Báo cáo khoa học: "Growing Finely-Discriminating Taxonomies from Seeds of Varying Quality and Size" docx
Bạn đang xem bản rút gọn của tài liệu. Xem và tải ngay bản đầy đủ của tài liệu tại đây (251 KB, 8 trang )
Proceedings of the 12th Conference of the European Chapter of the ACL, pages 835–842,
Athens, Greece, 30 March – 3 April 2009.
c
2009 Association for Computational Linguistics
Growing Finely-Discriminating Taxonomies from Seeds
of Varying Quality and Size
Tony Veale
Guofu Li
Yanfen Hao
Abstract
! "
#
$"%%
!
1 Introduction
&
'
!&
(')*++,-
./(%01,,*-
"%(.et al.*++,-!2
$
! &
3(!!4
51,,6-!
)
3! %'0
78 # 3
(!! 59 (*+:6-
52;- #
!
#
!0
"%
2
!0
(5 1,,<-! 4 "%
#!"
=
!
grown
seeds!
&
7>?%/%8
!""%
# 2
835
&7
8
sharp!;
(-
@"%
X-ness! /
!
0
1
A
!<
=
#
!>
B
#
!&
6!
2 Related Work
#
(>
2*+::-!;5
(*++1-
#
4
(*+++- !
3(
-
! & KnowItAll 2 et al.
(1,,<-
5(!!7%0%0*%01C8-
#
! " (1,,D-
5
#
=
7 E
%/%87%/%E8
(-(-
(-!
?%(1,,<-
##
!&
"%
=
!>
! F et al.
(1,,B-
! >
=
!F et al. (1,,:-
5 (*++1-
!;
7%/%
%/%
E8
E
%/%
%/%
!
!&
$
reckless
(E-(%/%
-
=
!
&
F et al!(1,,:-!"
#
7>?
%/%
E87>?
E
%/%
8!>
(>?
- # !
&
"%"!Fet al!(1,,:-
3 states countries (
836
-singersfish(
-
food
sweet (G51,,D-!4
"%
"%"%
!
3 Seeds for Taxonomic Growth
>
&
3
HI
3
0
J
0
3
=
3
0
! &
Icola, carbonated, drinkJ!;
cola
(treatrefreshment-#
7E8
#7
E8!
#
#! "
$"
%%!
3.1 WordNet
& "%
!;"%
{feline, felid}
{true_cat, cat} {big_cat, cat}
!
5
"%
6,K Xess
"%(ess,
ess, ?ess!-
female !
%
"%
! >
Ilioness, female,
lionJ Iespresso, strong, coffeeJ
Imessiah, awaited, kingJImessiah, expect-
ed, delivererJ!
3.2 ConceptNet
"%
!
%('1,,<-
"""
!
%
(-
!'
>
% espresso
strong coffee("%-
bagelJewish word(usemen-
tion-!'expressionism
artistic style ("%
artistic movement- explosion
suicide attack(-!
%
"%
! " %
A,,,,>
(78-
(!!7 8-
(!!78-
"%!
& IWyoming, great,
stateJ Iwreck, serious, accidentJ Iwolf,
wild, animalJ!
3.3 Web-derived Stereotypes
G5(1,,D-
#
7>?%/%8
!&
#!!
#
!5
*BL
837
(!!787
8!-
6,,,
1,,, 3 ! 5
G59
!
&
=
Isurgeon, skilful, ?JIvirus, malicious, ?J
Idog, loyal, ?J!&
#
!
3.4 Overview of Seed Resources
%
!&"%
#
!&
%
"%
!>
G5
# 3
3
!>#
&*!
"% %
M
*111D **AA 6B*1
M
B*A*< *:,: *66::
M
<!*1 *!6 1!B6
M
1A,B BB, **D1
&*$&
!
""%
$
(-(
-
(
-! 4
#
!B
#
!
4 Bootstrapping from Seeds
&
!
NN
!&
&
3
HI
3
0
J
#
# (E
-$
*! 7
3
E
8
1! 7
3
0
E8
#
3
0
!
