An Approach to Machine Translation Method
based on Constructive Process Theory
SATORU IKEHARA
MASAHIRO MIYAZAKI
SATOSHI SHIRAI
AKIO YOKOO
ABSTRACT
The multi-level-translation method for
machine translation is proposed based onthe Constructive Process Theory. It iscomposed of two sub-methods: a separate-recombine method and a multi-level transfermethod, which are based on an analysisof the speaker's recognition of both thesubject and the object, and a simultaneousanalysis of the syntax and meaning.
The separate-recombine method analyzes
subjective expressions to extract emotionsand intentions, and then recombines themin the target language. The multi-leveltransfer method conveys the remainingobjective expressions in the target languageby three levels of transfer, based on thedegree of sentence structure abstractness.
Here, the meanings aspect of an
expression's structure is considered toachieve accuracy in machine translationcompared with the conventional one.
1. INTRODUCTION
Unlike natural science, which dealswith physical phenomena, research on natural
language deals with a mental product. Many
explanations for language have been proposedbased on different interpretations of themental function. For instance, Saussure's
structuralism (1)(2) assumes that the mentalfunction exists in an " a priori conceptualsubstance", and states that the content ofthe mental function is a linguistic norm. Neither Saussure's, structuralism nor
ordinary grammar (3), can explain homographicexpressions (why the same sentence structureoften has different meanings). Chomsky(4)(5)proposed that a more abstract structure mustbe considered to be the meaning of anexpression. He introduced a "deep structure",which assumes a common thought pattern forevery one. Chomsky's idea focuses on thecontent of an expression, but he separatedthe content from the object so that thereflection theory was ignored and a dualism(6)(9) was asserted, in which form opposedcontent. However, the state the objects,and how recognition actually occurs arereflected by the form of expression, so thatthe form and the content are interdependent.In addition, the speaker's recognition isconnected to an expression. Therefore, adeep structure need not be assumed as asemantic structure separated from a surfacestructure. On the contrary, the relationbetween the object, the speaker's recognitionof it and the expression should be consideredas the true meaning.
Most current research on machine
translation (8)(9) follows some form oftransformational generative grammar, inwhich a common structure of meaning isassumed and represented in an intermediatelanguage independent of ordinary languages.
On the contrary, if the "process constructionof a language" (10)(11) such asan object,recognition and an expression is considered,it becomes apparent that only the object iscommon to the languages and that the mannerof recognition differs from person to personas well as from language tolanguage.
Therefore, it is difficult to assume a deepstructure as having a meaning common todifferent languages. This difference inrecognition structure (12)(13) between theoriginal language and the target languageshould be considered in high quality machinetranslation.
2. PROCESS CONSTRUCTION OF A LANGUAGE
2-1 The Process Construction Theory
Language is metaphysically explained as
a complex body of the complete world byformalism and structuralism. An expressionis explained by the functions or forms ofeach part of its parts.
Chomsky's explanation pays sufficientattention to the content of expressions.However, the relation between the object andthe speaker's recognition is not considered.Prior to this, the Process ConstructionTheory proposed by Tokieda (10), claimed thatlanguage is composed of three processes:object, recognition and expression. Theseprocesses are combined by the law ofcausality. The state of an object reflectsthe speaker's recognition, therefore, the waythe speaker recognizes that object results inan expression. This mental link fromobject to recognition and finally toexpression illustrates the ProcessConstruction Theory. The differencesbetween Chomsky and Tokieda's theories areshown in Fig. 1.
2-2 Speaker Recognition and Representation
(1) Recognition of the Subject and the Object
The world recognized by a speaker iscomposed of a subject, namely the speakerhimself, and objects ( Fig. 2). A speakerrecognizes both his own state and the stateof other objects, and connects them toexpressions. Based on this, Tokiedaexplained that a Japanese sentence iscomposed of subjective expressions andobjective expressions. By his definition,subjective expressions directly representsubjective emotions and intentions. Japaneseadverbs and "joshi" (post-positional wordsfunctioning as auxiliaries to main words)are used for these expressions.
Objective expressions representconceptualized objects. Nouns, verbs andadjectives are used for these expressions.
When a speaker conceptualizes the subject(the speaker himself), the subject can berepresented by these objective expressions.
The relationship between subjective andobjective expressions has been pointed out byPort-Royal (15) for Indo-European languages.
(2) Connection between a Recognition and an Expression
An object can be broken down intosubstance, attribute and relation, all ofwhich have many structures. These structuresare reflected in expressions through thespeaker's recognition. The substance of anobject has a hierarchical relation between apart and the whole for example. Attributesare related to the substance. Relations aremainly constructed of three further kinds ofrelations between substances, attributes andrelations.
