第２次短縮版

Ｔｈｅ　Ｃｏｎｓｔｒｕｃｔｉｖｅ　Ｐｒｏｃｅｓｓ　Ｔｈｅｏｒｙ　ａｎｄ

ｔｈｅ　Ｍｕｌｔｉ－Ｌｅｖｅｌ　Ｔｒａｎｓｌａｔｉｏｎ　Ｍｅｔｈｏｄ

　　　　　 ＮＴＴ　Ｉｎｆｏｒｍａｔｉｏｎ　Ｎｅｔｗｏｒｋ

　　　　　 Ｓｙｓｔｅｍｓ　Ｌａｂｏｒａｔｏｒｉｅｓ

　　　　　 Satoru IKEHARA＊, Satoshi SHIRAI, Kentarou OGURA,

　　　　　　　　　 Akio YOKOO and Hiromi NAKAIWA

　　　 * NTT Information Network Systems Laboratories　　　　　　　　　　　　　　1-2356 Take Yokosuka-shi Kanagawa-ken　238-03, Japan

　　　　 Tel:81-468-59-2515　　　Fax:81-468-59-3428

========================================================================

　 Computational linguistics has developed as a academic study, but a

means of dealing with existing linguistic processing, it is all but

useless and there are no examples of its having achieved appreciative

results.　The basic reason for this is believed to be that computationallinguistics has historically segregated its contents from language and

has turned it into an abstract form as a problem of formality and theory.

　 In the ALT-J/E Japanese into English Machine Translation System, a

critical posture is assumed toward the position of computational

linguitics. The system has suggested and realized a Multi-Level

Translation Method grom the view points of handling realistic language. This Method has been devised with the background of the Constructive

Process Theory which is one of the traditional Japanese grammar known

as the Tokieda Grammar.　It features handling of subjective expression

which as aware of the speaker and thinks through the meaning of

expression structures.　Technically, this method is supported by

meaning analysis which is based on linguistic knowledge.　As linguistic

knowledge for meaning analysis, a word meaning dictionary consisting of

400,000 words and a semantic structure dictionary including 15,000

sentence type have been prepared.　These have been written under a

system of about 100 types (in furtherdetail, these would amount 500

types) of syntactic attributes and about 3,000 types of semantic

attributes,　If rules were to be written by the same atrribute system,

it would consist of such a structure that this knowledge would be

========================================================================

　 Computational linguistics underwent development and progress since

generative grammar was propounded by Chomsky [Chomsky 1956, 1965].　

Examples of this having achieved results in natural language processing

[Tomita 1987] were rare and the gapbetween reality remains large.　A

great deal of thought has been directed toward identifying the causes

for this.　In most cases, the impasse confronted by theory and

limitations are found in the initial presises of such theories.

　 Primarily computational linguistics expresses language by a

theoretical model that has conformity and has sought to process this by

a unified method.　Yet, in reality, lamnguage is certainly not anything

that abounds in such conformity. Natural language is something like a

custom or habit that has spomtaneously grown and developed within its

own group of languages.　It reflects the anything but an object that is

perfectly and beautiflly adusted. It is a means of expression the

various distortions and conflicts in society.　With the changes in

society, language itsself undergoes changes.　Thus, to become

completely committed to language would make it impossible to pursuue

and unchangeable model.　In contrast with natural language processing

attempts to grasp language as a method of reality, computational

linguistics would appear to separate itself from reality and become

　 For example, deep structure is suggested as a structural element of

linguistic model incomputational linguistics.　But this is nothing

other than a sel-contradictory argument [Cresswell 1973, Mendeleson

1979 and Bresnan 1982] that ignores the relation between the target

objectives and the speaker in the language of reality.　There is the

strange phenomenon of hypothesizing a meaning to enable explanation of

a meaning.　Further, by manual translation, as in the saying,

"traslation begins with conversion of idioms to idioms" and importance

is attached to handling of idiomatic ecpressions.　In contrast, in

computational linguistics, idioms tend to be rejected as being beyond

the scope of principles of compositional semantics.　It would be

unnecessary to fall back on to the example of idiomatic expression, the

number of persons who believe that compositional linguistics is valid

in compositional semantics is limmited in number.　Yet computational

linguistics continued to stand by steadfastly by this assumption.

　If languages are a social custom, there should be a method of

processing that matches them.　Researches of natural language

processing require the pursuit of every one of these words.

