Construction of an Objective Hierarchy of Abstract Concepts           
via Directional Similarity  
Kyoko Kanzaki �  Eiko Yamamoto � Hitoshi Isahara 
Computational Linguistics Group, 
National Institute of Information and Communications 
Technology 
3-5 Hikari-dai, Seika-cho, Souraku-gun, Kyoto, Japan,  
{kanzaki, eiko, isahara}@nict.go.jp 
Qing Ma 
Faculty of Science  
and Technology 
Ryukoku University 
Seta, Otsu,520-2194, Japan 
qma@math.ryukoku.ac.jp
Abstract 
The method of organization of word mean-
ings is a crucial issue with lexical databases. 
Our purpose in this research is to extract word 
hierarchies from corpora automatically. Our 
initial task to this end is to determine adjec-
tive hyperonyms. In order to find adjective 
hyperonyms, we utilize abstract nouns. We 
constructed linguistic data by extracting se-
mantic relations between abstract nouns and 
adjectives from corpus data and classifying 
abstract nouns based on adjective similarity 
using a self-organizing semantic map, which 
is a neural network model (Kohonen 1995). 
In this paper we describe how to hierarchi-
cally organize abstract nouns (adjective hy-
peronyms) in a semantic map mainly using 
CSM. We compare three hierarchical organi-
zations of abstract nouns, according to CSM, 
frequency (Tf.CSM) and an alternative simi-
larity measure based on coefficient overlap, to 
estimate hyperonym relations between words. 
1. Introduction 
A lexical database is necessary for computers, 
and even humans, to fully understand a word's 
meaning because the lexicon is the origin of lan-
guage understanding and generation. Progress is 
being made in lexical database research, notably 
with hierarchical semantic lexical databases such 
as WordNet, which is used for NLP research 
worldwide.  
When compiling lexical databases, it is impor-
tant to consider what rules or phenomena should 
be described as lexical meanings and how these 
lexical meanings should be formalized and stored 
electronically. This is a common topic of discus-
sion in computational linguistics, especially in 
the domain of computational lexical semantics. 
The method of organization of word meanings 
is also a crucial issue with lexical databases. In 
current lexical databases and/or thesauri, abstract 
nouns indicating concepts are identified manually 
and words are classified in a top-down manner 
based on human intuition. This is a good way to 
make a lexical database for users with a specific 
purpose. However, word hierarchies based on 
human intuition tend to vary greatly depending 
on the lexicographer, and there is often dis-
agreement as to the make-up of the hierarchy. If 
we could find an objective method to organize 
word meanings based on real data, we would 
avoid this variability. 
Our purpose in this research is to extract word 
hierarchies from corpora automatically. Our ini-
tial task to this end is to determine adjective hy-
peronyms. In order to find adjective hyperonyms, 
we utilize abstract nouns. Past linguistic research 
has focused on classifying the semantic relation-
ship between abstract nouns and adjectives 
(Nemoto 1969, Takahashi 1975).  
We constructed linguistic data by extracting 
semantic relations between abstract nouns and 
adjectives from corpus data and classifying ab-
stract nouns based on adjective similarity using a 
self-organizing semantic map (SOM), which is a 
neural network model (Kohonen 1995). The rela-
tive proximity of words in the semantic map in-
dicates their relative similarity.  
In previous research, word meanings have 
been statistically modeled based on syntactic in-
formation derived from a corpus. Hindle (1990) 
used noun-verb syntactic relations, and Hatzivas-
siloglou and McKeown (1993) used coordinated 
adjective-adjective modifier pairs. These meth-
ods are useful for the organization of words deep 
within a hierarchy, but do not seem to provide a 
solution for the top levels of the hierarchy.  
To find an objective hierarchical word struc-
ture, we utilize the complementary similarity 
measure (CSM), which estimates a one-to-many 
relation, such as superordinate–subordinate rela-
tions (Hagita and Sawaki 1995, Yamamoto and 
Umemura 2002).  
In this paper we propose an automated method 
for constructing adjective hierarchies by connect-
ing strongly related abstract nouns in a top-down 
fashion � within a semantic map, mainly using 