#
#$
A! 7E
3
0
E8
<! 7E
3
E
8
&
! ;
7
8 #
Ilemonade, cold, beverageJIlemon-
ade, refreshing, beverageJ!&
(
3
-
!
"&
#
#O ex-
pand(T')!"
) >0! /
# 1,,
#B,
#!
838
"
#
StK
t
S
. &
K
,
S
=S
K
*
S
=K
,
S
∪
{T
∣
T ' ∈S ∧ T ∈expand T ' }
K
t*
S
=K
t
S
∪
{T
∣
T ' ∈K
t
S
∧ T ∈expand T ' }
"#
3
!
# ex-
pand(T') !;
Fet al.(1,,:-
reckless bootstrapping
#
!&
3
!"
"% near-miss$
I
3
0
J"%
0
(-
0
( -!&
#
"%
#
"%(
"%-!&
$
K
tP
S
=K
t
S
∪
{
T
∣
T '∈K
t
S
∧
T ∈ filter
near−miss
expand T '
}
;*1
!
Q*
%
ND
<,
"%N
! & "% near-
miss
#!
;*$)#
B!
;1$)
#B
!
4.1 An Example
cola
$Icola, refreshing, beverageJ!>
cola
effervescent beverage
sweet beverage nonalcoholic beverage
!>
sugary foodfizzy drinkdark mixer!
> sensitive
beverage everyday beverage common
drink!>
irritating food unhealthy drink!>
stimulating
drinktoxic foodcorrosive substance!
cola
*<*<A1
D1A+A<*,1
B!
refreshing beverage
champagne
lemonadebeer!
0 1 2 3 4 5
0
200000
400000
600000
800000
1000000
1200000
1400000
1600000
1800000
WordNet
Simile
ConceptNet
Bootstrapping Cycle
# Triples
0 1 2 3 4 5
0
50000
100000
150000
200000
250000
300000
350000
WordNet
Simile
ConceptNet
Bootstrapping Cycle
# Terms
839
5 Empirical Evaluation
&"% near-miss
(0
-
(
-
(
3
- #
!&
#
#
3
#
3
!;
>
0 (1,,B-
#
O
"
%!
>0
<,1 1*
"%!&#
( hot
red!-#R(a|an|
the) * C
i
(is|was)R
3
!
not #
#
(0
-(
Temperature hot
-!&#+<+:+
<,1!&
'&/
F(1,,1-!4
<,1
1* >
0
1* "%
!&
"%
(-
!
3B6!DL
"%!4
61!DL!.0>
,!61D
,!AA:B*A<B
<,1!
*
We replicate the above experiments using the
same 402 nouns, and assess the clustering accur-
acy (again using WordNet as a gold-standard)
after each bootstrapping cycle. Recall that we use
only the D
j
fields of each triple as features for the
clustering process, so the comparison with the
WordNet gold-standard is still a fair one. Once
again, the goal is to determine how much like the
human-crafted WordNet taxonomy is the tax-
onomy that is clustered automatically from the
discriminating words D
j
only. The clustering ac-
curacy for all three seeds are shown in Tables 2,
3 and 4.
Cycle E P # Features Coverage
1
st
!A1D !61+ +,D 66L
2
nd
!1BA !D*1 *<:1 DDL
3
rd
!1D1 !D*D 1**< :1L
4
th
!A*1 !6<, 1<DA :AL
5
th
!1:+ !6:< 1DB1 :AL
&1$WordNet
(2200-
Cycle E P # Features Coverage
1
st
!**B !:<1 A6A <*L
2
nd
!1BB !D1< D:D B+L
3
rd
!1:6 !6+< *A61 D<L
4
th
!1D+ !6+< *:BA D+L
5
th
!1++ !6DA 11D< :1L
&A$ConceptNet
Cycle E P # Features Coverage
1
st
!1B< !D*6 :AD B+L
2
nd
!1:, !D*1 *AA: DAL
3
rd
!1:+ !6+A *+<< D+L
4
th
!A*A !66, 1A*1 :1L
5
th
!*BD !:<A 16*< :1L
&<$Simile
& <,1
#casuarina, cinchona, do-
decahedron concavity>
*
"
!"=
,!61D61!DL!
