a0a2a1a4a3a6a5a7a5a9a8a11a10a4a8a11a12a7a3a14a13a15a8a11a16a18a17a19a16a18a10a18a20a22a21a24a23a26a25a18a12a27a13a15a8a29a28a30a3a31a1a32a3a31a17a33a21a35a34a35a16a18a17 -a3a6a21a24a23a26a25a18a12a27a13a15a8a29a28a30a3a31a1a30a34a35a16a24a36a33a17a33a5a38a37a39a3a6a5a9a25a14a21a35a16a18a17a40a13a15a41a42a25a18a8a11a43
a44a33a25a18a45a46a3a31a17a32a13a15a8a11a12a47a37a47a25a18a41a33a3a14a28a6a8a11a16a18a43a49a48a30a50a52a51a40a5a9a8a11a17a33a53a52a3a35a44a33a25a18a1a11a10 -
a54
a43a7a53a31a3a31a17a42a8a4a55a7a8a11a17a33a53a52a44a33a25a18a45a46a3a31a17a32a13a15a8a11a12a2a56a57a3a14a58
Kyoko Kanzaki  Qing Ma  Masaki Murata  Hitoshi Isahara 
 
Computational Linguistics Group 
Communications Research Laboratory 
{kanzaki, qma, murata, isahara}@crl.go.jp 
 