These components and partial structuresare combined in many ways to create atotal structure in the speaker's mind. Themanner of recognition depends on theviewpoint of the speaker. Every languagehas its own framework for representing suchrecognition. Thus, the original structure ofthe object is not reflected directly in theexpression, but through the speaker'srecognition.
The difference between languages is thedifference between the frameworks of thespeaker's recognition. The relationshipbetween subjective expressions and objectiveexpressions in the original language shouldbe reconstructed and expressed in thetarget language.
2-3 Definition of Meaning
(1) A New Definition
Many explanations of the meaning of anexpression have been proposed. Saussure (1)distinguished "lang" as having common socialsignificance and "parol" as having individualmeaning based on individual experiences.Taking a different point of view, Chomskymade a distinction between semantic deepstructure and syntactic structure. Althoughtheir perspectives are different, they agreethat content is independent of the object,and that "lang" and deep structure are bothcommon to human beings or specific humangroups. A lot of current machine translationresearch is based on these ideas and assumesthat deep structure is meaning. There arealso cases where deep structure is defined asthe state of an object, which differs fromthe speaker's recognition of it.
On the contrary, we define the meaningof an expression as the relation between theobject, the recognition and the expression.This definition is based on the fact that howthe subject and an object actually exist isconnected to an expression through thespeaker's recognition. Accordingly, themeaning (i.e. relationship) cannot existwithout the expression. The ordinary meaningof a word as defined in a dictionary is notactually a meaning, but rather a languagenorm. Only when a word is used in anexpression does the relation to a recognitionand an object arise giving the word a truemeaning.
(2) The Meaning of Syntactic Structure
A speaker consolidates his recognitionand represents it by an expression, usingrules to form words, phrases and clauses.This consolidation is based upon languagenorms supported by a meaning. That is, thestate of the object reflects a speaker'srecognition, and the speaker's recognition isreflected in an expression. This meansthat the syntactic structure is integral withthe object and the recognition, and that thesyntactic structure is part of the meaning.Unlike transformational generative grammar,where structure (a surface structure) is
distinguished from a meaning (a deepstructure), a surface structure can beconsidered part of the meaning. Therefore,transforming an expression, strictlyspeaking, changes the meaning. Transformationcannot leave the meaning unchanged.
Transformations in ordinary languageprocessing are transformations to alternateapproximate expressions.
Thus, the element composition method,
which tries to compose the whole meaningfrom its parts, neglects the meaning ofthe syntactic structure. The meaning of apart of an expression can be correctlyinterpreted only in the context of wholesentences. Accordingly, the rules of themeaning of a word used in the expression,can be determined only for that particularcontext.
3. MULTI-LEVEL MACHINE TRANSLATION METHOD
A competent sentence is defined as asentence that can be translated in isolationusing only linguistic knowledge, namely,grammar and a dictionary. The multi-levelmachine translation method (MLMT) is proposedfor competent sentences.
3-1 Separate and Recombine Method
Japanese is classified as anagglutinative language where "joshi" andadverbs are used for subjective expressions.By contrast, English is an inflectionallanguage, with subjective expressions usuallyrepresented by inflections. Thus, Japanesesubjective expressions do not directlycorrespond to English ones, and it isdifficult to translate word for word.
For this reason, the speaker's emotionsand intentions are first classified intocategories and then analyze what categoriesthe subjective part of a given Japanesesentence represents. Thus, the originalJapanese sentences are transformed into basicJapanese sentences. The Japanese sentencesremaining after the subjective expressionsare extracted are objective expressions. These expressions are translated intobasic English sentences by the multi-leveltransfer method described in the nextsection. Finally, the speaker's previouslyextracted emotions and intentions arerecombined with the basic English sentences.Adverbs and prepositions are added and nounsand verbs inflected. In this way, informationabout the subjective expressions separatedfrom the original Japanese sentences isrecombined in creating theEnglish sentences.
3-2 The Multi-Level Transfer Method
The state of the object is representedin a basic Japanese sentence (an objectiveexpression) from the speaker's viewpoint.
The speaker's recognition of an object hasseveral structures and these structures arereflected in the structure of the objectiveexpressions. If we strictly adhere to thethat changing an expression changes themeaning and that an integrated process ofsyntactic structure and meaning are neededfor accurate translation, then a matchingEnglish expression is needed for everypossible Japanese expression. Clearly, thisis not practical because of the infinitenumber of expressions. Thus, sentencestructures are classified into three levels,according to the strength of the link betweenthe sentence structure and the meaning. Sentence structures are transformed by amethod suitable to the level.