CSM. We compare three hierarchical organiza-
tions of abstract nouns, according to CSM, fre-
quency (Tf.CSM) and an alternative similarity 
measure based on coefficient overlap, to estimate 
hyperonym relations between words. 
2. Linguistic clues to extract adjective hy-
peronyms from corpora 
In order to automatically extract adjective hy-
peronyms we use syntactic and semantic relations 
between words.  
There is a good deal of linguistic research fo-
cused on the syntactic and semantic functions of 
abstract nouns, including Nemoto (1969), Taka-
hashi (1975), and Schmid (2000). Takahashi 
(1975) illustrated the sentential function of ab-
stract nouns with the following examples. 
a.  Yagi  wa  seishitsu  ga  otonashii. 
(goat) topic (nature) subject (gentle) 
        The nature of goats is gentle 
b.   Zou    wa   hana   ga     nagai. 
    (elephant) topic  (a nose) subject  (long) 
         The nose of an elephant is long 
He examined the differences in semantic func-
tion between “seishitsu (nature)” in (a) and “hana 
(nose)” in (b), and explained that “seishitsu (na-
ture)” in (a) indicates an aspect of something, i.e., 
the goat, and “hana (nose)” in (b) indicates part 
of something, i.e., the elephant. He recognized 
abstract nouns in (a) as a hyperonym of the at-
tribute that the predicative adjectives express. 
Nemoto (1969) identified expressions such as 
“iro ga akai (the color is red)” and “hayasa ga 
hayai (the speed is fast)” as a kind of meaning 
repetition, or tautology.  
In this paper we define such abstract nouns 
that co-occur with adjectives as adjective hy-
peronyms. We semi-automatically extracted from 
corpora 365 abstract nouns used as this kind of 
head noun, according to the procedures described 
in Kanzaki et al. (2000). We collected abstract 
nouns from two year's worth of articles from the 
Mainichi Shinbun newspaper, and extracted ad-
jectives co-occurring with abstract nouns in the 
manner of (a) above from 100 novels, 100 essays 
and 42 year's worth of newspaper articles, includ-
ing 11 year's worth of Mainichi Shinbun articles, 
10 year's worth of Nihon Keizai Shinbun (Japa-
nese economic newspaper) articles, 7 year's wor-
th of Sangyoukinyuuryuutsu Shinbun (an eco-
nomic newspaper) articles, and 14 year's worth of 
Yomiuri Shinbun articles. The total number of 
abstract noun types is 365, the number of adjec-
tive types is 10,525, and the total number of ad-
jective tokens is 35,173. The maximum number 
of co-occurring adjectives for a given abstract 
noun is 1,594. 
3. On the Self-Organizing Semantic Map  
3.1  Input data 
Abstract nouns are located in the semantic map 
based on the similarity of co-occurring adjectives 
after iteratively learning over input data. 
In this research, we focus on abstract nouns 
co-occurring with adjectives. In the semantic 
map, there are 365 abstract nouns co-occurring 
with adjectives. The similarities between the 365 
abstract nouns are determined according to the 
number of common co-occurring adjectives. We 
made a list such as the following. 
OMOI (feeling): ureshii (glad), kanashii (sad), 
shiawasena (happy), … 
KIMOCHI (though): ureshii (glad), tanoshii (pleased), 
hokorashii (proud), … 
KANTEN (viewpoint): igakutekina (medical), 
rekishitekina (historical), ... 
When two (or more) sets of adjectives with 
completely different characteristics co-occur with 
an abstract noun and the meanings of the abstract 
noun can be distinguished correspondingly, we 
treat them as two different abstract nouns. For 
example, the Japanese abstract noun “men” is 
treated as two different abstract nouns with 
“men1” meaning “one side (of the characteristics 
of someone or something)” and “men2” meaning 
“surface”. The former co-occurs with “gentle”, 
“kind” and so on. The latter co-occurs with 
“rough”, “smooth” and so on. 
3.2  The Self-Organizing Semantic Map 
Ma (2000) classified co-occurring words using 
a self-organizing semantic map (SOM). 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   
We made a semantic map of the above-
mentioned 365 abstract nouns using SOM, based 
on the cosine measure. The distribution of the 
words in the map gives us a sense of the semantic 
distribution of the words. However, we could not 
precisely identify the relations between words in 
the map (Fig 1). In Fig. 1 lines on the maps indi-
cate close relations between word pairs. In the 