　Accordingly, we have sought a new approach amidst the results of

conventinonal linguistics, promoting a new methodology.　The background

for our research work is the Constructive Process Theory of language

[Tokieda 1941] by Motoki Tokieda, who followed up on the work initiated

in the 18th century by Norinaga Motoori[Motoori 1779].　Our machine

translation research efforts have been based on this theory and has

resulted in a Multi-Level Translation method[Ikehara et. al. 1987,

1989, 1989] and a test product Japanese to English experimental machine

　 In this paper, thinking of the conventional computational lingustics

will be regarded in a critical manner and in its place, the

significance of the multi-level translation method based on the view

point of the processinf along this formula will be discussed.

　（１）Background of Computational Linguistics

　 There has been two types of logics, dialectic and formal logic since

the era of ancient Greece.　Dialectics which was advanced by Hegel

[Muramatsu 1959] asserts that actual existence should be thought of as

a meaning theory.　Hence, it deals with contents.　On the other hand,

formal logic[Quine 1950] insisted on purity and handling of an ideal

world, and cut itself off from the realistic world.

　 Symbolic logics[Hilbert et. al. 1928], computational logics[Allwood

1971] and computational linguistics are also derived from the latter,

formal logics.　There appears an effort to apply computational

linguistics to natural language processing.　But natural language is a

language that deals with the realistic world.　This would seem to be

the cause of various problems. Especially, Frege's principle [Nomoto

1986], which is compositional semantics and constitutes a major

assumption in computational linguistic, is not valid in a natural

　（２）Differences of View-points between Japanese and English

　 Let us consider the differences between the Japanese and Ind-

European languages.　The Ind-European languages have sentence patterns

such as 5 patterns which are relatively independent from contents.　

These languages can be classified as a suitcase type lamguage.　

Structualism was developed from the viewpoint of importance of forms

and structures of expressions.　Chomsky took notice of the importance

of contents to overcome the limitation of structualism.　But he still

　 In contrast, Japanese language is relatively free from forms and

does not have a rigid sentence structure.　Japanese can be classified

as a 「furoshiki」(wrapping cloth) type language.　Forms are dependent

on contents.　In other words, the shape of contents appears in a form

of expression.　Thus, contents are important in Japanese processing.

　 There are 20 different types of grammars in Japanese.　And four of

them are well known.　Hashimoto Grammar and Tokieda Grammar form a

contrast.　Hashimoto grammar[Hashimoto 1946] which was adopted for use

in school by the Ministry of Education of Japan is strongly influenced

by European philology.　Tokieda grammar[Tokieda 1941] which was derived

from Norinaga Motoori[Motoori 1779] is the most traditional one.

　 As nature is a complexed body of processes that develops with

conflicts as the motivating force, so also is language.　Tokieda

established the Constructive Process Theory of language which placed

importance on the contents based on this idea.

　２．２　Outline of Constructive Process Theory of Language

　（１）Comparison to Generative Grammar

　 Fig.1 shows the differences between the Constructive Process Theory

　 Constructive Process Theory asserts that a language should be

considered as a complex body of processes comprised of objects,

speaker recognition and expressions.　Objects reflect recognition.　

This process is explained by the reflection theory.　Recognitions are

combined into expressions using norms, namely the rule of languages.　

Linguistic norms are national rules which have been fostered in each

community.　These norms are referred to as a grammar in the broad

　　 ａ） Constructive Process Theory ( M.Tokieda )

　　　　　　　ｂ） Generative Grammar ( N.Chomsky )　　　　　　　　　　

　　　　　 Fig.1 Comparison between Generative Grammar 　　　　　　　　

　　　　　　　　 and Constructive Process Theory　　　　　　　　　　　　

　 In contrast, Chomsky interpreted language in the form of a dual

structure, a surface structure and a deep structure.　The existence of

deep structure has been assumed as the meanings of surface structure.　

This is certainly a strange theory where because meanings are assumed

for the purpose of explaining meanings.　More recently, he appears to

be attempting explanation based on mystic concepts.　He says language

cannot be explained without assmuming that humans are born with a basic

　（２）Constructive Process of a Language

　 Let us now see what happens when we regard language as a causal

sequence between object, recognition and expression.　The world of

objects consists of a speaker and other objects.　Objects are composed

from substances, attributes and relations.　The speaker recognizes the

object world and at the same time, recognizes things about himself.　In

recognition of objects, conceptualization takes place.　