 
Abstract 
We treat nouns that behave adjectively, which we 
call adjectival nouns, extracted from large corpora. 
For example, in “financial world” and “world of 
finance,” “financial” and “finance” are different 
parts of speech, but their semantic behaviors are 
similar to each other. We investigate how adjectival 
nouns are similar to adjectives and different from 
non-adjectival nouns by using self-organizing 
semantic maps. We create five kinds of semantic 
maps, i.e., semantic maps of abstract nouns 
organized via (1) adjectives, (2) adjectival nouns, 
(3) non-adjectival nouns and (4) adjectival and 
adjectival nouns and a semantic map of adjectives, 
adjectival nouns and non-adjectival nouns organized 
via collocated abstract nouns, and compare them 
with each other to find similarities and differences. 
1 Introduction 
In this paper, we propose a method for fundamental 
research to construct an organized lexicon, in which 
we classify words depending on not only their part 
of speech, but also their semantic categories. We 
applied both a neural network model and a linguistic 
method, that is syntactic information, to a large 
corpora and extracted necessary information. To 
extract semantic information of words such as 
synonyms and antonyms from corpora, previous 
research used syntactic structures (Hindle 1990, 
Hatzivassiloglou 1993 and Tokunaga 1995), 
response time to associate synonyms and antonyms 
in psychological experiments (Gross 1989), or 
extracting related words automatically from corpora 
(Grefensette 1994). Most lexical classification is 
based on parts of speech, as they have very 
important semantic information. For examples, 
typically, an adjective refers to an attribute, a verb 
refers to a motion or an event, and a noun refers to 
an object. However, in real data, a semantic function 
of a part of speech is not defined rigidly, as shown 
in the above examples. In spite of different parts of 
speech, they sometimes represent the same or very 
similar semantic functions. For examples, there are 
the following Japanese examples:  
yuushuu_na       seiseki 
 (excellent)   (an academic record) 
an excellent academic record    
sugure_ta               seiseki 
(excel and suffix of “adnominal”) (an academic record)  
an excellent academic record  
“Yuushuu_na (excellent)” is an adjective and 
“sugure_ta (excel)” is a verb, but they represent the 
same meaning and same semantic function, that is, 
an evaluation of an academic record. In English 
there are the following examples;  
   financial world 
   world of finance  
In these examples, “financial” and “finance” are 
different part of speech, but represent same meaning 
and same semantic function, that is, one of domains.  
On the other hand, there are examples in which 
only semantic function is the same, but the part of 
speech and meaning of the words are different. For 
examples,  
   kandai_na    kihuu      no    hito 
    (gentle)   (disposition)  (of)  (person) 
       a gentle person  
   shinshu   no    kihuu     no    hito 
  (initiative)  (of) (disposition)  (of)  (person) 
       a person of initiative  
In Japanese “kandai_na (gentle)” is an adjective 
and “shinshu (initiative)” is a noun. They have 
different parts of speech and meanings, but the same 
semantic function, that is, they represent 
characteristics of a person. In terms of a semantic 
function of representation of characteristics, both 
“kandai_na (gentle)” and “shinshu (initiative)” are 
classified in the same category. In this work we call 
this type of noun an “adjectival noun.” 
It is important for developing high quality natural 
language processing systems to establish an 
objective method to represent relationship between 
words not only by part of speech but also by 
semantic functions. However, it is very difficult to 
extract this type of linguistic phenomena from real 
data automatically. We used syntactic and semantic 
patterns in our previous work (Isahara and Kanzaki 
1999) in order to extract these types of examples 
from large corpora semi-automatically. In this work, 
by using syntactic information, we are collecting 
adjectives and adjectival nouns in the “noun + NO 
(of + Noun)” structure that we supposed to have the 
same semantic functions. We examined how 
adjectives and adjectival nouns extracted from 
corpora are similar or different in the real data and 
how non-adjectival nouns unlike adjectival nouns 
are different from adjectives in order to verify the 
usefulness of self-organizing semantic maps for 
lexical semantics. 
In Section 2, we explain our methodology, based 
on linguistic information. In Section 3, we describe 
a self-organizing semantic map. In Section 4, we 
describe the similarities between adjectives and 
adjectival nouns and the differences between 
adjectival nouns and non-adjectival nouns by 
comparing two different self-organizing semantic 
maps. In Section 5, we give our conclusion. 
2 Methodology 
Isahara and Kanzaki (1999) classified semantic 
relations between adjectives and their head nouns 
from the viewpoints of syntax, semantics and 
computational treatment. Among various types of 
semantic relations extracted in this research, there is 
a case in which the meanings of adnominal 
constituents are semantically similar to the features 
of their head nouns. Let us consider the Japanese 
phrases, “kanashii kimochi (sad feeling)” and 
“yorokobi no kimochi (feeling of delight)” as 
examples.  
kanashii kimochi 
(sad)  (feeling) 
{adjective} {noun} 
sad feeling  
yorokobi no     kimochi 
(delight)  (of)      (feeling) 
{noun}   {postpositional}  {noun}  
feeling of delight 
(The English translation of the “noun + no” 
examples should be read from right to left.) 
One meaning of “kimochi (feeling)” represents 
the semantic element, [mental state]. In the above 
examples, the adjective, “kanashii (sad),” and “noun 
+ no” structure, “yorokobi no (delight + no),” 
represent the concrete contents of their head noun 
“kimochi (feeling),” i.e. they are descriptors of the 
mental state: “kimochi (feeling).” The head noun, 
“kimochi (feeling),” is a cognate object for 
“kanashii (sad)” and “yorokobi no (delight + no).” 
Therefore, even though “kanashii (sad)” and 
“yorokobi no (delight + no)” belong to different 
parts of speech (adjective and noun phrase), they 
must be classified as the same semantic category, 
since both carry the same type of meaning. 
As for data, necessary expressions are extracted 
from large corpora: 10 year’s worth of Japanese 
newspapers — the Mainichi Shinbun from 1991 to 
2000, 100 novels — Shincho-bunko, and 100 kinds 
of essays. We extracted 134 abstract nouns used as 
this kind of head noun semi-automatically by using 
syntactic patterns that Isahara and Kanzaki(1999) 
and Kanzaki et al. (2000) used in their paper. The 
total number of adnominal constituents appearing 
with these head nouns in the corpora was 47,248, 
and the number of different adnominal constituents 
was 28,063. We got the list of pairs of a head 
(abstract) noun and its adnominal constituents 
(Table 1). These adnominal constituents are 
classified into three types, i.e. adjectives, adjectival 
nouns and non-adjectival nouns. 
 
