--- AN APPROACH TO A SEMANTIC ANALYSIS OF METAPHOR --- 
Fumio MIZOGUCHI*, Akihiko YAMAMOTO*, 
* Department of Industrial Administration 
Tokyo University of Science 
Noda, Chiba 278, Japan 
Abstract 
The present study deals with conflict resolution process in metaphorical 
interpretation for the noun phrase. In order to make the problem mbre explicit, we 
have reviewed the knowledge representation with conflict both from cognitive 
psychology and artificial intelligence. Then, we propose a semantic model which is 
obtained from the notion of Linguistics as Chemistry. That is, the model called 
"Semistry" is introduced so as to interprete a metaphor semantic bonds between 
nouns. By using production system couped with contex free parser (ELINGOL), the 
working system called META-SIM is constructed to analyze the noun phrase metaphor. 
Finally, there are discussions on a role of metaphor in human cognitive processing. 
I. Introduction 
A conflict resolution in semantic analysis 
is regarded as an important problem in natural 
language processing. In case of a human 
cognitive system, this kind of problem was 
discussed in the realm of behavioral decision 
making theories, such as Festinger's (1957) 
theory of cognitive dissonance and Abelson's 
(1968) Psychological implication. And also, 
it was discussed in the field of cognitive 
psychology which dealt with the human 
information processing mechanism, especially 
long-term memory (LTM) representation. 
The work of Kintsch (1969) in his 
structure of semantic memory was useful for 
insight into the conflict resolution in lexical 
item stored in LTM. He particularly made his 
attention on the problems on how one 
semantically unacceptable a sentence. 
If one selects the problem from artificial 
intelligence field, McDermott's (1974) TOPLE in 
"Ring formalism" was suggestive to a design of 
inference mechanism which could interprete 
unacceptable knowledge in a simple world model. 
This formarism also guided us about a 
construction of lexical data in natural 
language processing. 
By following these current issures on 
knowledge representation, the present study 
focuses the problem of conflict resolution in 
semantic analysis of Metaphor both from 
viewpoints of cognitive psychology and 
artificial intelligence. For this purposes, 
we propose a new semantic model which can deal 
with the metaphorical interpretation together 
with the inference mechanism. And then, we 
will demonstrate examples of metaphorical 
analysis which is based on the proposed 
semantic model. 
2. Semantic representation 
with conflict knowled~ 
Metaphor plays an important role in our 
understanding of language and of the world we 
describe through language. Thus, there have 
been a number of researches on the nature of 
metaphor since the time of Aristotle. 
Recently, historical survey on metaphor was 
made by Ortony, Reynolds & Alter (1978) on 
their paper titled "Metaphor: Theoretical and 
Empirical Research". Their main concerns on 
metaphor are to develop a model of metaphoric 
comprehension both from Psychological reaction 
time study and "Schema" based theoretical 
framework. To quote their paper: 
"The structure of a shema is of a series 
of variables together with relationships among 
them. There are constraints on the values 
that the variables may take, but these 
constraints are rarely absolute, although some 
values are typical than others. This kind of 
representation system appears to offer greater 
flexibility for matching incoming information 
to previously stored knowledge, and with this 
flexibility comes a better prospect for dealing 
with nonliteral use of language. The 
metaphorical interpretation would be achieved 
by finding that schema or those schemata that 
matched the input in such a way as to minimize 
the required changes in variable range. " 
Certainly, the idea of schema described 
above is one of convergences on the knowledge 
representations proposed by Rumelhart & Ortony 
(1977) and Bobrow & Norman (1975). Here, the 
procedure for the metaphorical interpretation 
is a kind of pattern-matching process which is 
based on a semantically acceptable 
interpretation. In order to make a discussion 
more explicit, let us consider the same type of 
problem which deals with knowledge aquisition 
through the conflict or contradict resolution. 
For this purpose, we focus our attention 
to the semantic structure which is stored in an 
understanding system. 
In case of Kintsch model, this problem is 
to define a lexical entry used in the semantic 
memory. And further, the notion of 
acceptability of sentence is defined by the use 
of production rules which are applied to the 
set of propositions stored in memory system. 
If there exists a production rule for a set of 
136 
proposition, the sentence is semantically 
acceptable. And if not, the sentence is 
semantically unacceptable. Therefore, if two 
sentences are contradictory, the memory system 
must decide which one to keep and which one tO 
disregard by choosing the one with 
acceptability. In case of McDermot's TOPLE, 
this kind of contradiction is resolve through 
the notion of "ring" which can absorb the 
confliction. This process is accomplished 
through a procedure called DOUBT. By this 
procedure, the system can find the allowable 
course of action to take to patch up a ring. 