(1) Specific Recognition Structure Level Idiomatic expressions have meaningswhich cannot be determined from individualwords causing extreme difficulty intranslation by an element compossitionmethod. These kind of expressions therefore
are completely transferred as matched pairsof Japanese and English expressions byidiomatic expression transfer rules.
(2) Individual Recognition Structure Level
More general structures are classifiedinto this category. In specific recognitionstructures, the words are completely fixed. However in an individual recognition
structure, a single word is fixed, and theother words are restricted by their semanticattributes. When a declinable word isfixed, the contents of other "bunsetsu"(Japanese clauses) connected to thedeclinable word are restricted in the use of"joshi" and the semantic attributes of nouns. A case grammar(16) could be applied forthese structures. However, a Valentzpattern (17) transfer method is more suitableas it does not require use of a deepstructure, which is difficult to defineprecisely. In a case grammar, some of themeaning of a syntactic structure is missed inthe process of deep case selection, whereasthe meaning of a syntactic structure can betransferred intact using the Valentz patterntransfer method.
This paper uses the Valentz PatternTransfer method augmented by restrictions ofa word attributes. This method is supportedby a system of precise and mutuallyexclusive semantic attributes of words.It can transfer meanings which cannot becategorized by case grammar.
Individual recognition structures arepaired for corresponding Japanese andEnglish expressions registered in a pattern
dictionary similiarly to idiomatic transfer
rulls.
(3) General Recognition Structure Level
Compared with both special andindividual recognition structures, which arethought to be comprised by a pattern ofspecial words or special and associatedwords, the general recognition structuredeals with a more comprehensive patterns. Inthis level, a word is not fixed. For example,patterns may be classified by verb type: i.e.instantaneous or stative, etc. Generalpatterns corresponding to groups of verbs areprepared. With this method rough translationis inevitable due to generalization.
Specificity in a structure correlateswith the quality of translation. The rule forthe three methods is that the more specificthe structure, the higher the quality oftranslation that can be expected. Thesemethods are applied to basic Japanesesentences (subjective expressions) in theorder described above. If no patternsrelevant to a given Japanese sentence canbe found in the dictionary of idiomaticexpression or a semantic Valentz pattern,then a general pattern is used but loses thehight translation quality. With the expansionof pair pattern dictionary, the translationquality is expected to improve.
3-3 Structure of the MLMT Method
The MLMT method consists of twosub-methods: a separate and recombine methodfor subjective expressions, and a multi-leveltransfer method for objective expressions(Fig. 3).
This translation process is similar tomanual translation (Fig. 4). Here, a humantranslator first feels for himself, thespeaker's experience as described by a givensentence. This process is supported bythe Japanese norm that connects a speaker'srecognition to Japanese expressions. Thus,the translator understands the state of theobjects and the speaker's emotions andintentions towards the objects. In the MLMTmethod, the original Japanese sentence isseparated into descriptions of the state ofthe objects and the speaker's emotions. Thestate of the objects is represented bya basic Japanese sentence (an objectiveexpression) and the speaker's emotions arerearranged in a reference table.
In human translation, the state of theobjects is then reorganized in the frameworkof English, and the speaker's emotions arerecombined with it. Similarly, in theMLMT method, the meanings of objects aretransferred into English by the three levelsof the transfer method. The speaker'semotions, arranged in a reference table,are recombined to give the final Englishexpressions.
In this method, syntactic structure andmeaning are represented, by idiomaticpatterns or Valentz patterns. They can notonly be used for Japanese and Englishtranslation, but also for Japanese sentenceanalysis which resultsin fewer ambiguitiesthan the ordinarymethod. Moreover, transferrules are highly independent of each other;therefore, the consistency check is limitedto a smaller range facililating the expansionof the translation system.
4. Conclusion
Based on the constructiv process theory
of natural language, the multi-level machinetranslation method was proposed.
For machine translation, the importanceof separating recognitions concerning subjectand object, and retaining the meaningassociated with syntactic structure wasshown. The MLMT method consists of two sub-methods, which correspond to these twoideas: a separate and recombine method forsubjective expressions, and a multi-leveltransfer method for objective expressions.
Ideally, to handle a syntacticstructure and its meaning as one unit andthus to produce high quality translation,all possible expressions should beidentified and included in the transferrules. The open-ended characteristics ofnatural language make this technicallyimpractical. As a technical compromise,expression structures are classified intopatterns corresponding to abstractionlevels of speaker recognition, and subjectiveexpressions are separated from the originalsentences to improve the ratio of matchingpatterns.
The MLMT method was proposed fortranslating competent Japanese sentencesinto English. Proposed ideas aboutseparating subjective expression andobjective expressions, and the importance ofthe meaning of syntactic structure can beapplied commonly to natural languages, thenMLMT method will also operate with othernatural languages.