Table 1: Example of gathered data 
Noun Adnominal constituents 
kimochi 
(feeling) 
shiawasena (happy),  
hokorashii (proud), 
kanashii (sad), … 
joutai 
(status) 
aimaina (vague),  
ansei no (repose + no), …  
kanten 
(viewpoint) 
gakumontekina (academic), 
anzensei no (safety + no), … 
… … 
 
We classified these head nouns according to the 
similarities of sets of their adnominal constituents 
by using a self-organizing system in a neural 
network model. This means that we co-classified 
both head nouns, i.e. abstract nouns, and adnominal 
constituents at the same time. 
3 Self-Organizing Semantic Map 
In this section, we explain self-organizing 
semantic maps by a neural network model. For the 
analysis of the similarities between adjectives and 
adjectival nouns, we make some semantic maps 
based on these adnominal constituents. We use a 
self-organizing semantic map to classify words 
because it distributes words onto a two-dimensional 
plane and is therefore a visible and continuous 
representation. This feature is very feasible to 
classify word meanings, because they cannot be 
always classified into an explicit category as 
hierarchical clustering does. As for the clustering 
ability of self-organizing semantic map compared 
with the multivariate statistical analysis and 
hierarchical clustering method, it is almost the same 
as the hierarchical clustering method and superior to 
multivariate statistical analysis (Ma 2001). 
The semantic map we construct in this paper is 
one on which nouns, with their adnominal 
constituents as attributes, are mapped in a semantic 
order; i.e. nouns with similar meanings are mapped 
on (i.e. best-matched by) nodes that are 
topographically close to each other, and words with 
meanings that are far apart are mapped on nodes 
that are topographically far apart. 
3.1 Learning Data 
As we mentioned above, we used the list in Table 1 
as learning data. Table 1 shows some example data 
that was gathered manually and in which the 
adnominal constituent is a descriptor of its head 
noun, i.e. a kind of cognate noun. 
3.2 Encoding 
The semantic map of nouns is constructed by first 
defining each noun as the set of its adnominal 
constituents. From Table 1, for example, we can 
define “kimochi (feeling)” as the set of its 
adnominal constituents, i.e. “kimochi” = 
{“shiawasena (happy),” “hokorashii (proud),” 
“kanashii (sad),” “kinodokuna (unfortunate),”…}. 
Suppose there is a set of nouns w i ( i = 1, … , a0 ) 
that we are planning to use for self-organizing. Any 
noun w i can be defined by a set of its adnominal 
constituents as 
 
w i = { a1( i ) , a2( i ) , … , aa1i( i ) }      -------(1) 
 
where a j( i ) is the jth adnominal constituent of wi  
and a2 i  is the number of adnominal constituents of 
wi . One method of encoding nouns so that they can 
be treated by SOM is to use random coding, which 
is a common method used for constructing SOMs 
(see details in Kohonen (1997)). By several 
preceding computer experiments, however, we 
found that this method is not suitable for our task. 
We therefore used a new method as described 
below. 
Suppose we have a correlative matrix (Table 2) 
where d i j is some metric of correlation (or distance) 
between nouns w i and w j . We can encode noun w i 
from the correlative matrix as 
V (w i ) = [ d i 1 , d i 2 , … , d i 
a3 ] 
T .  ---------(2) 
 
The V (w i ) a4  ℜ a3 is the input to the SOM, i.e. x 
= V (w i ) and n = a5 . 
 
Table 2: Correlative matrix of nouns 
 w 1  w 2 …  wa6 
w 1     
w 2     
  
 
 
 
 
wa6     
 
d 11   d 12 …   d 1a6 
d 21   d 22 …  d 2a6 
 
 
 
 
 
da61   da62 …  d wa6 
 
 
In this paper, di j  is measured by 
 
 
                                 If i = j 
d i j =                               -------(3) 
 
 
 