In this connection, Tanaka (1980) in his SRL, 
this procedure is carried out through the use 
of production rule called "without 
description". In SRL representation which is 
an extension of Bobrow & Winograd's KRL, 
knowledge is organized around conceptual 
entities with associated descriptions and 
procedures. Therefore, by embedding various 
procedures to knowledge, lexical item is 
represented by knowledge unit with associated 
descriptions and procedures. One type of 
conceptual entities is the use of hiearchcal 
relation which was actively utilized by the 
work of Carbonell's SCHOLAR (1970). In SRL, 
hiearchical concept is accomplished through 
part-whole and class-inclusion relations. And 
further, the conflict resolution was made by 
the use of the without description. In 
contrast to TOPLE, a bird like Pengin is 
represented by the following way. 
(PENGIN 
unit 
(self (a TORI without 
(hasp = TOBU ))) 
(part-of nil) 
other descriptions ) 
Fig. l SRL's description for pengin 
Although we made a quick overview of the 
related topic on the knowledge with conflict 
resolution, it is necessary to consider a 
semantic model which can accept semantically 
conflict knowledge. That is, any lexical item 
stored in the understanding system should 
process a change of meaning through the 
metaphorical use. 
In this section, we will propose a new 
semantic model in which semantic elements are 
compared to chemical elements. Here, chemical 
elements.are refered to the dynamic aspect of 
meaning. In a sense, the theory can be 
considered as an extension of Arnold Zwicky's 
1973 paper, "Linguistics as Chemistry", in 
Anderson & Kiparsky (Eds.), A Festschrift for 
Morris Halle. In this connection, some 
preliminary work on "Linguistic Chemistry" was 
carried out by Harada & Mizoguchi (1977) in 
which semantics and lexical elements were also 
compared to chemical elements (such as 
molecules, atoms, protons, neutrons and 
elecrtons). A large part of syntax is now 
compared to a theory of semantic "bondage". 
The semantic equivalent of 'chemical reaction' 
is a theory of semantic amalgamation. The 
analogy with chemistry may not be completely 
felicitous, but at the present moment it is a 
least useful in shaping a new theory of 
semantics. 
The first step for constructing a 
chemically interpreted model of semantics, or 
"Semistry", so to speak, is to study the 
bondage among atoms and molecules. For this, 
it is necessary to develop a theory of valence. 
Valence is defined as the capacity of an atom 
to enter into chemical (or semantic) 
combination with other atoms. It is possible 
to assign a value to the valence displayed by 
an atom in particular compound. This notion 
must be the reader who is well-informed of the 
European tradition of "Valenzgrammatik". 
Here, however, we will develop a theory of 
valence totally independently of European 
tradition. 
Before going into a detail of Semistry, 
let us show you a concrete exsample which is 
selected from Schank (1973) of his Coneepual 
Dependency theory (CDT for short). Here, 
"Semantic primitives of CDT" are compared to 
chemical elements. In the chemical elements, 
there are three types of chemical bondages; i. 
Single bond 2. Double bond 3. Tripple bond. 
If we look at CDT representation of a sentence 
through the viewpoint of Semistry, we will 
recognize a Similarity between chemical 
molecules and CD structure. From this 
insight, we can make a analogy of semantic 
'isomer', depending on a mode of bondage 
between the semantic primitives. For 
exsample, CD structure of PP(picture producer) 
and ACT (action) is represented in the Fig.2 in 
which two way dependency is interpreted as a 
double bond in case of Semistry. If one of 
the valence shifts to another pair of primitive 
as shown in Fig. 2, then the structure is called 
semantic resonance. 
In case of PP with the extra valence, some 
modifiers will be possible to link the 
activated part of PP. If the activation will 
occur at the ACT, the extra valence part will 
be embedded with the related case in CDT. 
Since it is not the purpose of the present 
paper to develop an impecable account of 
Semistry, let us take another example from a 
lexical item which is related to the present 
study. 