cosine-based semantic map, there is no clear cor-
respondence between word similarities and the 
distribution of abstract nouns in the map.    
To solve this problem we introduced the 
complementary similarity measure (CSM). This 
similarity measure estimates one-to-many 
relations, such as superordinate–subordinate 
relations (Hagita and Sawaki 1995, Yamamoto 
and Umemura 2002). We can find the 
hierarchical distribution of words in the semantic 
map according to the value of CSM (Fig 2). In 
the CSM-based SOM, lines are concentrated at 
the bottom right hand corner, that is, most ab-
stract nouns are located at the bottom right-hand 
corner.  
Next, we find hierarchical relations between 
whole abstract nouns, not between word pairs, on 
the map automatically. 
4. How to construct hierarchies of nominal 
adjective hyperonyms in the Semantic 
Map 
4.1 Similarity measures, CSM and Yates’ 
correction 
A feature of CSM is its ability to estimate hi-
erarchical relations between words. This similar-
ity measure was developed for the recognition of 
degraded machine-printed text (Hagita and Sa-
waki, 1995). Yates’ correction is often used in 
order to increase the accuracy of approximation. 
Hierarchical relations can be extracted accurately 
when the CSM value is high. Yates’ correction 
can extract different relations from high CSM 
values. When the CSM value is low, the result is 
not reliable, in which case we use Yates’ correc-
tion. 
According to Yamamoto and Umemura (2002), 
who adopted CSM to classify words, CSM is cal-
culated as follows. 
))(( dbca
bcad
CSM
++
−
=  
Yates’ correction is calculated as follows. 
))()()((
)2/|(|
2
dbcadcba
nbcadn
Yates
++++
−−
=  
Here n is the sum of the number of co-
occurring adjectives; a indicates the number of 
times the two labels appear together; b indicates 
the number of times “label 1” occurs but “label 
2” does not; c is the number of times “label 2” 
occurs but “label 1” does not; and d is the num-
ber of times neither label occurs. In our research, 
each “label” is an abstract noun, a indicates the 
number of adjectives co-occurring with both ab-
stract nouns, b and c indicate the number of ad-
jectives co-occurring with either abstract noun 
Figure 1. The Cosine-based SOM of word similarity Figure 2. The CSM-based SOM of word similarity
(“label 1” and “label 2”, respectively), and d in-
dicates the number of adjectives co-occurring 
with neither abstract noun. We calculated hierar-
chical relations between word pairs using these 
similarity measures. 
4.2 Construction of a hierarchy of abstract 
nouns using CSM and Yates' correc-
tion 
The hierarchy construction process is as fol-
lows: 
1) Based on the results of CSM, “koto (mat-
ter)” is the hyperonym of all abstract nouns. 
First, we connect super/sub-ordinate words 
with the highest CSM value while keeping the 
super-subordinate relation.  
2) When the normalized value of CSM is 
lower, the number of extracted word pairs be-
comes increasing overwhelmingly, and the reli-
ability of CSM diminishes. Word pairs with a 
normalized CSM value of less than 0.4 are lo-
cated far from the common hyperonym “koto 
(matter)” on the semantic map. If we construct a 
hierarchy using CSM values only, a long hierar-
chy containing irrelevant words emerges. In this 
case, the word pairs calculated by Yates' correc-
tion are more accurate than those from CSM. We 
combine words using Yates’ correction, when the 
value of CSM is less than 0.4. When we connect 
word pairs with a high Yates’ value, we find a 
hyperonym of the super-ordinate noun of the pair 
and connect the pair to the hyperonym. If a word 
pair appears only in the Yates' correction data, 
that is, we cannot connect the pair with high 
Yates’ value to the hyperonym with high CSM 
value, they are combined with “koto (matter)”. 
3) Finally, if a short hierarchy is contained in a 
longer hierarchy, it is merged with the longer 
hierarchy and we insert “koto (matter)” at the 
root of all hierarchies. 
4.3  Results 
The number of groups obtained was 161. At its 
deepest, the hierarchy was 15 words deep, and at 
its shallowest, it was 4 words deep. The 
following is a breakdown of the number of 
groups at different depths in the hierarchy.  
The greatest concentration of groups is at 
depth 7. There are 140 groups from depth 5 to 
depth 10, which is 87% of all groups. 
 