where a7 i and a7 j are respectively the numbers of 
the adnominal constituents of wi and wj , and cij is the 
total number of common adnominal constituents of 
both wi and wj . The term di j is therefore a 
normalized distance between wi and wj in the context 
of the number of adnominal constituents they have 
in common; the smaller di j is, the closer wi and wj 
are in terms of their adnominal constituents. 
4 Experimental Result 
4.1 Comparisons of Word Distribution on 
Semantic Map via Adjectives with ones 
via Adjectival Nouns and via 
Non-adjectival Nouns. 
In Section 4, we examine adjectival nouns extracted 
from corpora, whose behaviors are similar to 
adjectives. In order to verify the data extracted from 
corpora manually by using syntactic method we 
prepare for four kinds of self-organizing semantic 
maps. One is a semantic map of head nouns 
( a7 i 
a8
c i j ) + ( a7 j 
a8
c i j ) 
 a7 i + a7 j 
a8
c i j 
0,             otherwise 
a9
co-occurring with adjectives the second is a 
semantic map of head nouns co-occurring with 
adjectival nouns that we extracted from corpora, the 
third is a semantic map of head nouns co-occurring 
with non-adjectival nouns and the final one is a 
semantic map of head nouns co-occurring with both 
adjectives and adjectival nouns. As we mentioned in 
section 2, head nouns distributed on four maps are 
abstract nouns that represent the concrete content of 
adnominals, e.g., “feeling” co-occurring with 
“happy” and so on. 
We compare a semantic map of head nouns via 
co-occurring adjectives (Figure 1) with three other 
maps, that is, a semantic map via adjectival nouns 
and a semantic map via non-adjectival nouns and a 
semantic map via adjectives + adjectival nouns. And 
then we mark the points of words that are similarly 
distributed between the semantic map via adjectives 
and one of other maps (Figure2, 3, 4).  
  Input data for neural network model was a list 
that we mentioned in section 2. In the ordering 
phase, the number of learning steps was 10,000, and 
in the adjustment phase it was 100,000. After 
learning the input data, a two-dimensional array, in 
which a hexagonal topology type of neighborhood, 
the area that the winner node influences in the 
learning stage was used. A self-organizing semantic 
map of head nouns via adjectives is shown in Figure 
1. We translate some Japanese words on the map 
into English for the reader’s convenience. 
 
Figure 1. Self-organizing semantic map of head 
nouns via adjectives 
 
 
 
  
 
 
 
 
 
 
 
 
Figure 2. A semantic map via adjectives marked 
via comparison with classification by adjectival 
nouns. 
 