[Object World]　　　　　　　　　　　　　　　　　　　　　　　　　　　　

　　　　　　　 Fig.2 Relation between Recognition and Expression

　　　　　　　　　　　　　　　　　　　　 (Subjective)　　　　　　

　　　　　Fig.3　Processing of Subjective Expressions 　　　　　　

　　　　　　　　 and Objective Expressions　　　　　　　　　　　　

　 There are two occasions regarding recognition of self.　One is the

occasion where speaker represents himself directly in the expression

without conceptualizing himself.　On other ocassions, he conceptualizes himself and represent it as an object.　When he conceptualizes himself, the phenomena of self disunion take place.　A separate imaginary self

other than the actual self appears.　And the self conceputualizes the

　 Recognition related to subjects and objects is represented by two

types of forms in linguistic expression as shown in Fig.3.　One is a

subjective expression and the other is an objective expression.　

Subjective expressions represent speaker emotions and intentions.　

These expressions are expressed in the Japanese language by post

positionals and adverbs. Objective expressions represent conceptualized

objects.　And these expressions are generally expressed by nouns and

　 This relation is somewhat different in English.　Inflections usually express subjective expressions in English.

　 A similar view is witnessed in the grammar of Port Royal[Lancelot et.al. 1660] in France 300 years ago.　The difference is an important

　３．１　Relational Meaning Theory

　 Let us consider the structure of languages.　If we assume that

meaning is a substance, it is one of either "object", "recognition" or

"expression" and the fourth element of "interpretation".

　 If we say that meaning is born out of forms, we would arrive at a

"formal semantic theory".　But this would be ideological.　Next, if we

say that forms are born out of contents, this would result in a "object semantic theory" or "recognition semantic theory".

　 If we assume that objects and recognitions are meanings, how should

we explain the sentence that has been erroneously written.　This would

result in a sentence that is incorrect in which something that is

opposite is written, yet stating that the meaning is correct.　Further, the object and recognition can never remain the same indefinitely and

will undergo changes.　Then, the meaning of the sentence will change

independently from expressions.　Motoki Tokieda chose eclecticism and

　 Tsutomu Miura[Miura 1967] took after the Tokieda grammar, but

improved the theory and advocated the "relation meaning theory".　He

proposed that structural processes from object to recognition are

combined to make an expression.　He explained the meanings of

linguistic expression as the relationship between object, recogntion

and expression.　This kind of relationship is concrete and specific.　

As long as the expression exists, so will the meanings.　When

expression is cancelled out, so will all relationships, and the

　 The concept of regarding relationships as meaning resembles recent

situation semantics[Barwise et al. 1981], but actually, the two are

entirely different.　In situation semantics, the meanings of

expressions and the meanings related to the location in which the

expression is placed are completely mixed up.　The Miura grammar

features a clear distinction between linguistic expression and

　３．２　Meaning Analysis and Meaning Understanding

　 Meaning understanding would involve re-tracing such a relationship

between object, recognition and expression and to relive and experience the status of object and recognition.

　 Here, let us consider meaning processing in two step.　The first

step would be the process which designates the rules or conventions

used in linguistic expression.　Conventions regarding languages are

complex and liable to be construed in many ways.　The listener must

identify the convention used by the speaker.　Let us define this act

of identifying the convention as "meaning analysis".

　 The second step would be the process of identifying the recognition

of the speaker and the status of the object that is tied in with the

expression.　Here, this action is refered as "meaning understanding".　

Thus meaning processing is comprised of two processes, "meaning

　 The knowledge that is required for meaning analysis would be

linguistic knowledge, that is, conventions or norms regarding

languages.　In contrast, meaning understanding would require knowing

the status of objects and this would require general knowledge, that

is, worldwide knowledge and knowledge in specialized fields.

