Extractio   of o ,,   eman :m iMormation from 
Ordinary Lnghsh Dictionary and its Evalua,tlon 
Jnn-ichi NAKAMURA, Makoto NAGAO 
Department of Electrical Engineering, 
Kyoto University, 
gosh ida-honma.chi, Sakyo, Kyoto, 606, jAPAN 
The ;~ ~tomatlc e~tractimt o~ scntar~tie ilthlrmatio~, ca-. 
pec{a.\]ly ~emrmtic rel~tionships be~wemt words, from sat o~:- 
di~l~'y I~h~glish dictionaty it~ described. For the extra,trim b 
the mag,~etic tape re:eaton or' I, DOCE (l,oagman f)ictic,- 
~ry of (Jmttempotary E~glish, 1978 editimt) is loaded 
b~to a ~:elatk,aa\] database system. Developed exgractimt 
pro!.,,~:;u~ts a.mdyze a definition se:uteltce in I,I)OCE with a 
pa,~te:r~t ~m~tching based algo*ithm. Si,tce this ~lgofithm is 
no t pe*\[e(:t, the *esalt of ¢\]te extra.el.trot h a~; been corn IJa.~:ed 
with sem~/ltic b, formatimt (sema.ntic markers) which gba 
ma.gnetic tape version o~ LI)OCI~ eontaiim. The zesnlt oi 
comparismt i~'~ a,b;o discussed for evaluating the !:e\]iability 
of aucl~ ;,.tt alttontatic e×traetion. 
A large <lictionary database is tm important conq>onent of a uat.. 
nral langl~age processing systmn. We already kuov+ sy'~dac~ic ill.- 
tbrm:~tion which should be and can be stored in a large dictionary 
d ~tabase for ~ practical application such as a machine hanslatiou 
system, tlowever, we still need 'more research on :;emanlic intbr- 
m~tion which can be prepared for a large system. As a first step 
to construct ~ large scale semantic dictionary (lexical knowledge 
base) the authors of this paper have inspected a machine read- 
able ordinary English dictionary LDOCE, Longrnan l)ictionary 
of Contempo,:ary English, 1.978 edition \[Procter 1987\]. 
Extr~m{ing semantic information ti'om an ordinary dictionary 
is ~n interesting research topic. One of the o~ims of automatic ex-- 
traction is to produce a thesaurus. No'~l, for example, proposed 
the idea of thesaurus production fl'om LDOCE in \[No~l 1982\]. 
Amsler also showed the result of automatic thesaurus production 
from a techniaeal encyclopedia \[Amsler 1987\]. Boguraev and AI-- 
shawi have st ~ldied the. utilization of LDOCE for natural language 
proce~,:sfing re'.;earches in geuerM \[Alshwai 1987,Bognraev 1987\]. 
in this paper, the automatic extraction of scntantic reh~tion- 
ships between words l¥om I,DOCE is described. For the ex-- 
traction, the m~gnctic tape version of 1,1)OCE is loaded into 
relational d~.tabase system. Developc~ extraction programs ana- 
lyze the definition sentence in LDOCE with a pattern matddng 
ba~d algorithm. Since this algorithm is not perfect, the result 
of the extrac;;ion ha'~ been compared with semantic information 
(sem~mtic markers) which the magnetic t~pe version of bDOCF, 
co~rtains. The result of comparison is also discussed ibr evaluat- 
ing the reliaLility of such ~n automatic extraction. 
,~ ~,DB Version of LDOCE 
hi genera.l, a dictionary consists of a complex dat~ structure: 
various relationships between words; grammatical information; 
usage l~otes, etc. '\['here.fore, we need a special database ma.nage- 
mer*t sy.~tem to handle dictionary data. l, br inst~nme, \[Nag~m 1980\] 
shows mlch a system for retrieving a Japanese dictionary. In this 
pa.per, however, the anthors are mainly interested in tile defini- 
tion and tile sample sentence parts of I,I)()(?E, ira;read of com- 
plex relati, ms among inlbrmation in the dictionary. 