For examples, “viewpoint,” “standpoint,” “side” 
on the right hand in Figure 1 are co-occurring with 
“medical,” “musical,” “economical,” “political” and 
so on, and “mind,” “thought,” “mood” are 
co-occurring with “delightful,” “sad,” “happy,” 
“proud” and so on.  Some sets of head nouns are 
not classified enough because the number of the 
co-occurring adjectives is not enough, or we could 
not extract enough of the collocation that we treated. 
After we made a semantic map of head nouns via 
adjectival nouns we compared it with a semantic 
map of head nouns via adjectives (Comparison1). 
We marked the words in Figure 1 located similarly 
between two semantic maps (See Figure 2). We 
defined “the common sets of words” as words 
located similarly between two semantic maps, i.e., 
marked words, and they are located within three 
neighborhoods on the semantic map (See also 
Figure 2).  
And we also examined the data of non-adjectival 
nouns as same as the above experience and we 
marked the common sets of words between two 
maps, that is, a semantic map via non-adjectival 
nouns and a semantic map via adjectives 
(Comparison 2).  
Each square on Figure 2 and 3 refers to a word in 
Figure 1. The black marked squares indicate a 
common words appearing on both a map via 
adjectives and a map via another data and a circle 
surrounding squares is common sets of words on 
both maps. 
In Figure 2 we marked 51 words among 134 
abstract nouns (38% of all the abstract nouns) on a 
semantic map via adjectives (Figure 1), which are 
common in a classification of words on two 
self-organizing semantic map, i.e., two semantic 
a0a2a1a4a3a2a5a7a6a8a3a2a9
a6a8a1a11a10a13a12a14a0a16a15a14a15a17a6a8a3a2a9
a15a17a6a19a18a16a0
a10a13a3a2a6a8a9a20a5
a1a4a6a8a9a21a18
a5a7a22a20a3a2a23a21a24a25a22a20a5
a1a4a3a2a3a25a18
a22a20a0a25a26a25a12a14a5
a27
a0a25a26a25a23a20a5a7a28
a29a14a22a21a26a25a12a14a1
a15a17a22a21a26a16a18a16a3a2a30
a6a8a9a4a3a2a9a20a0a2a31a15a32a10a13a33a19a26a25a29a14a0
a15a17a0a2a9a20a5a7a6a8a1a4a0a2a9a20a5
a10a20a26a34a15a14a15a17a6a8a3a2a9
a6a8a9a20a29a14a33a8a6a8a9a21a26a25a5a7a6a8a3a2a9
a33a8a3a2a3a2a35
a24a25a0a16a15a17a5a7a23a20a12a14a0
a12a14a0a2a1a4a6a8a9a21a18a16a0a2a12
a15a17a5a7a28a16a33a8a0
a26a16a18a25a36a13a26a25a9a20a5a17a26a16a24a25a0
a37
a6a19a24a25a23a20a12a14a0
a0a2a9a20a5a7a22a20a23a32a15a17a6a19a26a34a15a17a1
a5a17a26a25a33a8a0a2a9a20a5
a26
a27
a6a8a33a8a6a8a5a7a28
a36a34a6a8a0a2a30a38a10a13a3a39a6a40a9a13a5
a15a17a5a17a26a25a9a21a18a25a10a13a3a14a10a34a6a40a9a13a5
a15a17a6a19a18a16a0
a37
a6a8a0a2a33a19a18
a37
a12a39a26a25a1a4a0a2a30a41a3a2a12a14a35
a26a25a12a14a0a25a26
a15a17a5a17a26a25a5a7a23a20a12a14a0
a6a8a1a11a10a13a33a8a6a8a29a39a26a25a5a7a6a8a3a2a9 a15a17a1a4a0a2a33a8a33
a29a14a3a2a33a8a3a2a12
a0
a37a19a37
a0a2a29a14a5
a29a42a6a40a12a42a29a42a23a13a1a43a15a7a5a17a26a2a9a13a29a42a0
a29a14a3a2a12a14a9a20a0a2a12
a5a7a22a20a0a44a1a4a3a2a1a4a0a39a9a13a5
a3
a37
a5a7a12a14a23a20a5a7a22
a29a14a22a21a26a25a12a39a26a25a29a14a5a7a0a2a12
a6a8a12a14a3a2a9a32a15a17a6a19a18a16a0
a33a8a6
a37
a0
a5a7a0a2a1a11a10a13a0a2a12
a33a8a0a39a36a34a0a2a33
a26a34a15a7a10a13a0a2a29a14a5
a10a13a3a16a15a17a6a8a5a7a6a8a3a2a9
a6a8a9a21a18a16a6a8a29a39a26a25a5a7a6a8a3a2a9
a30a41a22a20a0a2a9
a10a13a3a16a15a14a15a17a6
a27
a33a8a0
a3a2a12a39a18a16a0a2a12
a18a16a0a25a24a25a12a14a0a2a0
a15a17a5a17a26a16a24a25a0
a15a17a5a17a26a25a28a15a17a5a17a26a25a5a7a0
a6a8a9
a29a14a3a2a9a21a18a16a6a8a5a7a6a8a3a2a9
a26a2a10a34a10a13a0a25a26a25a12a39a26a25a9a20a29a14a0
a37
a0a2a0a2a33a8a6a8a9a21a24
a27
a26a16a18a45a29a14a3a2a9a21a18a16a6a8a5a7a6a8a3a2a9
a26a25a9a20a29a14a6a8a0a2a9a20a5
maps via adjectives and via adjectival nouns and 
these 51 words can be classified into 16 common 
sets of words. 
 
 
Figure 3. A semantic map via adjectives marked 
via comparison with classification by 
non-adjectival nouns. 
 