In the analysis of lexical structure, 
words are not really defined in the standard 
dictionaries in any precise way in case of the 
human cognitive system. There are various 
means to be employed to indicate their meaning 
more or less vaguely, but these means are 
usually sufficient for the cognitive 
processing. They may be extralinguistic means 
(such as diagram) or linguistic definition, 
both explicit and implicit. Neither of these 
--137 
modiiier case slots 
< > 
t t 
4--.PP === At'r--a< > 
"~--I'P --- ACT"~ < > 'i 
f 1 
4---PI' --- AC'I'--~ "" ~ 
pp --- ACT ~ < > 
t 
4-. p p --- ACT --~ 
i" 1 < > < > 
modifiers for I't' 
selaan tie resonance 
extra valetlce 
case slots ~or ACT 
Fig. 2 View o\[ semistry into Cu theroy 
tot Semantic resonance representatioB. 
are of much use for the construction of 
metaphor processing system. 
In case of metaphorical analysis, the 
lexical item must be defined with an inclusion 
of semantically unacceptable feature. This is 
represented by the following way as shown in 
Fig°3. 
The format of lexical item is adapted by the 
use of distributed semantic links (or Single 
bonds) between words. That is, a word or 
lexical item is surrounded with semantic 
features S I, S 2, ... S n. 
These bonds between word and semantic 
features are usually single bonds with 
homogeneous tention. In that case, the 
resonance is observed among the semantic 
features of word. In case of metaphorical 
semantic analysis, especially, noun-noun phase, 
Sl 
Sn I /$2 
$1 
Sn _ J .w $2 
S4 sQ,,,antic f~t.re, 54 
$I Sn-1 
5 4 52 
lelnantlc bond th¢oucJh S} 
.sn-1 
Sn~ ii/52 I Sn l,/ 
s/M " 2 ~: t ~SI 
S 4 $2 
Fig,3 I{epresentation of lexical item 
in semantic bond. 
the first noun modifies the second. So, the 
resonance is broken and the first noun in 
metaphorical relationship must include the 
meaning that is interpreted by the second. 
Therefore, in order to determine the meaning of 
a noun phase, there must be an intersection of 
meaning between M* and M. If such 
intersection exsists between the first and the 
second, the double bond is constructed in the 
Fig.3. In this way, word definition can be 
turned by adding procedure for unacceptable 
semantic link. The process is regarded as 
semantic change of meaning from Semistry's 
viewpoint. The change of meaning in metaphor 
is classified by the following categorical 
transformation as is shown in Table i. 
Thus, the idea of Semistry is proposed so 
as to meet the present purpose of metaphor 
semantic analysis. The experimental system 
188 - 
i. From Object to Human 
Contextual transformation 
M* = Objec t M = Human 
2. Bond between Objectand Human 
Link transformation 
M* = Object M=Human 
3. Transformation from Human body to Object, 
Location 
M* = Human body M = Object & Location 
4. Animal & Location's Personification 
M* = Animal & Location 
M = Human 
5. Pseudo-personification 
M* = Object & Animal & Location & Abstract 
M = personification's Object 
6. From Abstract to Concrete Object 
M* = Abstract Object M = Concrete Object 
Table I. Metaphorical Transformation 
called META-SIM is designed and tried out 
through the use of ELINGOL developed by Tanaka 
et al (1978). 
3a Metaphor analysis of noun phras e 
In this section, we show the case study 
based on the idea shown before. At the first 
stage, we analyzed a noun-noun metahpor using 
ELINGOL couped with production system designed 
with a viewpoint of standard control structure. 
The present studies focus on a noun-phrase in 
metaphorical use in Japanese, such as 
I. Metaphor 
"noun + nouN" 
II. Simile 
i. M* no m* SIM M no m.(m* of M* SIM m of M) 
2. M* no m* SIM M(or m).(m* of M* SIM M(or m)) 
3. M*(or m*) SIM M no m.(M*(or m*) SIM m of M) 
4. M*(or m*) SIM M(or m) 
In the above notation, SIM represents a 
similarity between two nouns in Simile, and a 
noun denoted a small letter is a part of noun 
denoted a capital letter. 
In this case study, we use a ELINGOL 
(Extended Linguistic Oriented Language) for the 
parser (systactic analyzer). The ELINGOL is a 
contex free parser extended at ETL, and it has 
a semantic processing parts that the user can 
write any semantic processing program in terms 
of LISP. 
Dictionarx 
The description in dictionary used in this 
case study is as in Fig. 4 
Each dictionary item consists of four parts, 
the first is item of the word, the second is 
the syntactic category of the word, the third 
is the part used in case of some ambiguities, 
(MQMIJI NQU~ (HIL O) 
"(MQMIJI 
UMIT 
<SELF SHOKUBUTSU) 
(PRRT-OP MID 
(~EM-FERTURE 
<SIZE = ,,4,~) 
CPRRT-OF-FEATURE (H~ (SIZE ~ CHIISQI))))) ) 
(TE IIOUN (HIL O) 
" (TE 
UNIT 
(SELF NIL> 
(PRRT-OP NI NBEM) 
(SEM-FEBTURE 
(SIZE = ~) 
(P-PROPERTY = ~,e-~.) 