For the sake of efticiency (including the cost of sy:;i.cm dc 
vetopment) of \[,DOCE retrieval, we have decided to u,';e a cow 
ventional relational databa.~;e management system (I{I)BM). 'iFhe 
Ii.DBM which we use is running on the rnaiafr~mm eomputel of 
Kyoto University Data Proce~ing Center (Fujitsu M782, (),q/IV 
l"4 MSP, FACOM AIM/RDB). 
For loading the magnetic version of LDOCE into this t\[I)li/i\[VI, 
we have extracted the following fields from I,I)OCI,;: 
1. IIead Word (IIW); 2. Part-.of-Speech (PS); 3. Deft- 
nition Number (DN); 4. Grammar (?ode (GC); 5. Box 
Code (BC); 6. Definition (Die); and 7. ~~trnt)le Se.w 
tence (SP). 
The Box Code field contains various information such as semantic 
restrictions, etc, which are explained in section 4.1. 
The fields I through 5 are ahnost tile same as the origi-. 
hal LDOCE data. (Several special characters are removed or 
changed into standard characters for simplicity of retrieval. The 
syllable division mark (.) is removed. Some of the font control 
characters are changed into '<' and '>.') 
'l'he definitions and the sample sentences are separated into 
a clause or a sentence. For example, definition 1 of the verb to 
abandon is: 
to leave completely ~md tbr ever; desert 
in the originM data. This definition is transfl)rmed into Lwo set)- 
state ebmses in the RI)B version: 
1. to leave completely and for ever 
2. desert. 
Since every data in the RDB is repres(mted in a tabular 
form, we have made three t~bles for the RDB version of I,\])OCF, 
(I,DOCE/RDB, see table I regarding their its record format): 
4'59 
t. Grammar Code and Box Code 'Fable (LDB.D1). 
2. Definition Table (LDB.D2, see table 2). 
3. Sample Sentence Table (LDB.D3). 
3 Extra(:tion of Semantic Information 
One form of semantic information useful for natural language 
processing is a thesaurus (or semantic network), which basically 
describes semantic relations between words. "1'o automatically 
produce the thesaurus from LDOCE, two programs have been 
dcveloped: 
1. Key Verb extraction progra m. 
'2. Key Noun and Function Noun extraction program. 
These programs and the result of extraction are discussed in this 
section. 
3.i Key Verb Extraction Program 
Most of the definitions of verbs in LDOCE are described as: 
to VERB ... 
Usually VERB in tlfis pattern expresses a 'key concept' of the 
defined verb. Therefore, we c,.U this VERB a Key Verb. 
For example, the verbs semantically related to the verb t0 hit 
have the tollowing definitions: 
e strike: to hit 
Table 2: Definition 
HW | PS DN / 
abandon v 1 
abandon v 1 
abandon v 2 
abandon v 3 
abandon v 4! 
abandon n 0 
abandon n 0 
abandoned adj 0 
Table {LDB.D2) of LDOCE/RDB 
DF 
to leave completely and fQr ever 
desert 
to leave (a relation o~ I~iend) in a though~ 
less or cruel way 
to give up, esp. without finishing 
to give (oneself) up completely to a feelo 
lag, desire, etc. 
the state when one's feelings and acgions 
axe uncontxoned 
freedom from control 
given up to a life that is though~ ~(~ be 
immoral see also ABANDON (2,4) 
* beat: to hit many times, esp. with a stick 
® kick: to hit with the foot 
® knee: to hi~ with the knee 
From this pattern of definitions, we can draw figm'e 1 which 
shows the semantic hierarchy around to kit: to beat, ~o kick and 
lo knee are specialized verbs of to kit 
~Ib expand this hierarchy, a program to extract the key verbs 
from a definition is developed. Table 3 (LDBV.D2) shows some 
examples of this extraction. In table 4, the frequency of key verbs 
is listed. Most frequently used key verb is l0 make. Note that 
~o make and to cause are used to define causative and transitive 
verbs respectively. 