Then, we compared the semantic maps organized 
via non-adjectival nouns with semantic map via 
adjectives to find how different these maps are. 
Thirty-five marked words from 134 abstract nouns 
(26%) and 14 common sets of these words, which 
are common between two maps, i.e., semantic maps 
via adjective and via non-adjectival nouns, are 
distributed on semantic maps via adjectives (Figure 
3).   
There are 12% more common words and 3 more 
common sets in Comparison 1 than in Comparison 
2. However, there is a question of why the map 
organized via non-adjectival nouns still has sets of 
words common to the map organized via adjective. 
Are there any similarities of behaviors between 
adjectives and non-adjectival nouns? We 
investigated the common co-occurring head nouns 
in Comparison 2 precisely, and found two facts that 
caused the existence of these common sets of words 
in Comparison 2. 
One is that some co-occurring words that we 
classified as non-adjectival nouns are nouns that we 
must classify as adjectival nouns. Another is that 
some non-adjectival nouns refer to people and they 
are possessors of the modified abstract nouns. For 
examples, “emotion,” “mood” and “thought” are 
common sets in both maps. Co-occurring adjectives 
are “delight,” “sad” and “happy,” however, 
co-occurring non-adjectival nouns are “watashi-no 
(my),” “haha-no (mother’s),” “sensei-no 
(teacher’s),” and so on. From this fact, we can 
conclude that the existence of common 
classifications of head nouns between these two 
semantic maps does not always mean semantic 
similarity between adjectives and non-adjectival 
nouns. 
From these observations we made a semantic 
map of head nouns by using both adjectives and 
adjectival nouns. If the adjectival nouns work 
similarly to the adjectives, using both adjectives and 
adjectival nouns will not influence the distribution 
and classification of words on the semantic map via 
adjectives. On the other hand, if the data of semantic 
phenomena between adjectives + adjectival nouns 
and adjectives only are completely different, the 
distribution and classification of words on the 
semantic map via adjectives will be influenced by 
the addition of adjectival nouns. We mark the point 
of the common words between them on the 
semantic map via adjectives (Figure 4). 
 
 
 
 
 
 
 
 
 
 
 
Figure 4. A semantic map via adjectives marked 
by the common words between classification 
by adjectives and adjectival nouns and 
classification by only adjectives. 
 
 
Eighty-three words among 134 words on this map 
are classified similarly to the words on the map 
organized via adjectives, and there are 21 similar 
sets of words. This result shows that the distribution 
of the abstract nouns on the semantic map is not 
affected by the addition of adjectival nouns. 
Therefore, the semantic roles of adjectival nouns for 
abstract nouns are similar to those of adjectives. 
4.2  A Semantic Map Distributed by 
Adjectives, Adjectival Nouns and 
Non-adjectival Nouns Organized via 
Head Nouns. 
In this section we made the semantic map of 
adjectives, adjectival nouns and non-adjectival 
nouns organized via collocation with abstract nouns 
to see the semantic distances between them. As for 
marks on the map, a0 , a1  and a2  indicated, in 
turn, adjective, adjectival nouns, and non-adjectival 
nouns. 
 
 
 
 
Figure 5. Semantic map of adjectives, adjectival 
nouns and non-adjectival nouns organized via 
collocation with abstract nouns. 
 