(METR-NOUtl (HS OF SH~U~UT~U)))) ) 
Fig.4 Dictionary 
the fourth is the part for 
representation which the word has. 
knowledge 
In this 
case study, the knowledge of each word is 
expressed in terms of SRL knowledge 
representation, as in Fig.4. 
Here, the framework of knowledge 
representation is constructed by a set of 
semantic feature and properties, such as "TE 
(hand)" and "MOMIJI (maple)" in Fig.4. In the 
above representation, there are some special 
slots or semantic feature • The SELF slot 
represents a semantic category of the noun for 
the top node in part-whole relation network. 
PART-OF slot represents a upper node of the 
noun in part-whole relation network. In 
SEM-FEATURE, PART-OF-FEATURE represents some 
special feature of the components of the noun, 
and MATA-NOUN represents a restriction of the 
compared parts for the modifier noun category. 
Grammar 
The description in grammar used in this 
case study is as shown in Fig. 5. 
(HPK (HDU. OD) (.IL 0) (LG)) 
(NPP (HPK NOUH> {NIL 0) (CONS <LG) (LIST (RG>))) 
(IND (JB RBJV) <NIL O) O) 
(HPl (MPP IMD) (NIL O) £hG)) 
{ttPH CI'tOUM IND) eMIL 0) (LG)) 
(MP (NDUH HDUH) (NIL O) (NP~_EM (LG) (RG))) 
(HP (NPN NPP) (MIL O) (MTSEM3 (h~) (RG))) 
(tIP IMPI HPP) (NIL O) (MTSEMI (LG) (RG))) 
(~P (MPN HOUN> (NIL O) (MTSEM4 ~'LG) (RG))) 
{MP (NPI NOUN) (NIL O) (MTSEM23 (LG> (RG))) 
(HPL (NPK HPI) (MIL 0) (CONS (LG) (RG))) 
(NP (NPL NOUM~ (NIL O) (MTSEMI (CDR (L6)) 
(COMS (CAP (LG)) (LIST (RG))))) 
(SENTENCE ~{P (~,IL 0~, {~G)) 
($EtITEHCE (SENTEI~CE END) (r11L 0) <LG)) 
0 
Fig. 5 Grammar 
Each grammar consists of four parts, the first 
and the second parts represent a contex free 
rule of A --- B (C), the third is used in case 
of some ambiguities, and in the fourth part, we 
describe any semantic processing procedures. 
In Fig. 5, the fourth part describe a LISP 
function for metaphorical semantic processing 
which is considered in the next section. 
Procedure fo_j Metaj0horical semantic processing 
First, a input string must be parsed 
through ELINGOL and produce a parsing tree 
which is one of the control structure for 
semantic analysis. 
In order to interprete a noun phrase, a 
meaning of a phrase is constructed by seeking 
the semantic relation between noun and noun in 
--139 
the noun phrase. So, at first, two nouns to 
be interfered must be chosen, the choice is 
desided in terms of a syntactic structure and 
semantic part-whole relation network, because, 
in Japanese, there are many paraphrase only one 
noun phrase that has same meaning. 
Then, a new semantic interpretation is 
obtained from a intersection which is 
accomplished through the search of the two noun 
definitions. When an intersection occurs, the 
system focuses the matched semantic features 
extracted in the search to construct an 
interpretation. Thus, the search process 
corresponds to the conflict resolution process 
to produce the "infered meaning". In this 
way, interpretation of metaphorical use is 
accomplished. 
Here, we show the detailed semantic 
procedure for each cases shown before. 
(I) noun-i + noun-2, Metaphor 
Top level function : NPSEM 
Procedore : 
By metaphorical interference between noun-I and 
noun-2, metaphorical semantics is obtained from 
a intersection of semantic features between two 
nouns. 
(II - i) M* no m* SIMM no m, Simile 
Top level function : MTSEMI 
Procedure : 
First, by comparing noun semantic between M* 
and m* to that of M and m, the system can 
decide the semantic of "M* no m*" and "M no m". 
Then metaphorical semantics is obtained by 
contrasting noun phrase semantic between the 
semantic of "M* no m*" and that of "M no m". 