Table 1: Record Format and Size of LDOCE/RDB 
l)\]: Grammax Code and Box Code Table (74,130 records) 
BC 
Name \[ llead Partof IDefinitionlGrammar Box / 
\]_Word Speech Number Code Code\[ 
I-A~r~h~--~2oy ~r~s) ~i,o4) I 
L Index J_IIlIW IIPS I I1DN.. I IIGC I1BC| 
D2: Definition Table (84,094 records) 
~At t-~(~o~-- I~ PS DN DF Name \[ tlead art of Definition DeFinition 
L. Word )eech Number 
fibute I char(20 at(10) char(10) ' "vatchar(250) 
\[___I),dex J_ I2HW \[2PS 12DN --. 
D3: Example Table (46,122 records) 
ame I ~ttead \[ Part of \[Definition I SamPle 
/ Word j Speech \[ Nnmbe~ Al-  a g 1 cha ( o) I char(lO) 
I Index t I3~W 1 13eS___.___~_ J___RP s I - 
hit ----~ strike 
/1',. 
many times/f \[ ~ith the knee 
/ w 
/ 
beat kick knee 
Figure 1: Semantic Hierarchy axround 'hit' 
Table 3: Definition and Key Verb Table (LDBV.D2, part) 
HW KV PS_ DN DF 
abase make v 0 to make (someone, esp. oneself) 
have less self-respect 
abase make V 0 make humble 
abash cause v 0 to cause to feel uncomioxtable o~ 
ashamed in the presence of others 
abate become v 1 (of winds, storms, disease, pain, 
etc.) to ~eeome less strong 
abate decrease v 1 decrease 
abate mane v 2 <lit> to make less 
abate bring v 3 <law> to bring to an end (esp. i~ 
the pht. <abate a nuisance> ) 
460 
Table 4: l,~ret 
KV 
make 
be 
~ive 
put 
take 
~tove ' 
have 
bc-co~ite 
go 
set 
uency of Key Verbs 
COUNT(KV) 
1311 
875 
641 
505 
446 
388 
383 
374 
336 
263 
208 
'lYaveming these relations between delined verb and key verb, 
a thesaurus (network) of verbs has been obtained approximately. 
Most of the verbs in this tlu.~uurus make s tree-like structure 
shown in figure 1. Ilowever, several 'loops' are found. A 'loop' 
exprea~es a cyclic definition: ~o welcome is defined by t0 greet, 
and to greet is defined by lo welcome. In the network, six typical 
cyclic definitions are: 
do: do (the verb to do does not have a key verb.) 
cha~tge: dtange, move~ come, become 
~ go: go, leave 
get: get, receive 
stop:: stop, cease 
o let: let, allow, permit 
Note that there are many other cyclic definitions in the network. 
However, most of them have a link to another verb; at least one 
of the verb in a cyclic definitions is defined by another verb. 
Since no reader of LDOCE cml understand the meaning of 
these verbs only from the dictionary, these may be a kind of bug 
of the dictionary. However, these cyclically defined verbs seem to 
correspond to semanlic primitives, which are first introduced to 
AI works by \[Sehank 1975\]. Semantic primitives may be defined 
outside of linfuislic words. Details of the result of extraction are 
discussed in \[Nakamura 1986\]. 
3o2 Key Noun and Function Noun Extraction 
Program 
We cau apply a similar algorithm to definitions of nouns, al- 
though the pattern of definitions of nouns is mo~e complex than 
that of verbs, ln~cting definitions with LDOCE/ttDB, most 
of them a~e, classified into two forms: 
1. {determiner} {adjective}* \]Key Noun {adjective phrase}* 
2. {determiner} {adjective}* le-hnction Noun of Key Noun 
{adjective phrase}* 
The first one is a simple form and many of them express is-a 
relations between a defined:noun and a key noun. For example, 
abandon: the w~aSe when one's feelings sad actions 
axe u;acontroIled 
shows that 
abandon is-a slate. 
The second form expresses more complex ~mantic relations 
between nouns. 
abbey: the group of people living in such a building 
shows that 
abbey is-a-group-of people. 