　 Let us consider about the conventions involved in words of the

sentence (a).　The meaning of the word "高い(takai)" is "expensive",

"high", "noble", "loud among others", but here, it is used with the

meaning "expensive".　"Abura" has various conventions as "fat", "oil",

"lard", but here the convention expressing the meaning of "oil" is

used.　Thus, meaning analysis is identification of the convention which

the speaker actually used among numerous conventions that exist in the

　　 sono misewa takai aburawo utte　　iru　　

　(a) その店は　高い　　油を　売って　いる。　　

　　that store expensive　oil　 sell　　　　　　　　　

　 Meaning analysis cannot be achieved by merely staring at the words

one by one.　Look at the sentence (b).　Literal trnslation of

"senotakai" is "back is high".　In English, one word "tall" is used to

convey this meaning.　But looking at "back" and "high" separately will

not bring the concept of "tall".　The same can be said of the other

expression.　"Sell" and "oil" jointly means "idle away his time" in

this case.　Thus, conventions regarding words can never be decisive if

　　 seno takai otokoga　sigotochuu　　 aburawo　utte　 iru　　

　(b)背の　高い　男が　　仕事中　　　　油を　　売って　いる。

　　　 The tall man　 during work hour　is idling　

　 Structrures of objects reflect themselves into speaker recognition.

These structures are also combined to expressions.　This means that the

structure of an expressinon is a part of the meaning.　Structures

cannot be separated from meanings because structures also represent

meanings.　Thus, in meaning analysis also, there is the need to think

the relationship between structure and meanings.　More specifically, it

is important to grasp the structure of expressions as units of meaning.

　 Next, let us consider about meaning understanding.　As discussed in

the foregoing, meaning understanding involves the identification of the speaker's recognition and the status of object as tied in with the

　 To know the speaker's recognition and the status of object, the

listener must be equipped with a certain world in his mind.　This

world must have certain elements in common with the world which the

speaker is depicting.　To restructure a world corresponding to the

expressions of the speaker, he will be required to link up the elements of the speaker's expressions and the elements of his own world.　In

other words, meaning understanding can be considered as linking the

elements of the linguistic expressions and within the world model of

the listener.　This action of linkage will structure a new portion in

　３．３　Meaning processing and Knowledge

　 In contrast that liguistic knowledge is required for meaning

analysis, meaning understanding will require worldwide knowledge.　

Compared to linguistic knowledge, worldwide knowledge is massive and

considerable amount of difficulties will be involved in research.

　　　　　　　Fig.4 Meaning Processing and Application　　　　　　　　　

　 In machine translation, the contents of source text to be translated are diversified and it would be difficult indeed to prepare a worldwide knowledge.　Machine translation can be acceptable even if the computer

cannot entirely understand the contents of the text as long as the

translation results can be finally understood by humans.　Replacing the

linguistic conventions used in the Japanese language by English

conventions and leaving it up to the judgment of the average person.　

This is the basic stance assumed by our research efforts in machine

　 First, we want to realize a translation based on meaning analysis.　

Next, we will extract the expressions and concepts which could not be

translated by meaning analysis and attempt a meaning understanding with a limited world model.

　 It is advisable to deal with meaning processing with areas such as

telephone number information and database inquiries, because the target

world can be contracted into a comparatively small range.　

　４．１　Problems of Intermediate Language Method

　 Conventional translation methods[MT Summit-I,II,Ⅲ 1987,1989,1991]

can be classified into two types, the pivot method and the transfer

method.　There are many which seek to achieve the pivot method.　

However, there have been none that can be stated to have truly realized

the pivot method.　This is because universal intermediate language that

can be commonly applied to many language cannot be designed.

　 Natural language has primarily originated by reflecting the

perspectives and thinking of the group of people using such language.　

It would appear unreasonable to consider a intermediate language that

would be commonly applicable to every language which differs in

　 When a translation between languages which are very close each other is considered, or if a rough ranslation is to suffice, the pivot method may be satisfactory.　But in either case, pivot method seems to have a

basic misunderstanding of languages in its background.

　 The transfer method can be stated as being more realistic compared

to the pivot method, because it does not assume that intermediate

language is universal.　However, if intermediate language can be

thought of as a meaning expression such as deep structure, the same

difficulties as in the case of the pivot method would arise.　The

meanings of surface structure cannot necessarily be retained by deep

　４．２　Basic Concepts of Multi-Level Translarion Method

　 The method of translation based on the foregoing meaning analysis is the Multi-Level Translation method.　The term "Multi-Level" indicates

that translation is being conducted at various levels of abstraction.

　 The need to observe differences between subjective expression and

objective expression have previously been mensioned.　And, the

necessity of consideration about the meaning of structures have also

been pointed out.　The Multi-Level Translation　method has been

　 The first point is considered as follows.　Japanese and English are

very different in presenting subjective expressions.　Therefore,

subjective expressions are segregated from objective expressions and

　 The second point is resolved by abstraction of expression structures.Three levels of abstraction have been proposed.　The first is related

to a level of idiomatic expressions.　The second concerns specific

structures such as loosely coupled words.　The last concerns most

loosely coupled structure that can be represented by general rules.