840
0
#B *,,
4%!'
0 >
#B
!
#(
&*-
(S:1L-B ! &
yesteryear, nonce ( -
salient(3-jag, droop,
fluting, fete, throb, poundage, stinging, rouble,
rupee, riel, drachma, escudo, dinar, dirham,
lira,dispensationhoardairstream(
-riversidecurling!;
A<
$
#
!
;A$)
!
;<$0
!&
0>$
H,!61D!
4 "% %
6:L6DL
B
61!DL
0 >! 5
:<!AL
66!<L0>
and ( Tem-
peratureColor!-
D,!+L
!;
<,1
G5(1,,:-
(6+!:BL-!
4
#
!&
316*<
0 > B*A<B
!&
0>!
6 Conclusions
&
#
B
!%#
! 4
"%
0>!"
!
&
#! ;
G5(1,,D-
O=
!G
5(1,,:-
1 2 3 4 5
0.40
0.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
WordNet
Simile
ConceptNet
Bootstrapping Cycle
Coverage
1 2 3 4 5
0.40
0.50
0.60
0.70
0.80
0.90
1.00
WordNet
Simile
ConceptNet
Poesio & Alm.
Bootstrapping Cycle
Purity
841
!
7D
j
C
i
8
7D
j
P
k
C
i
8=
C
i
D
j
! 7 8
!
>
:1L
0> !
B
A1:
<,1
:*!B+L !
!>
#
! &
#F et
al. (1,,:-
!
#
#
#
!
References
>&!2.!(*+::-!0!&
0!
In Proc. of the 26
th
>.>'
1*D 11<!
> >! 0 .! (1,,B-!
' "!
Proc. of the annual meeting of the Cognitive
Society?!
4 >! 5 )! (1,,6-! 2
"% .'Q
!Computational Linguistics,A1(*-$*A <D!
0!"?!(1,,D-!>>
# Q T
"! In Proc. of the 45
th
Annual Meeting of the
ACL::: :+B!
2!4.!(*+++-!;
! In Proc. of the 37
th
Annual
Meeting of the ACLBDN6<!
2 /! F ! ! .!
0> .!"!!&!
U>! (1,,<-!"
F> ( -! In
Proc. of the 13
th
WWW Conference*,,N*,+!
5F!?!(*+:6-!52;$>.
0! In Proc. of the 5
th
National Con-
ference on Artificial Intelligence 16D 1D*
00!>>
>!
50!(1,,<-!"%$"V
Proc. of GWC’2004, the 2
nd
Global WordNet con-
ference.4!
5 .! (*++1-! > #
! In Proc. of the
14
th
Int. Conf. on Computational Linguistics
BA+NB<B!
F G! Q ! &! >!
(1,,B-!&.$
! Int. Jour-
nal of Web and Grid Services*(1-1<, 166!
F )! (1,,1-! '&/$ > !
Technical Report 02-017.!
$OO !!!OSOO!
FW!Q2!52!(1,,:-!
' "5
0')!In Proc. of the 46
th
Annu-
al Meeting of the ACL.
'!4!)Q!G!(*++,-!4
$
3!%U$> "!
'5!0!(1,,<-%$>0
Q&! BT Technology
Journal11(<-$1** 116!
.)!4Q!;!)!
.F!?! (*++,-! "%$
!!?'
A(<-$1ABN1<<!
%!0>!(1,,*-!&
!In Proc. of the 2
nd
International Con-
ference on Formal Ontology in Information Sys-
tems (FOIS-2001)!
Q!?!%>!U!(1,,<-!'
! Advances in Neural Information Process-
ing Systems*D!
G&!5U! (1,,D-!.'/
; !In Proc.
of the 45
th
Annual Meeting of the ACLBDN6<!
G&!5U! (1,,:-!>;F
Q )
.!In Proc. of Coling 2008, The
22
nd
International Conference on Computational
Linguistics.!
842
