

When NN O O
the NN O O
scene NN O I-Generic
exhibits NN O O
a NN O O
persistent NN O I-OtherScientificTerm
dynamic NN O I-OtherScientificTerm
behavior NN O I-OtherScientificTerm
in NN O O
time NN O O
, NN O O
such NN O O
an NN O O
assumption NN O O
is NN O O
violated NN O O
and NN O O
detection NN O I-Task
performance NN O O
deteriorates NN O O
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
propose NN O O
a NN O O
new NN O O
method NN O I-Generic
for NN O O
the NN O O
modeling NN O I-Task
and NN O I-Task
subtraction NN O I-Task
of NN O I-Task
such NN O I-Task
scenes NN O I-Generic
. NN O O

Towards NN O O
the NN O O
modeling NN O I-Task
of NN O I-Task
the NN O I-Task
dynamic NN O I-Task
characteristics NN O I-Task
, NN O O
optical NN O I-OtherScientificTerm
flow NN O I-OtherScientificTerm
is NN O O
computed NN O O
and NN O O
utilized NN O O
as NN O O
a NN O O
feature NN O I-OtherScientificTerm
in NN O O
a NN O O
higher NN O I-OtherScientificTerm
dimensional NN O I-OtherScientificTerm
space NN O I-OtherScientificTerm
. NN O O

Inherent NN O O
ambiguities NN O I-OtherScientificTerm
in NN O O
the NN O O
computation NN O I-Task
of NN O I-Task
features NN O I-Task
are NN O O
addressed NN O O
by NN O O
using NN O O
a NN O O
data-dependent NN O I-OtherScientificTerm
bandwidth NN O I-OtherScientificTerm
for NN O O
density NN O I-Task
estimation NN O I-Task
using NN O O
kernels NN O I-Method
. NN O O

Extensive NN O O
experiments NN O O
demonstrate NN O O
the NN O O
utility NN O O
and NN O O
performance NN O O
of NN O O
the NN O O
proposed NN O O
approach NN O I-Generic
. NN O O

-DOCSTART- NN O O

Information NN O I-Task
distillation NN O I-Task
aims NN O O
to NN O O
extract NN O O
relevant NN O O
pieces NN O O
of NN O O
information NN O O
related NN O O
to NN O O
a NN O O
given NN O O
query NN O O
from NN O O
massive NN O I-Material
, NN O I-Material
possibly NN O I-Material
multilingual NN O I-Material
, NN O I-Material
audio NN O I-Material
and NN O I-Material
textual NN O I-Material
document NN O I-Material
sources NN O I-Material
. NN O O

In NN O O
this NN O O
paper NN O O
, NN O O
we NN O O
present NN O O
our NN O O
approach NN O O
for NN O O
using NN O O
information NN O I-OtherScientificTerm
extraction NN O I-OtherScientificTerm
annotations NN O I-OtherScientificTerm
to NN O O
augment NN O O
document NN O I-Task
retrieval NN O I-Task
for NN O I-Task
distillation NN O I-Task
. NN O O

We NN O O
take NN O O
advantage NN O O
of NN O O
the NN O O
fact NN O O
that NN O O
some NN O O
of NN O O
the NN O O
distillation NN O I-OtherScientificTerm
queries NN O I-OtherScientificTerm
can NN O O
be NN O O
associated NN O O
with NN O O
annotation NN O I-OtherScientificTerm
elements NN O I-OtherScientificTerm
introduced NN O O
for NN O O
the NN O O
NIST NN O I-Task
Automatic NN O I-Task
Content NN O I-Task
Extraction NN O I-Task
-LRB- NN O I-Task
ACE NN O I-Task
-RRB- NN O I-Task
task NN O I-Task
. NN O O

We NN O O
experimentally NN O O
show NN O O
that NN O O
using NN O O
the NN O O
ACE NN O I-OtherScientificTerm
events NN O I-OtherScientificTerm
to NN O O
constrain NN O O
the NN O O
document NN O O
set NN O O
returned NN O O
by NN O O
an NN O O
information NN O I-Method
retrieval NN O I-Method
engine NN O I-Method
significantly NN O O
improves NN O O
the NN O O
precision NN O I-Metric
at NN O O
various NN O O
recall NN O I-Metric
rates NN O I-Metric
for NN O O
two NN O O
different NN O O
query NN O O
templates NN O O
. NN O O