 
We distributed three kinds of words, that is, 
adjective, adjectival nouns and non-adjectival nouns 
on the semantic map based on their head nouns, that 
is, abstract nouns. For example, “happy” has 
“feeling,” “mood,” “state,” and so on as 
co-occurring head nouns. When we made this map, 
we utilized words (adjectives, adjectival nouns and 
non-adjectival nouns) that collocate with 10 to 20 
abstract nouns, so that the input data for 
constructing semantic map is fair from the 
viewpoint of number of co-occurring words. We 
selected from them 100 adjectives, 100 adjectival 
nouns, and 200 non-adjectival nouns at random. 
This semantic map is shown in Figure 5. 
The semantic map shown in Figure 5 shows that 
there are three classes on the map. The upper half 
part of this semantic map indicates the adjective 
area, the bottom right half of this map is the 
adjectival noun area and the bottom left half of this 
map is the non-adjectival noun area. Semantic roles 
of adjectives are isolated from those of nouns, and 
semantic roles of nouns are divided into two areas, 
i.e. adjectival and non-adjectival. The 
self-organizing mechanism could separate the 
semantic roles of adjectival nouns from those of 
non-adjectival nouns. 
5 Conclusion 
We extracted adnominal constituents from corpora 
and created several self-organizing semantic maps 
by using them. 
First, we compared the semantic maps organized 
via adjectives and via adjectival nouns. The 
common sets of head nouns were 16 sets and 
common head nouns were 51 words in 134 head 
nouns, that is, 38% of the head nouns were 
classified similarly. 
Second, we compared the semantic maps 
organized via adjectives and via non-adjectival 
nouns. The common sets of head nouns were 14 sets, 
and the common head nouns were 35 words in 134 
head nouns, that is, 26% of the head nouns were the 
same classifications. 
Some sets of abstract nouns, head noun 
co-occurring with adjectives, are common with sets 
of abstract nouns co-occurring with non-adjectival 
nouns. However, based on the precise investigation, 
we could find that the semantic function of 
adjectives and non-adjectival nouns were different. 
Finally, we created a semantic map of abstract 
nouns by both adjectives and adjectival nouns. This 
is because we wanted to see how word distribution 
on the map changed when we added adjectival 
nouns to the data for self-organization. The common 
sets of head nouns were 21 sets and the common 
head nouns that did not change were 83 words in 
134 abstract nouns, that is, 62% of head nouns were 
not affected by the addition of adjectival nouns. This 
means that adjectival nouns are similar to adjectives 
in their semantic behavior for abstract nouns. 
Then, we showed the semantic map of adjectives, 
adjectival nouns and non-adjectival nouns organized 
via co-occurring abstract nouns. As these three 
kinds of adnominals were isolated on this map, we 
could find that the adjectival nouns had specific 
semantic roles that are different from those of 
non-adjectival nouns. 
From the above evidence, we considered that we 
a3a5a4a7a6a9a8a11a10a13a12a7a14a16a15a7a17a19a18a21a20a11a22a23a10a25a24a26a6a27a4a7a28a29a6a27a30a21a31
a12a32a14a21a15a32a17a33a18a34a20a35a22a13a10a32a24a25a6a23a4a19a28a36a6a23a30
a31
a12a32a14a21a15a19a17a37a18a11a20a35a22a38a15a32a30
could extract the adjectival nouns similar to 
adjectives, rather than non-adjectival nouns. 
In future work, we need addition and 
modification of input data and would like to use the 
accurate distribution of words by using some kind 
of information such as frequencies1. And then we 
will construct a semantic map of words from 
Japanese large corpora and link words according to 
semantic behavior while we verify our data 
extracted from corpora by using a neural network 
model.
 
References 
Hindle, D. (1990), Noun Classification From 
Predicate-Argument Structures, In the 
Proceedings of the 28th Annual Meeting of the 
Association for Computational Linguistics, pp. 
268-275. 
Hatzivassiloglou, V. and McKeown, R.K. (1993), 
Towards the Automatic Identification of 
Adjectival Scales: Clustering Adjectives 
According to Meaning, In the Proceedings of 
the 31st Annual Meeting of the Association for 
Computational Linguistics, pp. 172-182.  
Tokunaga, T. Iwayama, M. and Tanaka, H. (1995), 
Automatic Thesaurus Construction Based On 
Grammatical Relations, In the Proceedings of 
the 14th International Joint Conference on 
Artificial Intelligence (IJCAI), pp. 1308-1313. 
Gross, D. (1989), The Organization of Adjectival 
Meanings, In Journal of Memory and Language, 
vol. 28, pp. 92-106 
Grefenstette, G. (1994), Explorations in Automatic 
Thesaurus Discovery, Kluwer Academic 
Publishers. 
Isahara, H. and Kanzaki, K (1999) Lexical 
Semantics to Disambiguate Polysemous 
Phenomena of Japanese Adnominal 
Constituents, ACL  ’99. 
Kanzaki, K., Ma., Q. and Isahara, H. (2000), 
Similarities and Differences among Semantic 
Behaviors of Japanese Adnominal Constituents, 
Workshop on the Syntactic an Semantic 
Complexity in Natural Language Processing 
a0 We can use many Japanese newspapers as corpora, 
however we cannot use many kinds of texts, for 
examples, novels, essays and so on. Though we should 
try to use frequencies, if our corpora are not well 
balanced, we don’t know how much confidence to 
place on the result of frequencies. This is why we 
didn’t use the frequencies as the first step of our 
research. 
Systems, ANLP and NAACL. 
Ma, Q. Kanzaki, K., Murata, M., Uchimoto, K. and 
Isahara, H. (2001) Self-Organizing Semantic 
Map of Japanese Nouns, Information 
Processing Society of Japan, Vol. 42, No. 10. 