(II - 2) M* no m* SIM M(or m), Simile 
Top level function : MTSEM23 
Procedure : 
First, by comparing two noun semantics between 
M* and m*, the system can decide the semantic 
of "M* no m*", then metaphorical semantic is 
obtained by contrasting the semantic between 
"M* no m*" and M(or m). 
(II - 3) M*(or m*) SIMM no m, Simile 
Top level function : MTSEM3 
Procedure : 
First, by comparing noun semantic between M and 
m, the system can decide the semantic of "M no 
m". In this type, noun phrase contrasting has 
three types. The first type is in case that 
m* of M* is omitted because of m*=m. In this 
case, by comparing noun semantic between M* and 
m* (=m), the system can decide the semantics of 
"M* no m*", and then, metaphorical semantics is 
obtained by contrasting noun phrase semantic 
between "M* no m*" and "M no m". The second 
type is in case that m* of M* is omitted but m* 
is restricted by META-NOUN description m*' of 
m. In this case, by comparing noun semantic 
between M* and m*', the system can decide the 
semantic of "M* no m*'", and then, metaphorical 
semantics is obtained by contrasting noun 
phrase semantic between "M* no m ~''' and "M no 
m". The third type is other cases. In this 
case, by comparing semantic between M*(or m*) 
and that of "M no m". 
(II - 4) M*(or m*) SIM M(or m), Simile 
Top level function : MTSEM4 
Procedure : 
In this type, semantic procedure is as same as 
type (II - 3) without comparing noun semantic 
between M and m. 
Results of case studies 
Results of some case studies are shown in 
Fig.6, Fig. 7, and Fig.8. 
\MOCH I HODS° 
~EM/EMCE 
! 
SEHTEMCE .............................. EMD 
! ! 
MP ! 
! ! 
H 0 Ui"t ................. I ~OUN ! 
! ! ! 
MDCtl I HODF~ 
*,-I, METFIPHORICFIL It TERF'EREMCE *.i.,, 
,:MOl-H I 
I M\[T 
(SELF J INK \[313pJJTSU) 
(PART-OF ~tlL) 
CSEM-FEOTI.IRE 
(P-PROPERTY = NF~MEPAKO ~ YAWF~RAI.'A) 
<COLOR = WHITE) 
<iJEIGHT = ~.~.-i-) 
) ) 
( HODF~ 
Ui'l I T 
(SELF NIL) 
':PORT-OF DO(JBU TSU) 
/ (SEN-FEATURE 
(COLOR = $,-i.) 
(P-PROPERTY = ~*,~) ) ) 
--- I'IFITCHED SEMFIi'ITIC FEFIFURE --- 
P-PPOPERTY =====> ~P-PR\[IPERTY = IHRMERAI~A ~ YFIWORAKA) 
--- MOTCHE\]\] SEMROTIC FEF~TUPE --- 
COLOR =====> (ISOLOR = WHITE) 
--- PESLILT OF METFAPHOR --- 
(HODFI 
UNIT 
,'::ELF NIL) 
,PART-OF DOI_IPBU'T SLI) 
'. SEM-FESTURE 
':P-PROPERTY = I'3MEF'OWA & YAWARAKA) 
<COLOR = WHITE) 
)> 
L'46 M It.L I t ECOI'I\[IS. 
Fig. 6 Metaphor processing for "MOCHIHADA" 
Result shown in Fig°6 is to deal with noun-noun 
metaphor "MOCHIHADA (a soft white skin)". The 
intersection occurs at the semantic feature's 
description, then the slot of "P-PROPERTY" is 
filled in the semantic feature of "MOCHI (rice 
cake)". 
Next case study shown in Fig. 7 is to deal 
with "MOMIJI NO YOUNA TE (hand like as maple)" 
known as "Simile". 
Tree structure represents a result of parsing, 
and it is one of the control structure of 
semantic processes. In this semantic process, 
first, noun "MOMIJI (maple)" and "HA (leaf)" 
are interfered for noun phrase and produce a 
new unit which means "MOMIJI NO HA (leaf of 
maple)". Then, new unit "HA (leaf)" and noun 
"TE (hand)" of someone are interfered for 
metaphorical use, and produce a final result. 
And, other case study "DAIKON NO YOUNA 
HANAKO NO ASHI (leg of HANAKO like as DAIKON)" 
and its paraphrase "HANAKO NO DAIKON NO YOUNA 
ASHI" are metaphorically analyzed into the same 
semantic structures shown in Fig.8 and Fig.9. 
In this case, the syntactic analysis is much 
complex as is compared to the noun-noun phrase, 
and semantic process is as like as before. 
