File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/83/a83-1027_metho.xml
Size: 26,084 bytes
Last Modified: 2025-10-06 14:11:28
<?xml version="1.0" standalone="yes"?> <Paper uid="A83-1027"> <Title>TRANS FE R, GENE RAT I ON S WATASHI-WA HON-WO MOTSU ,,-..._./ S S AUX WATASHI-WA HON-WO MOTSU NAI MOTANAI</Title> <Section position="1" start_page="0" end_page="0" type="metho"> <SectionTitle> AN APPLICATION OF MONTAGUE GRAMMAR TO ENGLISH-JAPANESE MACHINE TRANSLATION </SectionTitle> <Paragraph position="0"/> </Section> <Section position="2" start_page="0" end_page="157" type="metho"> <SectionTitle> ABSTRACT </SectionTitle> <Paragraph position="0"> English-Japanese machine translation requires a large amount of structural transformations in both grammatical and conceptual level. In order to make its control structure clearer and more understandable, this paper proposes a model based on Montague Gramamr. Translation process is modeled as a data flow computation process. Formal description tools are developed and a prototype system is constructed. Various problems which arise in this modeling and their solutions are described. Results of experiments are shown and it is discussed how far initial goals are achieved.</Paragraph> <Paragraph position="1"> I. GOAL OF INTERMEDIATE REPRESENTATION DESIGN Differences between English and Japanese exist not only in grammatical level but also in conceptual level. Examples are illustrated in Fig.l. Accordingly, a large amount of transformations in various levels are required in order to obtain high quality translation. The goal of this research is to provide a good framework for carrying out those operations systematically.</Paragraph> <Paragraph position="2"> The solution depends on the design of intermediate representation (IR). Basic requirements to intermediate representation design are listed below.</Paragraph> <Paragraph position="3"> a) Accuracy: IR should retain logical conclusion of natural language expression. The following distinctions, for example, should be made in IR level: - partial/total negation - any-exist/exist-any &quot;- active/passive - restrictive use/ nonrestrictive use, etc.</Paragraph> <Paragraph position="4"> In other words, scope of operators should be represented precisely.</Paragraph> <Paragraph position="5"> <E2 her arrival makes him happy ~.. \[s needed paraphrasing <j> KARE WA KANOJO GA TOUCHAKU SHITA NODE URESHII.</Paragraph> <Paragraph position="6"> (he becomes happy because she has arrived) Fig.l. Examples of Differences between English and Japanese.</Paragraph> <Paragraph position="7"> <E>: English; <J>: Japanese.</Paragraph> <Paragraph position="8"> b) Ability of representing semantic relations: In English-Japanese translation, it is often the case that a given English word must be translated into different Japanese words or phrases if it has more than one word meanings. But it is not reasonable to capture this problem solely as a problem of word meaning disambiguation in analysis phase; the needed depth of disambPSiuation depends on target language. So it is also handled in transfer phase. In general, meaning of * given word is recognized based on the relation to other constituents in the sentence or text vhicb is semantically related to the given word. To make this poaslble in transfer phase, IR must provide a link to semantically related constituents of a given item. For example, an object of a verb should be accessible in IR level from the verb, even if the relation is implicit ~n the surface structure (as., passives, relative claus=a, and their combinations, etc.) C/) Prediction of control: given an IR expression, the model should be able to predict explicitIy what operations are co be done in what order.</Paragraph> <Paragraph position="9"> d) Lexicon driven: some sort of transformation rules ere word specific. The IR interpretation system should be designed Co deal with those word specific rules easily.</Paragraph> <Paragraph position="10"> e) Computability: All processing= should be effectively computable. Any IR is useless if it is not computable.</Paragraph> <Paragraph position="11"> 2. PRINCIPLE OF TP, ANSLATION This section outlines our solution Co the requirements posed in the preceding section. We employ MonCague Gram=mr (HonCague 1974, Dowry 1981) as a theoretical basis of translation model. Inter~edlate representation is designed based on intensional logic. Intermediate representation for a given natural language expression is obtained by what we call functional analysis.