 
 
 
 
 
 
The word that has the strongest relation with 
“koto (matter)” is “men1 (side1)”. The number of 
groups in which “koto (matter)” and “men1 
(side1)” are hyperonyms is 96 (59.6%). The larg-
est number of groups after that is a group in 
which “koto (matter)”, “men1 (side1)” and 
“imeeji (image)” are hyperonyms. The number of 
groups in this case is 59 groups, or 36.6% of the 
total. With respect to the value of CSM, the co-
occurring adjectives are similar to “men1 (side1)” 
and “imeeji (image)”.  
Other words that have a direct relation with 
“koto (matter)” are “joutai (state)” and “toki 
(when)”. They have the most number of groups 
after “men1 (side1)” among all the children of 
“koto (matter)”. The number of groups subsumed 
by “joutai (state)” group and “toki (when)” are 21 
and 19, respectively. Other direct hyponyms of 
“koto (matter)” are: 
ki (feeling): 6 groups �  
ippou (while or grow –er and er): 3 groups �  
me2 (eyes): 3 groups �  
katachi1 (in the form of): 3 groups �  
iikata (how to say): 2 groups  
yarikata (how to): 2 groups 
There is little hierarchical structure to these 
groups, as they co-occur with few adjectives. 
4.4 The Hierarchies of abstract concepts in 
the semantic map 
In the following semantic maps, where abstract 
nouns are distributed using SOM and CSM (see 
Section 3), hierarchies of abstract nouns are 
drawn with lines. The bottom right hand corner is 
“koto (matter)”, a starting point for the distribu-
tion of abstract nouns.  
Five main types of hierarchies are found from 
patterns of lines on the map, as follows: 
The first figure, Fig.3, is hierarchies of “kanji 
(feeling), kimochi (feeling) …” on the semantic 
map. The location of hierarchies of “yousu (as-
pect), omomochi (look), kaotsuki (on one’s face), 
…” is similar to this type of the location. Hierar-
chies of “sokumen (one side), imi (meaning), 
kanten (viewpoint),  kenchi (standpoint) …” on 
Depth 4 5 6 7 8 9 
Groups 3 16 27 32 23 23 
Depth 10 11 12 13 14 15 
Groups 19 7 3 4 3 1 
Table 1: The depth of the hierarchy by CSM
 
 
 
 
 
 
 
the map are shown in Fig. 4. The lines of the hi-
erarchies go up from the bottom right hand cor-
ner to the upper left hand corner and then turn 
towards the upper right hand � corner. The loca-
tion of hierarchies of “nouryoku (ability), sainou 
(talent) …” is similar to this one. 
The hyperonym of “teido (degree)” is “joutai 
(state)”. In Fig.5 these abstract nouns are located 
at the bottom of the map. The location of hierar-
chies of “kurai (rather than)” and “hou (compara-
tively)” are similar to this one. The hierarchies of 
“joutai (state), joukyou (situation), yousou (as-
pect), jousei (the state of affairs)” are shown in 
Fig.6. The lines are found at a higher location 
than the line of “teido(degree)”. The lines of the 
hierarchies of “joutai (state), ori (when), sakari 
(in the hight of), sanaka (while)” are similar to 
these lines. 
The lines of the hierarchies of “seikaku (char-
acter)”, “gaikan (appearance)”and “utsukushisa 
(beauty)” are similar to each other. We show the 
hierarchies of “seikaku (character)” in Fig.7. The-
se lines in Fig.7 are located from the right end to 
the upper left hand corner. From the following, 
we can find five main types of hierarchies. 
From the starting point “ koto (matter)”, 
-The hierarchies of “men (side), inshou (impres-
sion), kanji (feeling), kibun (mood), kimochi 
(feeling)” 
-The hierarchies of “men (side), sokumen (one- 
side), imi (meaning), kanten (viewpoint), kenchi 
(standpoint)” 
-The hierarchies of “joutai (state), teido (degree)” 
-The hierarchies of “joutai (state), jousei (situa-
tion)”  
-The hierarchies of “men (side), inshou (impres-
sion), seikaku (character) or gaikan (appear-
ance) or utsukushisa (beauty)”.  
The lines in Fig.8 are not peculiar, and appear 
in an area of the hierarchies of “seikaku (charac-
Fig.3: Hierarchies of  
“kimochi (feeling)” 
Fig.4:Hierarchies of 
“sokumen (one side)” 
Fig.5:Hierarchies of 
“teido (degree)” 
Fig8: Hierarchies of 
“kanshoku (feel)” 
Fig.6:  Hierarchies of  
“jousei (situation)” 
Fig.7:Hierarchies of 
“seikaku (character)” 
ter)” in Fig.7. As Fig.8 shows, the hierarchies of 
“men (side), inshou (impression), kanji (feeling), 
kanshoku (feel) or kansei (sensitivity)” are lo-
cated in the area of the hierarchies of “seikaku 
(character)”, above the hierarchies of “kimochi 
(feeling)” in Fig.3. 
5. Comparison of hierarchies of super-
ordinate nouns of adjectives. 
We compare the hierarchy mentioned above 
with ones obtained from two kinds of data. 
1) Hierarchies obtained by: 
CSM and Yate’s correction 
corpus occurrence data (no frequency). 
2) Hierarchies obtained by: 
Tf.CSM and Yate’s correction 
corpus frequency data. 
3) Hierarchies obtained by: 
Overlap coefficient and Yates' correction 
corpus occurrence data (no frequency). 
 