　　　　　　Fig.5　Multi-Level Translation Method　　　　　　　　　　　

Thus, this system is divided up into 2 major conversion paths.　The

one is the conversion of the speaker's sense or what is subjective.　

The other is the path for conversion of the description for the object. And this path splits up into 3 more paths, depending on the abstraction degree in conceptualizing objects.　Thus, the Multi-Level Translation

method has currently four translation paths.　

　４．３　Organization of Knowledge

　（１）Machine Translation and Knowledge

　 To realize a machine translation system, a knowledge of language is

essential.　Natural language processing can be stated as being a battle

with ambiguity from the beginning to end.　In the translation process,

various types of ambiguities arise, but this is regarded as being a

　 To overcome such ambiguity, it is necessary to study what knowledge

is lacking and to have such knowledge established as the rule or as a

dictionary.　It is important to consider the type of ambiguity and

relationship with the knowledge corresponding to it.　ALT-J/E system

placed importance on knowledge for meaning processing for the purpose

of solving ambiguity of sentence structure and of the meaning of

　 From the viewpoint of linguistic knowledge, linguistic conventions

are divided into those which are related to subjective expressions and

those related to objective expressions.　Of the two, the knowledge

related to subjective expressions are smaller in scale compared to

objective expressions and have been solved in terms of rules or

programming.　But the objective expressions being massive, we have

decided to compile them in the form of a dictionary.

　（２）Semantic Word Dictionary and Semantic Structure Dictionary

　 When the speaker expresses an object, abstraction is conducted and

the objectis conceptualized.　In language expressions, the concept of

object is expressed separately in terms of a concept of substance and a

concept of attribute.　Linguistic conventions which combine concepts to

expressions have therefore been compiled into 2 separate dictionaries,

a semantic word dictionary consisting of 400,000 words and a semantic

structure dictionary consising of 15,000 sentence forms.

　　　　　　　Fig.6　Knowledge for Meaning Analysis

　 The meanings of words and their translation are listed in ordinary

dictionaries.　But it does not serve to have the computer know how

they are to be used.　For example, the word "school" is used as a

"place" and also as an "organization".　As a means of desscribing these

knowledge, we have compiled the use of the meaning of words as a

　 Consider the relation between expression structures and meanings.　

The meanings of expressions are not necessarily represented by the sum

of the meaning of every word in the expression.　In fact, it would

appear that the meanings of linguistic expressions actually used cannot

be explained solely by the meaning of each word used in the expression.

Expression structures need to be considered as units of meanings in

natural language.　Based on such thinking, the abstraction of

expression structures centered around declinable words and compiled as

units of meanings has resulted in the semantic structure dictionary.　

　（２）Necessity of Detailed Meaning Attribute System

　 It is important to consider precision of knowledge description in

relation to the degree of ambiguity.　For example, an algorithm which

was valid in a world of 1,000 words may not necessarily be valid in a

world of 100,000 words.　Experiences indicate that increase of the

number of words causes rapid growth of ambiguities.　

　 Seeking to establish practical method, we have collected words

including proper nouns normally used day to day and have compiled a

dictionary of about 400,000 words.　Thus, it is necessary for semantic

attribute system to work with this environment.　We began with about

500 categories.　But this has resulted in that translation of verbs

cannot be precisely described.　With the consideration of our aim and

environment, we set up 2,800 categories for general noun semantic

attributes and 200 categories for proper noun semantic atrributes.

　 Here, let us look at the differences of description capabilities

between of our semantic attribute systems and other similar systems.

Case 1 is a typical example of Japanese to English machine translation

systems.　The number of semantic categories are between 30 to 50.　Case

2 is the current plan of EDR [EDR 1990].　Case 3 is an example of our

　　A comparison was made of the three cases.　And, the capabilities of

describing the translation rules for verbs are evaluated.　In the

comparison, the case of ALT-J/E has been assumed to be 100%.