</Paragraph> <Section position="1" start_page="156" end_page="156" type="sub_section"> <SectionTitle> 2.1 Functional Analysis </SectionTitle> <Paragraph position="0"> In functional analysis, input sentence is decomposed into groups of constituents and interrelationship among those groups are analyzed in terms of function-argument relationships.</Paragraph> <Paragraph position="1"> Suppose a sentence: I don't have a book. (l) The functional analysis makes following two points: a) (L) is decomposed as: &quot;I have a book&quot; / &quot;nOt&quot;. (2) b) In the decomposition (2), &quot;not&quot; is an operator or function co &quot;I have a book.&quot; The result of this analysis can be depicted as follows: ~ &quot;&quot;I have a book&quot; I (3) wherel >denotes a function and\[ Idenotes en argument. The role of &quot;not&quot; as a function is: &quot;not&quot; as a semantic operstor: it negates a given proposition; &quot;not&quot; is a syntactic operator: it inserts an appropriate auxiliary verb and = lexical item &quot;not&quot; into appropriate position of its argument. (4) This kind of analysis goes on further with embedded sentence until it is decomposed into lexical units or even morphemes.</Paragraph> </Section> <Section position="2" start_page="156" end_page="157" type="sub_section"> <SectionTitle> 2.2 Montague Grammar as a Basic Theory </SectionTitle> <Paragraph position="0"> Montague Grammar (MG) gives a basis of functlonel analysis. One of the advantages of MG consists in its interpretation system of function form (or intensional logical form). In MG, interpretation of an intenelonal logical formula is a mapping I from incenaional logical formulas to set theoretical domain. Important property is chat this ampping I is defined under the constrainC of compositlonality, that is, I satisfies: Z\[f(a,b .... )\]'I\[fl(Ha\],Z\[b\] .... ), (5) without regard to what f, a, b, etc. are. This property simplifies control structure and it also specifies what operations are done in what order.</Paragraph> <Paragraph position="1"> For example, suppose input data has a structure like: A For the sake of property (5), ~he interpretation of (6) is done as a data flow computation process as followa: By this property, we can easily grasp the processing stream. In particular, we can easily ~hooc trouble and source of abnormality when debugging a system.</Paragraph> <Paragraph position="2"> Due to the above property and others, Ln particular due to its rigorous framework based .)n Logic, MG has been studied in ~nformation science field (Hobbs 1978, Friedman |978, Yonezaki \[980, Nishida 1980, Landsbergen 1980, Moran 1982, Moore 1981, Rosenschein 1982, ...). Application of MG to machine translation was also attempted (Hauenschild 1979, Landsbergen 1982), but those systems have only partially utilized the power of MG. Our approach attempts to utilize the full power of MGo</Paragraph> </Section> <Section position="3" start_page="157" end_page="157" type="sub_section"> <SectionTitle> 2.3 Application of Montague Grammar to Machine Translation </SectionTitle> <Paragraph position="0"> In order to obtain the syntactic structure in Japanese from an intensional logical form, in the same way as interpretation process of MC, we change the semantic domain from set theoretical domain to conceptual domain for Japanese. Each conceptual unit contains its syntactic expression in Japanese. Syntactic aspect is stressed for generating syntactic structure in Japanese.</Paragraph> <Paragraph position="1"> Conceptual information is utilized for semantic based word choice end paraphrasing.</Paragraph> <Paragraph position="2"> For example, the following function in Japanese syntactic domain is assigned to * logical item &quot;not&quot;:</Paragraph> <Paragraph position="4"/> </Section> <Section position="4" start_page="157" end_page="157" type="sub_section"> <SectionTitle> 3.1 Definition of Formal Tools </SectionTitle> <Paragraph position="0"> e) English oriented Formal Representation (EFR) is a version of intensional logic, and gives a rigorous formalism for describing the results of functional analysis. It is based on Cresswell's lambda deep structure (Cresawell 1973). Each expression has a uniquely defined type. Lambda form is employed to denote function itself.</Paragraph> <Paragraph position="1"> b) Conceptual Phrase Structure (CPS) is a data structure in which syntactic and semantic information of a Japanese lexicel unit or phrase structure are packed.