140 
a/HELLO) 
BUN WO IRE TE KUDASBI 
\MOMIJIMOYOUMRTE. 
SENTENCE 
! 
SEHTEHCE- ! 
HP 
! 
MPM ................ HOUM 
! ! 
HOUM ........... INB ! 
! ! ! 
I JO .... ADJV ! 
! ! ! ! 
MOMIJI HD YOUMA TE 
*~-i, MOUM PHRASE INTERFERENCE 
<MOMIJI 
UNIT 
(SELF SHOKUBUTSU) 
(PART-OF HIE) 
(SEM-FERTURE 
(SIZE = o,~.,~ 
(PART-OF-FEATURE (HA (SIZE = CHIISRI))) 
)) 
(HA 
UNIT 
(SELF NIL) 
(PART-OF SHOKUBUT~U) 
(SEM-FERTURE 
(SIZE = .~.b.~,,~9 
(COLOR = *-~,.wO 
(KATRCAI = ~-q'~') 
)) 
--- MATCHED SEMANTIC FEATURE --- 
SIZE :====>(SIZE = CHIISRI) 
--- SEMANTICS OF MOMIJI NO 
(HA 
OMIT 
(SELF NIL) 
( PART-OF MONIJI ) 
(SEM-FEATURE 
(SIZE = CHIISAI) 
(COLOR = $.~I,> 
(KATACHI = ~.0 
>) 
$~ METAPHORICAL IMTERFEREMCE 
(HA 
UNIT 
(SELF MIL) 
( PART-OF MOMIJI ) 
(SEN-FEATURE 
(SIZE = CHIISRI) 
(COLOR = $'~"~'0 
(KRTFEHI = ~"'~O 
)) 
(TE 
UNIT 
(SELF NIL) 
(PART-OF MIMGEM) 
(SEM-FERTURE 
(SIZE = ~> 
(P-PROPERTY = o~,,,) 
(METR-AOUM (HA OF SH~UBUTSU>) 
)) 
--- MATCHED SEMAMTIC FEATURE --- 
SIZE =====>(SIZE = CHIISRI) 
--- RESULT OF METAPHOR -- 
(TE 
OMIT 
(SELF NIL) 
(PART-~F HINGED) 
(SEM-FEATURE 
(SIZE = CHIISRI) 
~P-PROPERTY = $-~,-o) 
(METPr-MOUM (HA OF SHOKUBUTSU)) 
)) 
971 MILLISECONDS. 
END 
! 
! 
! 
! 
I 
! 
! 
! 
,DR I KZ\]~IMOY~UMF~HFItIAF ortoA -H I. 
SEMTEACE 
f 
SELtIEttCE .............................. ELtI 
! I 
LIP ! 
I I 
ttPM ..................... MPP ! 
MOULt ......... IHD MPK ........ ttOUM ! 
! JO .... ADJV IIOUM .... JO ! ! 
! ! ! ! ! ! ! 
BRIKOM MO YOLIMA HAHAKO HO RSHI 
$$$ MOUH PHRASE \]HTERFEREMCE ~,~ 
(HANRKO 
UNIT 
(SELF MIMGEH) 
(PART-OF IIIL) 
(SEM-FEATUPE 
(SEI = ONMA) 
(M-PROPERTY = ~-,.~) 
)) 
(ASHI 
UNIT 
(SELF NIL) 
<PART-OF DOUBUTSU) 
(SEM-FERTLIRE 
(HUTOSA = $$~) 
iLtRGASR = 4,~.$) 
)) 
--- SEMAMTICS OF HRMAKO NO RSHI 
(ASHI 
UNIT 
(SELF MIL) 
( PART-OF HAMAKO ) 
(SEM-FEATURE 
<HUTOSA = 4,~0 
CMAGASA = ~) 
)) 
$~ METAPHORICAL INTERFERENCE 
(DAIKOM 
UNIT 
(SELF ~HOKUBUTSU) 
(PART-OF NIL) 
(SEM-FERTURE 
~HUTOSR = HUTOI) 
(COLOR = WHITE) 
)) 
(ASH! 
UMIT 
(SELF NIL) 
(PART-OF DOUBUTSU) 
(SEM-FERTURE 
(HIJTOSA = $'$'~; 
(HAGASA = $~,~) 
)) 
--- MATCHED SEMAMTIC FEATURE --- 
HUTOSA =====b(HUTOSA = HUTOI) 
--- RESULT OF METAPHOR --- 
(( ~GEM0000 . HRMAKO > 
UNIT 
(SELF AIMGEM) 
(PART-OF NIL) 
(SEM-FEATURE 
(gEl = OMMR> 
(M-PROPERTY = ~,~-'P) 
)) 
(ASHI 
UNIT 
(SELF NIL) 
( PART-OF I,~SEMO000 ) 
(GEM-FEATURE 
(HUTOSA = HUTOI) 
(HRGASA = ~.~-G,) 
)) 
I283 MILLISECONDS. 