Acknowledgment
The authors wish to thank the othermembers of our research group for helpingto implement this method.
References:
(1) E. F. Koerner: Ferdinand de Saussure,
Braunschweig: Friedr. Vieweg+Shon GmbH, 1973 (Japanese edition by K. Yamanaka, Taishukan, 1982)
(2) G. C. Lepscky: A survey of structural linguistics, Faber & Faber, 1970(Japanese edition by S.Sugata, Taishukan, 1975)
(3) Iwanami: Japanese 6 (Grammar T), 7 (Grammar U), 1977 (In Japanese)
(4) K. O. Apel: Noam Chomskys Sprachtheorie und die philosophie der Gegenwart, 1971, (Japanese issue by S.Iguchi, Taishukan, 1976)
(5) M. Kazita: The Trace of Transformational Theory (In Japanese), Taishukan, 1976
(6) N. Chomsky: Cartesian Linguistics, (Japanese edition by Kawamoto Misuzu,
1966)
(7) N. Chomsky: Language and Mind, New York
1968
(8) H. Uchida: Japanese to English Translation System ATLAS U, Nikkei
Electronics, 17-Dec., 1984
(9) K. Muraki: Japanese to English Translation System PIVOT (In Japanese), Nikkei Electronics, 7-Dec., pp. 195-220, 1984
(10) M.Tokieda: Kokugogaku Genron (principles of Linguistics) (In Japanese), Iwanami,
1941
(11) T. Miura: The Theory of Noesis and Linguistics (In Japanese), Vol. 1' 3,
Keiso-Shobo, 1967
(12) Y. Morita: Conception by Japanese
(In Japanese), Koki-sha, 1981
(13) T. Miura (ed.): Critique of Modern
Linguistics, (In Japanese), Keiso-shobo, 1981
(14) T. Anzai: Conception in English
(In japanese), Kodan-sha, 1983
(15) C. Lancelot and A. Arnauld: Grammaire
generale et raisonnee, les fondements
de l'art de parler, 1966 (Japanese edition by H. Minamikata, Taishukan, 1972)
(16) C. J. Fillmore: Toward a Modern Theory of Case and Other Articles, Holt, Rinehart & Winston Inc. , New York, 1975 (Japanese edition by H. Tanaka and M. Funakoshi, Sanseido, 1975)
(17) T. Ishiwata: Grammar and Meaning (In Japanese), T, Asakura-shoten, 1983
Satoru Ikehara
Senior Research Engineer, Supervisor inthe NTT Communications and InformationProcessing Laboratories. Since joining theECL system in 1969, he has developed aformal algebraic manipulation language,queuing network analysis theory, andnatural language processing system. He ispresently developing a machine translationsystem. He received bachelor's degree, andmaster's degree, and Dr. Eng. degree fromOsaka University in 1967, 1969 and 1983.He was awarded the dissertation prizein 1982 for his research on queuingnetwork analysis from the InformationProcessing Society. He is a member of theInstitute of Electronics, Information andCommunication Engineers of Japan, andInformation Processing Society of Japan.
MASAHIRO MIYAZAKI
Senior Research Engineer, in the NTT
Communications and Information ProcessingLaboratories. Since joining the ECL system in1969, he has developed the computer systemDIPS-11, performance evaluation theory forcomputer systems and Japanese-text-to-speech-systems. He is presently developing amachine translation system. He received abachelor's degree in 1969 and Dr. Eng.degree from Tokyo Institute of Technologyin 1986. He is a member of theInstitute of Electronics, Information andCommunication Engineers of Japan and theInformation Processing Society of Japan.
SATOSHI SHIRAI
Senior Research Engineer, in the NTT
Communications and Information ProcessingLaboratories. Since joining the ECL system in1980, he has developed Japanese analysissystems for natural language processingsystems. He is presently developing a
machine translation system. He receivedbachelor's and master's degrees from OsakaUniversity in 1978 and 1980. He is a memberof the Institute of Electronics, Informationand Communication Engineers of Japan, and theInformation Processing Society of Japan.
AKIO YOKOO
Research Engineer of the NTT NaturalLanguage Processing Laboratory in the NTT
Communications and Information Processing
Laboratories. Since joining the ECL system in1982, he has developed a frame representation
language and a natural language processingsystems. He is presently developing a
machine translation system. He received a
bachelor's in 1980 and master's degree in1982 from the University of Electro-Communications. He is a member of theInstitute of Electronics, Information andCommunication Engineers of Japan, theInformation Processing Society of Japan, andthe Japanese Society Artificial Intelligence.