　　　　 Table 1 Comparision of Semantic Attribute System　　　

　　　 *　Relative Amount of Writable Semantic Patterns　　　　

　　　　　(Valentz Pattern) for Japanese General Verbs　　　　

　　　 ** Japan Electronic Dictionary Reseach Institute, Ltd.　

　 The results of this comparison are as follows.　The capabilities of

description was 31% for case 1 and 59% for case 2.　Those results

indicate that when precision levels of semantic atrributes are low,

many essential rules cannot be written.　Our experiences also indicates

that, in Japanese to English machine translation, precision level for

some 3,000 types of semantic attributes are necessary for

differentiation of translation for verbs.　However, this is not

　 In machine translation research, some of the main problems include

the meaning (convention) in which each word is being used　(words that

are frequently used have many meanings), the distinction between

idiomatic and literal meanings, the sentence form in which sentences

with no subjects (or subject matter) and multiple sentences.　The

achievements of conventional computational linguistics is all but

helpless in dealing with realistic problems.　When natural language is

viewed as having developed as a natural phenomenon and as a social

custom, there is a different convention regarding the meaning and

method of use for each and every word.　The convention needs to be

discovered and compiled as a knowledge if we are to realize natural

language processing　This does not mean collection of knowledge to

match the algorithm to match knowledge [Nirenburg 1989].

　 Based such thinking, this paper points out the @roblems confronted

by conventional computational lingfustics, and discusses the

significance of a new machine translation system, the multi-level

　 As a concept, the multi-level translation method has been devised

with the background of the constructive process theory which is one of

the traditional Japanese grammar known as the Tokieda Grammar.　It

feature handling of subjective expression which is aware of the speaker

and thinks through the meaning of expression structures.　Technically,

this method is supported by meaning analysis which is based on

lingustic knowledge.　As lingustic knowledge for meaning analysis, a

word meaning dictionary consisting of 400,000 words and a semantic

structure dictionary including 15,000 sentence types have been

prepared.　These have been written under a system of about 100 types

(in further detail, these would amount to 500 types) of syntactic

attributes and about 3,000 types of semantic attributes.　If rules were

to be written by the same attribute system, it would consist of such a

structure that this knowledge would be mobilized and translation be

　The ALT-J/E has, by means of this system, accomplished new functions,

（１）Differentiation in the translation of the verb and noun

（２）Automatic rewiting function of original Japanese sentences

（３）Automatic supplementation of subjects and objects

These functions are mandatory in realizing a system of Japanese to

English machine translation that requires no pre-editing.

　 The main issues confronted by machine translation research are:

(1)Meaning Analysis of Complex Noun Phrases and Complex Compound Nouns,

(3)Global Structure Generation of English for long sentence translations.　 Our major research efforts are being concentrated in these areas.　

④Introduction of Common Sense about the World.　

　 Preparations for these studies are now under way.

The authors wish to thank Dr. M. Miyazaki, Professor of Niigata

University and other colleagues of machine translation for their

[Allwood 1971] J. Allwood, L.G. Anderson and O. Dahl:"Logics in

　 Linguistics, Cambridge Univ. Press, 1971

[Barwise et al. 1981] J. Barwise and J. Perry:"Situation and Attitudes",　 J. of Philosophy, Vol.78, pp.668-691

[Bresnan 1982] J. Bresnan ed.:"The mental Representation of Grammatical

　 Relations", Cambridge, Mass., The MIT Press, 1982

[Chomsky 1956] N. Chomsky:"Three Models for the Description of Language",　 IRE Trans, IT-2, 1956

[Chomsky 1965] N. Chomsky:"Aspects of the theory of Syntax", MIT Press,　　1956

[Cresswell 1973] M.J. Cresswell:"LOGICS AND LANGUAGES", Methuen Co. Ltd.,　 London, 1973

[EDR 1990] EDR:"Concept Dictionary", Technical Report-027, Japan

　 Electronic Dictionary Research Institute, Ltd., April, 1990

[Hashimoto 1946]S. Hashimoto:"Theory of Japanese Language: Kokugogaku

　 gairon(in Japanese)", Iwanami Bookstore, 1946

[Hilbert et. al. 1928] D.Hilbert and W. Akermann:"Grundzuge der

　 theoretischen Ligik" Translator M. Itou, Oosaka kyouiku Book, 1928

[Ikehara et al. 1987] S. Ikehara, M. Miyazaki, S. Shirai and Y. Hayashi:　 "Speaker's Recognition in Language and Multi-Level Translation

　 Method (in Japanese)", Transaction of Information Society of Japan,

[Ikehara 1989] S. Ikehara:"Multi-Level Machine Translation System",

　 Future Computer Systems, Vol.2, No.3, pp.1269-1279, 1989

[Ikehara et. al. 1989] S. Ikehara, M. Miyazaki, S. Shirai and A.