Fig. 7 Metaphor processing for 
"MOMIJI NO YOUNA TE" Fig.8 Metaphor processing for 
"DAIKON NO YOUNA HANAKO NO ASHI" 
141- 
\AAMAKDMODAIKONNOYOUMAASHI. 4. Summary and conclusions. 
SENTENCE 
! 
~:ENTENCE .............................. E,~, We have dealt with conflict resolution 
l 
tip ! 
NP I ..................... liOUN 
! ! 
NPP ............... I ND ! 
I ! ! 
HPK ....... MDUH I ! 
I ! I I 
NOUN .... JD I JO .... ADJV I 
! I ! ! ! I 
HRNAKD NO DBIKDN NO YOUNA ASH! 
e~,,4. NOUN PHRASE INTERFERENCE 
(HANAKO 
UNIT 
<SELF NINGEM) 
CPART-DF NIL> 
<~EM-FERTURE 
(SEI = ONHA) 
(M-PROPERTY = ~.~-) 
>> 
<ASH! 
UNIT 
<SELF NIL) 
(PART-OF DOUBUTSU) 
(gEM-FEATURE 
(HUTDSB = o.,,,.,~,O 
<NRGASA = N.> 
)) 
--- SEMANTICS OF HRMAKD NO R:SHI 
<BSHI 
UNIT 
<SELF NIL> 
< PART-OF HAMRKO ) 
(SEM-FEBTURE 
<HUTOSB ~ --,,,~) 
(MRGASA ~ ~) 
>) 
METAPHORICAL INTERFERENCE 
(DBIKDN 
UNIT 
<SELF SHDKUBUTSU) 
<PART-OF NIL> 
<SEN-FEATURE 
(HUTOSB = HUTDI> 
<COLOR = WHITE) )) 
<RSHI 
UNIT 
<SELF NIL) 
<PART-OF DDUBUTSU) 
(~EM-FEBTURE 
(HUTOSR = ~'~0 
>> 
--- MATCHED SEMANTIC FERTURE --- 
HUTOSR =====><HUTO3A = HLITQ!) 
--- RESULT OF METRPHDR ---- 
(4 ,I,GENO00I . HSMSkO ) 
UNIT 
<SELF NIMGEH> 
<PART-OF NIL) 
<3EM-FEATURE 
<SET = DAMS) 
(M-PROPERTY = 4w,,4-) 
>) 
(RSHI 
UNIT 
(SELF NIL) 
( PART-OF ~wE~EMO001 ) 
<~EM-FEATURE 
(HUTDSR = HUTOI) 
(MRGRSA = ,D.~.,,,) 
)) 
742 MILLISECONDS. 
Fig.9 Metaphor processing for 
"HANAKO NO DAIKON NO YOUNA AHI" 
process in metaphorical interpretation for noun 
phrases. In order to make the discussion more 
explicit, we have reviewed the problem on 
i conflict resolution both from cognitive 
psychology and artificial intelligence. 
Especially, we have made our attention to the 
problem of knowledge representation in human 
long-term memory and AI system. In this 
connection, the procedure for dealing with 
semantically unacceptable knowledge is stressed 
for the understanding of metaphor. That is, 
we have considered the dynamic aspect of 
meaning for word or lexical item in metaphor. 
In order to penetrate the problem on 
representation of meaning in metaphor, the idea 
of "Semistry" is introduced so as to analyze 
the conflict resolution in semantic 
interpretation. The idea of Semistry has been 
derived from the notion of Zwicky's paper on 
"Linguistics as Chemistry" which is 
metaphorical interpretation on Chemistry. By 
applying the notion into semantic structure of 
lexical item, the dynamic aspect of meaning is 
explaind by introducing the idea of "semantic 
bonds" which have further constructed semantic 
resonance among semantic features. A usual 
meaning is determined from the single bonds 
between word and semantic features. In order 
to determine the meaning of noun-noun phrase 
metaphor, there must be an intersection of 
meaning between the first and the second nouns. 
This kind of intersection is accomplished 
through the procedure for finding the matched 
semantic properties of the first and the second 
nouns. 