</Paragraph> <Paragraph position="2"> i) example of CPS for a lexical item: EIGO:\[NP &quot;EIGO&quot; with,ZSAmLANGUAGE; ...,\] (9) category; lexical item; conceptual info.</Paragraph> <Paragraph position="3"> ; &quot;EIGO&quot; means English&quot; language. ii) example of CPS for phrase structure: \[NP \[ADJ &quot;AKAI&quot; with ... \] \[NOUN &quot;RINGO&quot; with ... \] with ... \] (i0) Transfer-generation process for the sentence (1) looks like: &quot;I don't have a book&quot; ~',,I have a book&quot; I ; &quot;AKAI&quot; means red, and &quot;RINGO&quot; means apple. c) CPS Form (CPSF) is a form which denotes operation or function on CPS domain. It is used to give descriptions to mappings from EFR to CPS. Constituents of CPSF are: i) Constants: CPS.</Paragraph> <Paragraph position="4"> ii) Variables: x, y, ... .</Paragraph> <Paragraph position="5"> (indicated by lower case strings).</Paragraph> <Paragraph position="6"> iii) Variables with constraints: e.g., (! SENTENCE x).</Paragraph> <Paragraph position="7"> ; variable x which must be of category SENTENCE.</Paragraph> <Paragraph position="8"> iv) Transformations:</Paragraph> <Paragraph position="10"> Using those description tools, translation process is modeled as a three staged process:</Paragraph> </Section> </Section> <Section position="3" start_page="157" end_page="159" type="metho"> <SectionTitle> 3. FORMAL TOOLS </SectionTitle> <Paragraph position="0"> Formal description tools have been developed co provide a precise description of the idea mentioned Ln the last section.</Paragraph> <Paragraph position="1"> stage I (analysis): anlyzes English sentence and extracts EFR form, stage 2 (transfer): substitutes CPSF to each lexical item in the EFR form,</Paragraph> <Paragraph position="3"> Fig.2. Example of Translation Process // Prefix notation is used for CPSF, described using Formal Tools. / and syntactic aspect is emphasized.</Paragraph> <Paragraph position="4"> stage 3 (generation): evaluates the CPSF to get CPS; generation of surface structure from CPS is straightforward.</Paragraph> <Paragraph position="5"> In order to give readers an overall perspective, we illustrate an example in Fig.2.</Paragraph> <Paragraph position="6"> Note that the example illustrated includes partial negation. Thus operator &quot;not&quot; is given a wider scope than &quot;always&quot;.</Paragraph> <Paragraph position="7"> In the remaining part of this section we will describe how to extract EFR expression from a given sentence. Then we will discuss the problem which arises in evaluating CPSF, and give its possible solution.</Paragraph> <Section position="1" start_page="158" end_page="159" type="sub_section"> <SectionTitle> 3.2 Extracting EFR Expression from Input Sentence Rules for translating English into EFR form </SectionTitle> <Paragraph position="0"> in .~ssociated with each phrase structure rules.</Paragraph> <Paragraph position="1"> For example, the rule looks llke: NP -> DET+NOUN where <NP>-<DET>(<NOUN>) (ii) where, <NP> stands for an EFR form assigned tu ~he NP node, etc. Rule (II) says chat EFR for an NP is a form whose function section is EFR for a DET node and whose argument section is EFR for a NOUN node. This rule can be incorporated into conventional natural language parser.</Paragraph> </Section> <Section position="2" start_page="159" end_page="159" type="sub_section"> <SectionTitle> 3.3 Evaluation of CPSF </SectionTitle> <Paragraph position="0"> Evaluation process of CPSF is a sequence of lambda conversions and tree ~ransformations.</Paragraph> <Paragraph position="1"> Evaluation of CPSF is done by a LISP ~ncerpreter- like algorithm. A problem which we call higher order problem arose in designing the evaluation algorithm.</Paragraph> </Section> <Section position="3" start_page="159" end_page="159" type="sub_section"> <SectionTitle> Higher Order Problem </SectionTitle> <Paragraph position="0"> By higher order property we mean that there exist functions which take other functions as arguments (Henderson 1980). CPSF in fact has this property. For example, an adjective &quot;large&quot; is modeled as a function which takes a noun as its argument. For example, large(database), &quot;large database&quot; (12) On the other hand, adverbs are modeled as functions to adjectives, For example, very(large), extremely(large), comparatively(large), etc. (13) The difficulty with higher order functions consists in modifiction to function. For explanation, let our temporal goal be regeneration of English from EFR. Suppose we assign