As both CSM and the Overlap coefficient are 
“measures of inclusion”, we compared CSM and 
Tf.CSM with the Overlap coefficient. 
The number of groups that were obtained by 
CSM, Tf.CSM and the Overlap coefficient are 
the following. 
Table 2. Total number of groups obtained from CSM, 
Tf.CSM and Ovlp (Overlap) 
 groups 
CSM 161 
Tf.CSM 158 
Ovlp 240 
The Depth of hierarchies obtained from CSM, 
Tf.CSM, and the Overlap coefficient are as fol-
lows: 
Table 3. The hierarchy depth for CSM, Tf.CSM,  
and the Overlap coefficient 
 
In the case of CSM, there are 32 groups at 
depth 7, which is the greatest number of groups. 
The greatest concentration of groups is at depth 5 
to 10. In the case of Tf.CSM, the greatest number 
of groups is 25 at depth 8. The greatest concen-
tration of groups is at depth 5 to 13. In the case of 
the overlap coefficient, the greatest number of 
groups is 61 at depth 5. The greatest concentra-
tion of groups is at depth 3 to 7. 
0
10
20
30
40
50
60
70
345678910112131415
CSM
Tf.CSM
Ovlp
 
 
 
From this result, we can see that hierarchies 
generated by Tf.CSM are relatively deep, and 
those generated by the Overlap coefficient are 
relatively shallow.  
In the case of the Overlap coefficient, abstract 
nouns in lower layers are sometimes directly re-
lated to abstract nouns in the highest layers. On 
the other hand, in hierarchies generated by CSM 
and Tf.CSM, abstract nouns in the highest layers 
are related to those in the lowest layers via ab-
stract nouns in the middle layers. The following 
indicates the number of overlapping hierarchies 
for CSM, Tf.CSM and Overlap. 
Table 4. The number of overlapping hierarchies 
among CSM, Tf.CSM and Overlap 
CSM&Tf.CSM 37 
CSM&Ovlp 7 
Tf.CSM&Ovlp 2 
CSM&Tf.CSM&Ovlp 7 
The hierarchy generated by Tf.CSM is the 
deepest, and includes some hierarchies generated 
by CSM and the Overlap coefficient. The hierar-
chy generated by CSM is more similar to the one 
made by Tf.CSM than that for the Overlap coef-
ficient: the number of completely corresponding 
hierarchies for CSM and Tf.CSM is 37, that for 
CSM and the Overlap coefficient is 7, and that 
for Tf.CSM and the Overlap coefficient is 2. The 
total number of hierarchies that correspond com-
pletely between CSM, Tf.CSM and the Overlap 
coefficient is 7, and the number of hierarchies 
which are generated by two of the methods and 
included in the third is 57. 
depth 3 4 5 6 7 8 9
CSM 0 3 16 27 32 23 23
Tf.CSM 1 5 10 18 13 25 11
Ovlp 32 56 61 57 21 7 2
depth 10 11 12 13 14 15 
CSM 19 7 3 4 3 1 
Tf.CSM 24 13 14 14 7 2 
Ovlp 2 0 0 0 0 0 
Figure 9. Distribution of hierarchy depth for CSM, 
Tf.CSM, and Overlap coefficient 
We investigated these 64 hierarchies precisely, 
checking adjectives appearing at each depth as 
indicated by an abstract noun in this paper.  In 6 
of these hierarchies, the same adjectives were 
found at all levels of the hierarchy. In 14 of the 
remaining 58 hierarchies, the same adjectives 
were found in all but the deepest level.  These 
20 hierarchies are the most plausible in the strict 
sense of the word. Below, we give examples of 
these hierarchies. In the next stage of this re-
search, we intend to investigate the remaining 44 
hierarchies to determine the reason for the differ-
ence in adjective content. 
The common hyperonym: koto (matter) --- 
men1 (side) --- 
sokumen (one side) --- 
imi (meaning) --- 
kanten (viewpoint) --- 
me2 (eyes) --- 
mikata (view) --- 
hyouka (evaluation) --- 