The proposed semantic model is designed 
and tried out for dealing with the noun phrase 
metaphor through the use of ELINGOL. Here, by 
parsing tree and LISP function in grammar, the 
inference system to resolve the conflict of 
semantic interpretation of metaphor was 
constructed. That is, the metaphor processing 
system would comprise a lexical item or word 
and associated inference mechanism to extract 
the meaning of metaphor. In order to proof 
the idea, the working system for the noun 
phrase is implemented by means of UCI-LISP 
(DEC-20) or HLISP (HITAC 8800-8700) and tried 
out with case studies. 
As the conclusions, we have shown a 
possibility for approaching a semantic analysis 
of metaphor from an actual working system. 
These cases are selected from the book called 
"A Stylistic Study of the Figuratives" 
(Nakamura, 1977). At the present state, 
number of items in dictionary is about 150. 
First, a new semantic model is proposed 
for dealing with metaphor. The idea of 
semantic resonance is introduced to explain 
"semantic bonds" which is derived from the 
comparison with Linguistic as Chemistry. 
Therefore, a role of metaphor is demonstrated 
in the present semantic model. 
Second, the ELINGOL is utilized to unify 
the syntactic processing with the associated 
inference mechanism to extract metaphorical 
interpretation. 
142 
Third, metaphor processing system called 
META-SIM is designed and tried out through case 
studies. We have discussed our experiences 
which was based upon the results of working 
system for metaphor processing system. 
Although the present system and case 
studies are restricted to the noun-noun phrase, 
the meaning of smaller phrase can be useful to 
build up semantic analysis of larger phrase of 
metaphor. In this sense, the present study is 
the first step toward the semantic analysis of 
metaphor which has not been explored in the 
natural understanding system. And further, 
the study of metaphor will give us about much 
more fruitful inference mechanism for 
interpreting semantically unacceptable 
sentence. In the future, the role of metaphor 
must investigate for both educational purpose 
and design philosophy of any understanding 
system. 
5_uAqknowledgements 
We indebted to many persons for 
suggestions and encouragement, especially to 
Dr. K.Fuchi and H. Tanaka of ETL Japan, 
T. Ishiwata of Ibaragi University, N.Terazu of 
Toyama University. We would like to special 
thanks to late Prof. S.I.Harada of Tokyo 
Metropolitan University for encouragement and 
effects for organizing research on discourse 
analysis for grant in aid for scientific 
research (Project number 211417, Agency number 
310702). The special care was offered to the 
authors from Fujimic Computer Center 
Corporation (DEC-20). 
8. Ortony,A., Reynolds,R.E. & Arter, J.A., 
Metaphor:Theoretical and Empirical research, 
Psychological Bulletin, 1978, Voi.85, 
pp.919-943. 
9. Rumelhart,D.E. & Ortony,A., The 
representation of knowledge in memory., In 
Anderson, Spiro & Montague (Eds.) Schooling and 
the acquisition of knowledge, Hillsdale,N.J., 
Lawrence Erlbaum Associates, 1977. 
i0. Schank, R.C., Identification of 
conceptualization underlying natural language, 
In Schank &Colby (Eds.), Computer models of 
thought and language. Freeman, 1973. 
ii. Tanaka,H. A Semantic Processing System for 
Natural Languedge Understanding, Research of 
the Electrotechnical Laboratry, No.797,July 
1979. 

References

i. Abelson,R.P., 1969, Psychological 
implication. In Abelson et al. (Eds.) Theories 
of cognitive consistency, Rand-McNally, New 
York. 

2. Bobrow,D.G. & Norman,D.A., 1975, Some 
principles of memory schemata, In Bobrow & 
Collins (Eds.), Representation and 
understanding, New York: Academic Press. 

3. Carbonell,J.R., Mixed-initiative 
Man-computer Dialogues, BBN Rep. No 1970, 
Cambr. Mass.,1970. 

4. Festinger,L., 1975, A theory of cognitive 
dissonance, Stanford, Ca.: Stanford Univ. 
Press. 

5. Kintsch,W., 1972, Note on the structure of 
semantic memory, In Tulving & Donaldson (Eds.), 
Organization of Memory, New York: Academic 
Press. 

6. McDermott,D., Assimilation of new 
information by natural language understanding 
system, Cambr. Mass., MIT AI Laboratory Tech. 
Rep. 291, 1974. 

7. Harada,S.I. & Mizoguchi,F., An 
introduction to linguistic chemistry, 
Unpublished maniscript 1977. 