A function noun, therefore, explicitly expresses the semantic re- 
lation between a head word and a key noun. 
With terras of a semantic network, defined nouns aml key 
nouns are nodes in a semantic network, and function nouns 
(when function noun is empty, its function noun is regared ~ 
kind) expre~ the name of a link between nodes. The following 
nouns (41 nouns, in total) are considered to be function nouns, 
which are mannally extracted. 
is-a: kind, type, ... 
o part-of: part, side, top, ... 
member~shlp: set, member, group, class, family, ... 
® action: act, way, action, ... 
state: state, condition, ... 
amount: amount, sum, measure, ... 
degree: degree, quality, ... 
• form: form, shape, ... 
A program to extract key nouns and function nouns from the 
definitions of nouns is developed, rl~ble 5 shows a part of the key 
noun and fmtction noun table in the LI)OCE/RI)B (LDBN.D2) 
generated by this program. 
As shown in table 6, the key noun of highest frequency is 
person (2174 times) and for function noun is type (1064 times) 
except null function noun (pattern 1). 
'lYaversing is-a relation, for example, a thesaurus has been 
obtained \[Nakamura 1987\]. Table 7 shows a part of the autmnat- 
ically obtained thesaurus, whose 'root' word is person: actor is 
a-kind-of person; comedian, ezlra, ham, and mime are a-kind-of 
actor; comedienne is a-kind-of comedian. 
4 Comparison between Result of Ex- 
traction and BOX Code 
The thesmlrus produced from LDOCE by the key noun and key 
verb extraction programs is all approximate one, and, obviously, 
contains several errors. The key noun of abbreviation 1, for ex- 
ample, is shorler in table 5, because the current program ignores 
ing-formed words. However, it should be making. (Even if we 
(:hanged the extraction algorithm, still we have a problem that 
making is not a simple noun, but a gerund. We need to define 
noun-verb semantic relations.) To evaluate the quality of the 
produced thesaurus, the noun part of the thesaurus has been 
compared with the semantic markers in LDOCE. 
461. 
Table 
(LDBN.D2, part) 
llW )N 
abandon 0 
al)~udon 0 
abbey 1 
M)bey 1 
~,bbey 2 
M)bey 3 
abbreviation t 
al)brcviatiou 2 
5: Definition, Key Noun 
*J?. ............ LC 
state 
freedom 
building 
convent 
people group 
house 
shorter act 
word form 
and Funclion Noun Table 
DF 
Table 6: l'5"equency of Key Nouns and 
KN COUNT(KN) lL FN 
to 1660 type 
668 II act 
{;55 II piece 
479 state 
294 part 
261 group 
255 any 
253 quality 
232 types 
226 set 
206 action 
205 kind 
something 
place 
alan 
materiaJ 
in 
people 
plan t 
substance 
money 
apparatus 
the state when one's feel- 
ings and actions are SIICOU-. 
h:olled 
freedom frora control 
(esp. formerly) a building in 
wMch Christian meu (monk 
<s> ) or women (nun <s> 
) live shut away from other 
people and work as a group 
for God 
monastery > or convent 
the group of people living in 
such a building 
a large church o~ house that 
was once such a building 
the act of making shorter 
a shortened \]orm of a word, 
often one used in writing 
Function Nouns 
COUNT(FN) 
36583 
1064 
838 
603 
557 
498 
327 
306 
247 
246 
208 
2OO 
182 
person 
... 