ippou (while or grow -er and er) --- 
ikioi (force) --- 
sokudo (speed) --- 
jikoku (time) --- 
6. Conclusion 
We have suggested how to make a hierarchy 
of adjectives automatically by connecting 
strongly-related abstract nouns in a top-down 
fashion. We generated a word hierarchy from 
corpus data by using a combination of two 
methods: a self-organizing semantic map and a 
directional similarity measure. As our directional 
similarity measure, we utilized the complement-
ary similarity measure (CSM). Then we com-
pared the hierarchy generated by CSM with that 
generated by Tf.CSM and the Overlap coefficient. 
In the case of Tf.CSM, the hierarchy is deeper 
than the others because there are more abstract 
nouns in the middle layer. In the case of the 
Overlap coefficient, the hierarchy is shallow, but 
there are more hyponyms in the lower layer than 
with the other two methods. As a result, the 
hierarchies generated by CSM have more com-
mon hierarchical relations than those generated 
by the other two methods. In future work, we will 
analyze common hierarchies made by the three 
methods in detail and examine differences among 
them in order to generate an abstract conceptual 
hierarchy of adjectives. We will then compare 
our hierarchy with thesauri compiled manually. 
After we have completed the experiment on Jap-
anese adjectives, we are keen to investigate dif-
ferences and similarities in adjective hypero-
nyms between Japanese and other languages such 
as English by means of our method. 
Acknowledgement 
We would like to thank Dr. Masaki Murata of 
NICT for allowing us to use his drawing tool. 

References  

Nemoto, K. 1969. The combination of the noun with 
“ga-Case” and the adjective, Language research2 
for the computer, National Language Research In-
stitute: 63-73/ 

Takahashi, T. 1975. A various phase related to the 
part-whole relation investigated in the sentence, 
Studies in the Japanese language 103, The society 
of Japanese Linguistics: 1-16. 

Kohonen, T. 1995. Self-Organizing Maps, Springer. 

Hindle, D. 1990. Noun Classification From Predicate-
Argument Structures, In the Proceedings of the 28th 
Annual Meeting of the Association for Computa-
tional Linguistics: 268-275 

Hatzivassiloglou,V. and McKeown,R.K. 1993. To-
wards the Automatic Identification of Adjectival 
Scales: Clustering Adjectives According to Mean-
ing, In the Proceedings of the 31st Annual Meeting 
of the Association for Computational Linguistics: 
172-182.  

Hagita, N. and Sawaki, M. 1995. Robust Recognition 
of Degraded Machine-Printed Characters using 
Complimentary Similarity Measure and Error-
Correction Learning| In the Proceedings of the 
SPIE –The International Society for Optical Engi-
neering, 2442: 236-244. 

Yamamoto, E. and Umemura, K. 2002. A Similarity 
Measure for estimation of One–to-Many Relation-
ship in Corpus, Journal of Natural Language Proc-
essing: 45-75.  

Hans-Jorg Shmid. 2000. English Abstract Nouns as 
Conceptual Shells, Mouton de Gruyter. 

Kanzaki, K., Ma., Q. and Isahara, H. (2000), Similari-
ties and Differences among Semantic Behaviors of 
Japanese Adnominal Constituents, In the Proceed-
ings of the Syntactic and Semantic Complexity in 
Natural Language Processing Systems, ANLP and 
NAACL. 

Ma, Q., Kanzaki, K., Murata, M., Uchimoto, K. and 
Isahara, H. 2000. Self-Organization Semantic Maps 
of Japanese Noun in Terms of Adnominal Constitu-
ents, In Proceedings of IJCNN’2000, Como, Italy, 
vol.6.: 91-96. 