-DOCSTART- NN O O

This NN O O
paper NN O O
presents NN O O
a NN O O
novel NN O O
representation NN O I-Generic
for NN O O
three-dimensional NN O I-OtherScientificTerm
objects NN O I-OtherScientificTerm
in NN O O
terms NN O O
of NN O O
affine-invariant NN O I-OtherScientificTerm
image NN O I-OtherScientificTerm
patches NN O I-OtherScientificTerm
and NN O O
their NN O O
spatial NN O I-OtherScientificTerm
relationships NN O I-OtherScientificTerm
. NN O O

Multi-view NN O I-OtherScientificTerm
constraints NN O I-OtherScientificTerm
associated NN O O
with NN O O
groups NN O O
of NN O O
patches NN O O
are NN O O
combined NN O O
with NN O O
a NN O O
normalized NN O I-Method
representation NN O I-Method
of NN O O
their NN O O
appearance NN O O
to NN O O
guide NN O O
matching NN O I-Task
and NN O O
reconstruction NN O I-Task
, NN O O
allowing NN O O
the NN O O
acquisition NN O I-Task
of NN O I-Task
true NN O I-Task
three-dimensional NN O I-Task
affine NN O I-Task
and NN O I-Task
Euclidean NN O I-Task
models NN O I-Task
from NN O O
multiple NN O O
images NN O I-Material
and NN O O
their NN O O
recognition NN O O
in NN O O
a NN O O
single NN O O
photograph NN O O
taken NN O O
from NN O O
an NN O O
arbitrary NN O O
viewpoint NN O O
. NN O O

The NN O O
proposed NN O O
approach NN O I-Generic
does NN O O
not NN O O
require NN O O
a NN O O
separate NN O O
segmentation NN O I-Method
stage NN O I-Method
and NN O O
is NN O O
applicable NN O O
to NN O O
cluttered NN O I-OtherScientificTerm
scenes NN O I-OtherScientificTerm
. NN O O

Preliminary NN O O
modeling NN O O
and NN O O
recognition NN O I-Task
results NN O O
are NN O O
presented NN O O
. NN O O

-DOCSTART- NN O O

Fast NN O I-Generic
algorithms NN O I-Generic
for NN O O
nearest NN O I-Task
neighbor NN O I-Task
-LRB- NN O I-Task
NN NN O I-Task
-RRB- NN O I-Task
search NN O I-Task
have NN O O
in NN O O
large NN O O
part NN O O
focused NN O O
on NN O O
2 NN O O
distance NN O I-OtherScientificTerm
. NN O O

Here NN O O
we NN O O
develop NN O O
an NN O O
approach NN O I-Generic
for NN O O
1 NN O I-OtherScientificTerm
distance NN O I-OtherScientificTerm
that NN O O
begins NN O O
with NN O O
an NN O O
explicit NN O O
and NN O O
exactly NN O O
distance-preserving NN O I-OtherScientificTerm
embedding NN O I-OtherScientificTerm
of NN O O
the NN O O
points NN O O
into NN O O
2 NN O O
2 NN O O
. NN O O

We NN O O
show NN O O
how NN O O
this NN O I-Generic
can NN O O
efficiently NN O O
be NN O O
combined NN O O
with NN O O
random-projection NN O I-Method
based NN O I-Method
methods NN O I-Method
for NN O O
2 NN O O
NN NN O I-Task
search NN O I-Task
, NN O O
such NN O O
as NN O O
locality-sensitive NN O I-Method
hashing NN O I-Method
-LRB- NN O I-Method
LSH NN O I-Method
-RRB- NN O I-Method
or NN O O
random NN O I-Method
projection NN O I-Method
trees NN O I-Method
. NN O O

We NN O O
rigorously NN O O
establish NN O O
the NN O O
correctness NN O O
of NN O O
the NN O O
methodology NN O O
and NN O O
show NN O O
by NN O O
experimentation NN O O
using NN O O
LSH NN O I-Method
that NN O O
it NN O I-Generic
is NN O O
competitive NN O O
in NN O O
practice NN O O
with NN O O
available NN O O
alternatives NN O I-Generic
. NN O O