accoIlllt nut 
C PA 
ace 
acto~ 
comedian 
comedienne 
extra 
sundry 
ham 
mime 
Table 7: Example of Th~aurus (person) 
_D N_ DF 
0 a person whose job is to keep and 
examine the money accounts of busi- 
nesses 
0 certified public accountant 
2 infml a person of the highest class or 
skill in something 
2 a person who takes part in something 
that happens 
1 an actor who a tells jokes or does 
amusing things to make people laugh 
0 a female comedian (1) 
2 an actor in a cinema film who has a 
very small part in a crowd scene and 
is 
0 c:vt~a (4) 
3 an actor whose acting is unnatu~ 
ral, esp. with improbable movements 
and expr 
3 an actor who performs without using 
words 
462 
/\ 
abs~l~act Concrete 
Iuanimafe a.mma.te (Q) @ 
Solk! (~as Human Pla~tt Animal 
Figure 2: ltierarchy of Semantic Markers i~ LDGCE 
4.1 Semantic Markers in L1)OeJ~ih £~0~ (C~de 
The magnetic version of LDOCE has a~ spech;~l field retatcd ~o 
semantic markers, which is called as BOX code tields, :,A~,,h,q@t 
it does not appear in the printed version of LI_)(){7~\]. Some o~! 
the BOX code field (called BOX1, tbr hlstance) express ~z-;ma~,~t~c 
restrictions for a noun governed by a verb or an adjective, ~,,d ~, 
semantic classification of a nolm. For exampl% the sema¢4ic re 
striction for a subject of the verb ~0 travel is marked ~_~ '~b~m~o~'; 
the noun person is classified as 'H? Th~ shows the,J, ~,h~ verb g0 
lravel may govern the noun per,~on in its snbjec~ po.~i~,io~. 'Lhe 
LDOCE uses 34 markers for expressing ~h~ restrictio~ ('~:~i,le 3). 
These semantic markers have a hierarci~y as shown in fi,% 
ure 2. ~br example, 'Human' , 'Plant', and 'A~dmaF are sub. 
elassificatior, s of 'animate (Q)? 
In the following part of this s&tion, the comparison betwee~ 
semantic markers of LDOCE and the thesaurus constrn&ed ti:o~, 
~he definitions of nouns in LDOCE is discussed ikon~ ~;he view 
Table 8: Semantic Markers in Box Code of Nouns and their 
Frequency (Part) 
type of (:ode 
A Animal 
B Female Animal 
C Concrete 
D Male Animal 
E 'S' + 'L' 
F Female tluman 
G Gas 
It ltuma~ 
I Inanimate 
J Movable 
K Male ('D' + 'M') 
L Liquid 
M Male Human 
N Not Movable 
O 'A' + 'II' 
P Plant 
Q Animate 
R Female ('B' + 'P) 
S Solid 
T Abstract 
U Collective + '0' 
V 'P' + 'A' 
W 'T' + T 
X 'T' + 'H' 
Y 'T' + 'Q' 
Z UNMARKED 
o.. 
total 
boxl DN=0,1 
43560 24906 
957 836 
26 15 
359 181 
27 21 
257 187 
453 314 
lit 79 
3457 2426 
42 26 
5794 3927 
2 2 
631 464 
875 603 
2144 1436 
69 42 
758 593 
23 14 
4 3 
1291 867 
16577 9668 
789 398 
20 15 
~03 61 
t97 108 
4t 18 
415 ~99 
...... ~0,~, '~'.,. N,m~,:, it/fi:~rkcA ~; Q (~nimate) and V (plant .{. animal) 
II\[W BI KN DF 
n~l~. developed under the influence 
tff man 
leu.~i tlta)l the usual size 
~,ure of b~ceds 
pta~kg~ V ~fe ghe very sta~:Al ~o~m~l of plant and ~- 
ixaal life that live in watee 
~,~M,~' K at&,r.d a male Eey_~9,Lyr.anidna ! 
*~ta!e g, mai~al a fcntale pe~2L(~iy!~l 
Oa.~'ea~ I\[ mothe~ the I~L~.~I._rn_p~_~ of a peraou 
poi,;; ~ff ti6::J ; derard~y. E.-:l>c~:ially the nous related to '.Animate', 
~ ~° ~ 
Nouns rdated to the concept animate have a relatively rumple 
st,nctnre in the thesaurus, us auimat~ is often used ~s an example 
(:d ~¢ the~uaar~_s.like system. Example~ of the words marked as 
':~~fimi~te (Q)' a~,(l rela~ed ~mims, c~pecia,lly marked ms 'plant q-- 
v.*d'md (V)', ~.re ,<~how~~ in table 9. 