　 Yokoo:"An Approach to Machine Translation Method Based on

　 Constructive Process Theory", Review of ECL, Vol.37, No.1, pp.34-49,

[Lancelot et. al. 1660] C. Lancelot and A. Arnauld:"Grammaire g'en'erale　 et raisonn'ee, les fondements de l'art de parler", Chez Pierre le

[Mendelson 1979] E. Mendelson:"Introduction to Mathematical Logics",

[Miura 1967] T. Miura:"Theory of Recognition and Language (in Japanese)",　 Keisou Shobou, 1967

[Motoori 1779] N. Motoori:"kotoba no tama-no-o (in Japanese)", 1779 =

　 see S. Oono: "Complete Works Series of Motoori Norinaga (in Japanese),　 Chikuma Publishing, 1970

[MT Summit-Ⅰ 1987] Proc. of MT Summit-Ⅰ, Sept.16-18, 1988

[MT Summit-Ⅱ 1989] Proc. of MT Summit-Ⅱ, Aug.16-18, 1989　

[MT Summit-Ⅲ 1989] Proc. of MT Summit-Ⅲ, Jul.2-5, 1991　

[Muramatsu 1959] K. Muramatsu:"Logics of Hegel(in Japanese)", Keisou

[Nirenburg 1989] S. Nirenburg:"KBMT-89-A Knowledge-Based MT Project at

　 Carnegie Mellon University", Proc. of MT Summit-II, pp.141-147

[Nomoto 1986] K. Nomoto:"Frege's Philosophy of Languages(in Japanese)",　　Keiso publishing, 1986

[Quine 1950] W.V. Quine: "Methods of Logic"　Translator Nakamura and

[Tokieda 1941] M. Tokieda:"The Principle of Linguistics (in Japanese)",　　Iwanami Bookstore, 1941

[Tomita 1987] M. Tomita:"An Efficient Augmented-Context-Free Parsing

　 Algorithm", Computational Linguistics, Vol.13, No.1-2, pp.31-46

　 Nevertheless the progress of computstional linguistics for these

30 years, they are not useful yet for practical language processing

and there are no example where noticeable effects are obtained.　The

reasons are considered that they historically cut off the actual

contents from languages and have focused the attention on the

relation between form and rogic of expressions.

Alpac Report:"Language and Machines: Computers in Translation and

Linguistics", Automatic Language Processing Advisory Committee,

National Academy of sciences, U.S. National Research Council, 1966

C.J. Fillmore:"TOWARD A MODERN THEORY OF CASE & OTHER ARTICLES (in Japanese)", Sanseidou Publishing, 1975

M. Nagao:"Evaluation of the Quality of Machine-Translated Sentences and

the Control of Language", J. of Information Processing Society of

M. Nagao:"Japanese View of Future of Machine Translation", Proc. of MT

F.D. Saussure:"COURS DE LINGUISTIQUE GENERALE", Japanese Edition,

H. Tomabechi:"Direct Memory Translation", 'Proceedings of IJCAI-87'

	Meaning = Relationship

Object		Recognition	Expression

	Deep Structure				transfer		Surface Structure
		Meaning						Expression

(Objective)		Direction
	Substance
		Addition
	Dynamic　 Attribute	Desire
		Guess


	Objective Expression	①Idiomatic E. T.		Objective Expression
		②Semantic V. P. T.
Multi-Level　　 Transfer Method			①Idiomatic Expres 　-sion Transfer ②Semantic Valentz　　　　Pattern Transfer　　　 ③General Pattern 　　　　Transfer
		③General P. T.

		Semantic Word Dictionary
Semantic　Attribute System		Caption Words　４００,０００　　< Total Word　600,000>
･General Noun　　　　　２,８００ Categories ･Proper Noun　　　　　　　２００ Categories

		Semantic Structure Dictionary
１５,０００ Rules Prepared for Declinable Words

No	Ｃａｓｅ	No. of Categories	No. of words	Evaluation (Example*)
１	Current System (J. to E.)	３０－５０ Categories	50,000 　 Words	３１％
２	ＥＤＲ Project**	５００ Categories	400,000 　 Words	５９％
３	ＡＬＴ－Ｊ／Ｅ	３，０００ Categories	400,000 　 words	１００％ (Assumed)