to &quot;large&quot; a lambde form like:</Paragraph> <Paragraph position="2"> which takes a noun and returns a complex noun by attaching an adjective &quot;large&quot;. If the adjective is modified by an adverb, say &quot;very&quot;, we have to modify (14); we have to transform (14) into a lambda form like:</Paragraph> <Paragraph position="4"> which attaches a complex adjective &quot;very large&quot; to a given noun. As is easily expected, it is too tedious or even impossible to do this task in general. Accordingly, we take an alternative assignment instead of (14), namely: large <- \[ADJ &quot;LARGE&quot;\]. (16) Since this decision cuases a form: \[ADJ &quot;LARGE&quot;\](\[NOUN &quot;DATABASE&quot;\]), (17) to be created in the course of evaluation, we specify what to do in such case. The rule is defiend as follows:</Paragraph> <Paragraph position="6"> ; which may read: is y:\[there is a uniquely specified object y referred to by an NP &quot;the table&quot;, such that y is a block which is restricted to be located on x.\] This lambda form is too complicated for tree transformation procedure to manipulate. So it should be transformed into equivalent CPS if it exists. The type of the lambda form is known from the context, namely one-place predicate. So if we apply the lambda form (20) to &quot;known&quot; entity, say &quot;it&quot;, we can obtain sentence structure like:</Paragraph> </Section> </Section> <Section position="4" start_page="159" end_page="159" type="metho"> <SectionTitle> SORE WA TSUKUE NO UE NO BLOCK DEARU </SectionTitle> <Paragraph position="0"> it a block on the ~able is (it is a block on the table) (21) From this result, we can infer that the lambda form (20) is equivalent to a noun:</Paragraph> <Paragraph position="2"> This rule is called an application rule.</Paragraph> <Paragraph position="3"> In general, evaluation of \[ambda form itself results in a function value (function as a value). This causes difficulty as mentioned above. Unfortunately, we can't dispense with lambda forms; lambda variables are needed to link gap and its antecedent in relative clause, verb and its dependants (subject, object, etc), preposition and its object, etc. For example, in our model, an complex noun modified by a PP: &quot;block on the table&quot; (19) PSs assigned a following EFR: Of course, this way of processing is not desirable; it introduces extra complexity. But this is a trade off of employing formal semantics; the same sort of processing is also done rather opaque procedures in conventional MT system.</Paragraph> </Section> <Section position="5" start_page="159" end_page="162" type="metho"> <SectionTitle> 4. MODELING TRANSLATION PROCESS </SectionTitle> <Paragraph position="0"> This section illustrates how English-Japanese translation process is modeled using formal tools. Firstly, how several basic linguistic constructions are treated is described and then mechanism for word choice is presented.</Paragraph> <Section position="1" start_page="160" end_page="162" type="sub_section"> <SectionTitle> 4.1 Translating Basic Constructions of English </SectionTitle> <Paragraph position="0"> a) Sentence: sentence consists of an NP and a VF. VP is analyzed as a one-place predicate, which constructs a proposition out of an individual referred Co by the subject. VP is further decomposed into intransitive verb or cranaltive verb + object. Intransitive verbs and transitive verbs ere analyzed as one-place predicates and two-place predicate, respectively. One-place predicate and two-place predicate are assigned a CFSF function which generates a sentence ouc of an individual and chat which generates a sentence out of a pair of individuals, respectively. Thus, a transitive verb &quot;constructs&quot; is assigned a CPSF form:</Paragraph> <Paragraph position="2"> ; given two individuals, this function attaches co each argument a case marker (corresponding to JOSHI or Japanese postfix) and then generates a sentence structure.</Paragraph> <Paragraph position="3"> The assignment (24) may be extended later to incorporate word choice mechanism.</Paragraph> <Paragraph position="4"> Treatment of NP in MonCague-besed semantics is significant in chat EFR expression for an NP is given a wider scope then Chat for a VP. Thus the EFR form for an ~P-VP construction looks llke: <~>(<w>), (25) where <x> means EFR form for x, x=NP,... .