The pro&aced thes~.mus contains more than 60% of the words 
mw&cd a>s eimple concepts, such as 'plant' (table 10), %.nimal', 
a~(t 'h..man (persm,~ in definitions)'~ i~ correct positions. As 
shown in t.ble 10, for example, 645 words are traversed from 
'£$:~ble 10: N(nms Related to (Living) Thing aml Plant 
(~ins) thi.~ .... phu~ (P) 
A 2 
D 2 
P 370 62.4% 
Q t 
other 270 
tot~ 645 
i, hc,~ i*~ tim pmduoed thesaurus; 370 words (62.4%) of these 
wosds a~c i~arked au 'Pin,it2 
l~owever, the produced tlmsaurus does not capture disjuneIive 
coucelAs ~a(h ~s %hiram or plant (V) ~ correctly. In the definition 
of cro~b','eed (table 9), the produced thesaurus only uses plant 
a~ v. key nom~, and ingores a~lffmal. This is a typical problem hi 
~.he current produced th~aurus. 
No~e tln..t the disth~ction between 'animate (Q)' and 'animal 
o~. pl~.nt (V} ~ (animate without human) .~enm to tie difficult for 
the lexico~r;i:aphe~'s: bl~ed is marked as Q; cwssbreed, however, is 
4£-i N~'a~s }Y~arked ~,~ ~abs~lYacU 
~.~ LDOC~3 really nouns (about 40%, table 8) are marked as 
':.<bsS,~h'ozC, ~md fltey are not classified into more detailed sub- 
cl~.~:~:o 0~ ~he other hand, fimction nouns work as a key for 
~b, ch~:dtic~tk,ia i~i the produced thesaurus, ha r~ction a.2, some 
of the function nouns are listed as action, star% amount trod 
degree. The~e function nouns classify abstract nouns. 
For example, there are 597 nouns whose function noun is ilct, 
and 584 nouns (97%) of them are marked as 'abstract'; there are 
398 nouns whose function noun is state, and 391 nouns (98%) 
of them are 'abstract.' The distinction between <state' and 'act', 
h)r instance, is useful for natural language processing in general. 
4.4 Nouns Marked as 'Inanimate' 
Some 'Inanimate' nouns are correctly identified in the produced 
thesaurus (table 11). Especially, 39% of nouns under the noun 
liquid have 'Liquid' markers, and 56~ of nmms under the noun 
gas have 'Gas' markers. 
However, many <Inanimate' nouns are defined by substance 
in LDOCE. Sub-classification of these noun is expr(~sed with a 
compound word (or an adjective) as shown in table 11: coke is a 
solid substance; fluorine is a non-metallic substance. Since the 
currect extraction program does not handle a compound word, 
the thesaurus cannot express these classification. 
4.5 Other Typical Nouns 
Several typical nouns in the produced thesaurus are also com- 
pared with markers of LDOCE. Because the current system can.- 
not distinguish senses of nouns, nouns which have several differ- 
ent senses causes a problem. A typical example is found in the 
definitions whose key noun is case. As shown in table 12, altache 
ease and tesl ease are both defined by case; these expr~ses corn 
pletely different concept. In 30 nouns whose key noun is case, 
Table 11: Examples of 
RW 
hydrogen 
water 
coke 
fluorine 
B1 KN 
L liquid 
S substance 
G' substance 
Nouns Marked as 'Inanimate' 
DF 
a gas that is a simple substance 
(ELEMENT), without colour or 
smell, that is lighter than Mr and 
that burns very emsiiy 
the most common liquid,' without 
colour, taste, or smell, wtlk:h falls 
from the sky as rain, forms rivers, 
lakes, and seas, and is drunk by peo- 
ple and animals 
the solid substance that remains af- 
ter gas has been removed from coal 
by heating 
a non-metallic substance, na~l, in 
the form of a poisonous pale 
greenish-yellow gas 
463 
Table 12: Nouns whose key noun is case 
HW B1 I KN DF 
attache case J case ryinga thinpapershard case with a handle, for car- 
test case T case a case in acourt of law which establishes 
a particular principle and is then as a 
standard against which other eases can 
be judged 
HW 
CaUVaS 
denim 
serge 
tweed 
J 
S 
J 
S 
Table 13: Nouns related the noun cloth 
L FF DF 
strong rough cloth used for tent, sails, bags, 
etc. 
