CONCEPTUAL TAXONOMY OF JAPANESE ADJECTIVES FOR UNDERSTANDING 
NATURAL LANGUAGE AND PICTURE PATTERNS 
Naoyuki OKADA, Aiko MIURA 
Information Science and Systems Engineering 
Oita University, Oita 870-11, Japan 
I. Introduction 
This paper presents a conceptual taxonomy of Japanese 
adjectives, succeeding that on Japanese verbs'. In this taxo- 
n~ny, natural language is associated with real world things - 
- matter, events, attributes - and mental activities - spirit- 
ual and sensual. Adjective concepts are divided into two 
large classes, simple and non-simple. Simple concepts cannot 
be reduced into further elementary adjective concepts, whereas 
non-simple ones can be. Roughly speaking, simple concepts are 
concrete and can be directly associated with physical and 
mental attributes, whereas non-simple ones are abstract and in- 
directly associated with them. 
2. Simple Adjective Concepts 
Verb concepts were well understood as "change" fro~ state 
S O to state S 1 as shown in Fig. 14 Adjective concepts are con- 
sidered to be captured as the "difference" between objects O O 
and 0 I. Ylg.2 shows how the difference in vertical length be- 
tween 00 and 01 brings about the concept of "high". Notice 
that surface structures often lack the expression of 00 like 
"yama-ga takai (the mountain is high)". Since the meaning of 
"high" cannot be expressed only by O 1, deep structures need 
O 0 as an object for comparison. 
- 209 - 
change 
State So State S, otoko-ga ie-kara deru. 
The man Koes out o-~hehouse. 
Fig. 1 The structure of verb concept 
jence 
Object Ol Object O0 
y~.a-ga ie-yori takai. 
The mountain is ~ than the house. 
Fig. 2 The structure of adjective concept 
2.1 Structural Patterns 
Adjective concepts can be regarded as the concept of a 
relation among constituents which produce difference and can 
be represented as follows: 
Ad(C, e, o, of, Or, o m, os, i, l, t, r, ...) (1) 
where each symbol in parentheses stands for a constituent. For 
example, c and s stand for an object for ccmparllon and a sub- 
~ect, respectively. Scae constituents are obligatory in the 
sense that they are indispensable in the recognition of differ- 
ence. Seven types of structural patterns are proposed, accord- 
ing to the combination of obligatory constituents. 
2.2 Semantic Contents 
The semantic contents of Ad in formula (1) are categoriz~ 
ed into physical, mental and abstract attribute. We classify 
each of them in detail because we cannot detect aDy difference 
- 210 - 
v 
without such information. In the case of verbs, twenty cate- 
gories such as displacement, defo~nstion, and spiritual act- 
ivity were proposed to classify the semantic contents of 
"changes". We modify them elightly to apply to the classific- 
ation of semantic contente of adjectives. For example, die- 
placement and deformation are modified to location and shape, 
respectively. 
3. Non-Simple Adjective Concepte 
Non-simple adjective concepts are not directly associat- 
ed with p~eisal or mental attributee~ therefore we emphasize 
an analysis of how they ere composed of simple ones. 
3.1 Complex Concepts-A 
For example, there is a complex adjective "hoso-nagai 
(thin-long)" which is registered as one word in ordinary 
Japanese dictionaries. The conne0ti ~ relation between two 
adjective concepts is a logical product in this case. The 
complex concept of A is defined ss follows: if two adjective 
concepts are connected by a rule and the connected concept is 
represented by a complex word of those adjectives, it is call- 
ed a "complex concept of A". Two types of connecting rules 
are proposed: logical product and syntactic relation. 
3.2 Complex Concepts-B 
This concept is usually represented by a simple word and 
is considered to consist of several elementary adjective con- 
cepts. No general rule can be found to connect elementary ad- 
Jective concepts. Semantic contents of the concepts are class- 
ified into such categories as linguistic activ£ty and character. 
3.3 Derivative Concepts 
Some concepts possess e function of deriving s new con- 
cept by operati~ others. They have two morphemic strUctttres, 
prefix ~ and ett~fix, which add certain i~ormetion to operands. 
For example, st~ffix "-tarash~" adds the meaning "indeed" to 
its operand. Such derivative information is categorized into 
five classes. 
- 211 - 
4. Classification 
There are approximately 530 adjectives considered suffic- 
ient for daily Japanese. We actually classified their con- 
cepts, n~nbering about 630, according to types and categories 
proposed in Sec.2 and 3. They were well classified into those 
types and categories. In the classification of simple concepts, 
"standardconcepte" were chosen from among similar concepts, 
amounting to approximately 270. They might be considered 
"primitives" in the sense that every other adjective concept 
can be derived from them by some rule. 
5. Extraction of Conceptual Features 
We extract conceptual features from each A d and its 
constituents, noticing the standard concepts with physical 
attributes obtained in Sec.4.The following analysis shows par@ 
of the feature extraction from taka..._~i (high)/ hikui (low). 
taka__~Chigh) / ~i_k_~Clow) 
Polysm, y (1) Structural pattern~ Ad(S,C), semantic 
contents; shape 
s, c : thing 
A d : The vertical length of s is larger/smaller 
than that of c. 
Example: TSky5 taw~-ga Ky6to taw~-yori takai. 
Tokyo Tower is higher than F~voto Tower. 
Polyeamy (2) Structural pattern;Ad(s,c~, semantic 
contents! location 
s, c : thing 
A d : The vertical distance of s from a standard plane is larger/smaller than that of c. 
Example: Hik~ki-ga Kumo-yori takai. 
The airplane is higher than the cloud. 
6. Sentence Production from Picture Patterns 
System SUPP, which accepts moving picture patterns, in- 
terprets the meaning of changes in them, infers about simil- 
arity among changes, end describes the results in Japanese 
and English sentences, was constructed as an application of 
- 212 - 
the conceptual taxonomy of verbs. SUPPwas improved to inter- 
pret physical attributes of piotnre patterns. Fourteen stan- 
dard concepts were chosen for this improvement. An example of 
experimental results ~a shown in F~4r.). 
(a) Input pattern 
Note. In I) through 3), input 
picture "onna" is compared to 
its model in machine. 
Fig. 3 
i) onna-ga nagai. 
The woman is long. 
2) onna-ga takai\[l). 
The woman is ~igh\[l\]: 
3) onna-ga 8k~. 
The woman is large. 
4) onna-ga ki-yori hikuiCl). 
The woman is qower\[l\]" than the tree. 
5) onna-ga ki-yori hlkui(2). 
The woman is lower\[2\] than the tree. 
6) onnu-ga ki-yorl ch~sqi. 
The woman is smaller than the tree. 
7) onna-ga tori-yori takaiC1). 
The woman is'higher\[l\]" than the bird. 
8) onna-gu tori-yori hikuiC2). 
The woman is lower\[2\] than the bird. 
9) onna-ga tori-yori Bki. 
The woman is larger than the bird. 
(b) Output sentences 
Sentence production from a picture pattern 
7* ConcltwioDa 
This taxonomy has made clear an outline of the whole of 
Japanese adjective concepts. I~ Japanese, the number of ad- 
Jectives is rather small compared to that of adjective verbs. 
This taxonomy is expected to give a basis to that of adject- 
ive verbs. 

Reference 

(1) N. Okada: eConceptual taxonomy of Japanese verbs for 
understanding natural language and picture patterns", 
l~'oc° COLING 80, Tokyo, 1980. 