a strong cotton cloth used esp. for jeans 
type a type of strong cloth, usu. woven from wo01, 
and used esp. for suits, coats, and dresses 
type a type of coarse woolen cloth woven form 
threads of several different colours 
16 nouns are 'movable (J)', and 14 nouns are 'absTract.' 
Difficulity of semantic marking is also found. For example, 
lexicographers could not mark 'movable (J)' and 'Solid' system- 
atically. For example, some nouns whose key noun is cloth are 
marked as 'Solid', and others are marked as 'movable (J)' (ta- 
ble 13). This is a problem in gathering of semantic information 
itself. 
5 Conclusion 
The extraction of semantic relations between verbs and nouns 
from LDOCE is discussed. Data from the magnetic version of 
LDOCE is first loaded into a relational database system for sim- 
plicity of retrieving. For the extraction of semantic relations, 
programs to find key verb, key noun, and function noun have 
been developed. Using these programs, the thesaurus is auto- 
matically produced. 
• b evaluate the quality of the noun part of the produced the- 
saurus, it is compared with the semantic markers in LDOCE. Al- 
though the produced thesaurus has several problems such as the 
difficulty of expressing disjunctive concepts, the comparison be- 
tween the produced thesaurus and semantic markers in LDOCE 
shows the possibility of sub-classifiCation of 'abstract' nouns. 
Acknowledgements 
The authors grateful to Prof. Jun-ichi Tsujii for his fruitful com- 
ments on. this work. We also wish to thank Mr. Motohiro Fuji- 
gaki, Mr. Nobuhiro Kato, and Mr. Keiichi Sakai who inspected 
LDOCE data carefully. 

References 

\[Alshwai 1987\] ALSHAWI, H., Processing Dictionary Definitions 
with Phrasal Pattern Hierarchies, Computational gin- 
guisiic~, Vol. 13 (1987). 

\[Amsler 1987\] AMSLSlt, R. A., How Do I Turn This Book On?, 
P~c. o\] Third Annual Con\]. o\] the UW Centre for the 
NOEl), pp. 75-88 (1987). 

\[Boguraev 1987\] BOGURAEV, B., Experiences with a Machine- 
Readable Dictionary, P~vc. of Third Annual Con\]. o\] 
the UW Centre \]or the NOEl), pp. 37-50 (1987). 

\[Nagao 1980\] NAGAO, M., TsuJII, J., UEDA, Y., TAKIYAMA, 
M., An Attempt to Computerized Dictionary Data 
Bases, Proc. of COLING80, pp. 534-542 (1980). 

\[Nakamura 1986\] NAKAMURA, J., FUJIGAKI, M., NAGAO, M., 
Longman Dictionary Database and Extraction of its 
Information, Report on Cognitive Approaches for 
Discourse Modeling , Kyoto University (1986) (in 
Japanese). 

\[Nakamura 1.987\] NAKAMURA, J., SAKAI, K., NAGAO, M., Au- 
tomatic Analysis of Semantical Relation between En- 
glish Nouns by an Ordinary English Dictionary, Im 
stitute of Electronics, Information and Communica- 
tion Engineers of Japan, WGNLC, 86-23 (1987) (in 
Japanese). 

\[Noel 1982\] MICHI~LS, A., NOi~L, J., Approaches to Thesaurus 
Production, Proc. o\] COLING82, pp. 227-232 (1982). 

\[Procter I987\] PltOCTI~R, P., Longman Dictionary of Contem- 
porary English Longman Group Lhnited, Harlow and 
London, England (1978). 

\[Schank 1975\].ScHANK, R. C., Conceptual Information Pro- 
cessing, New York, North Holland (1975). 