</Paragraph> <Paragraph position="5"> The reason is Co provide an appropriate model for English quantifier which is syntactically local but semantically global. For example, first order logical form for a sentence: &quot;this command needs no operand&quot; (267 looks Like:</Paragraph> <Paragraph position="7"> where operator &quot;not&quot;, which comes from a determiner &quot;no&quot;, is given a wider scope than &quot;needs&quot;. This translation is straightforward in our model; the following EFR is extracted from (26):</Paragraph> <Paragraph position="9"> \[f we make appropriate assignment including:</Paragraph> <Paragraph position="11"> In Engllsh-Japanese -,-'chine translation, this treatment gives an elegant solution to the :ranalation of prenominal negation, partial negation, etc. Since Japanese language does not have a synCactlc device for prenominal negation, &quot;no&quot; must be translated into asainly two separate constituents: one is a RENTAISHI (Japanese decerminer) and another is an auxiliary verb of negation. One possible assignment of CFSF looks like:</Paragraph> <Paragraph position="13"> In general, correspondence of ~P and individual is indirect in EFR. The association of an NF with its referent x is indicated as follows:</Paragraph> <Paragraph position="15"> one-place predlcaCe type ; <NP> stands for EFR expression for NP. (31) Most of ocher NP's correspond co ice referent more directly. The application rule reflecting this fact is:</Paragraph> <Paragraph position="17"> where, ix\] stands for a CPS for x.</Paragraph> <Paragraph position="18"> b) Internal structure of NP: the below illustrates the structure of EFR expression assigned process is determined by a CPSF assigned co <DET>, En cases of &quot;the&quot; or &quot;a/an&quot;, translation process is abic complicated. Et is almost the same as the process described in detail in section 3: firstly the <MODIFIER>s and <NOUN> are applied Co an individual like &quot;the chinE&quot; (the) or &quot;somechinE&quot; (a/an) and a sentence will be obtained; then a noun structure is extracted and appropriate RENTAISHI or Japanese determiner is attached.</Paragraph> <Paragraph position="19"> c) Other cases: some ocher cases are illustrated by examples in Fig.3.</Paragraph> <Paragraph position="20"> 4.2&quot;Word Choice Mechanism * In order to obtain high quality translation, word choice .~chanism must be incorporated at least for handling the cases like: i) subordinate clause: ; indirect question is generated first, then it is transformed into a sentence.</Paragraph> <Paragraph position="21"> Fig.3. Examples of Translation of Basic English Construction. <x>, {x}, \[x\] and &quot;x&quot; stand for EFR for x, CPSF for x, CPS for x, and CPB for Japanese string x, respectively.</Paragraph> <Paragraph position="22"> verb in accordance with its object or its agent, adjective-noun, adverb-verb, and preposition.</Paragraph> <Paragraph position="23"> Word choice is partially solved in the analysis phase as a word meaning disambiguation. So the design problem \[s to determine to what degree word sense is disamblguated in the analysis phase and what kind of ambiguities is left until transfer-generation phase. Suppose we are to translate a given preposition. The occurence of a preposition \[s classified as: (a) when it is governed by verbs or nouns: (a-l) when governmant is strong: e.g., study on, belong to, provide for; (a-2) when govern.ment is weak: e.g., buy ... at store; (b) otherwise: (b-I) idiomatic: e.g., in particular, in addition; (b-2) related to its object: e.g., by bus, with high probability, without/ING.</Paragraph> <Paragraph position="24"> We treat (a) and (b-l) as an analysis problem and handle them in the analysis phase. (b-2) is more difficult and is treated in the transfer-generation phase where partial semantic interpretation \[s done.</Paragraph> <Paragraph position="25"> Word choice in transfer-generatlon phase is done by using, conditional expression and attributive information included in CPS. For example, a transitive verb &quot;develop&quot; is translated differently according to its object:</Paragraph> <Paragraph position="27"> The following assignment of CPSF makes this choice poss ib le : To make this type of processing possible in the cases where the deep object is moved from surface object position by transformations, link information between verb and its (deep) object should be represented explicitly. The below shows bow it is done in the case of relative clause.</Paragraph> </Section> </Section> <Section position="6" start_page="162" end_page="163" type="metho"> <SectionTitle> 5 * EXPERIMENTS </SectionTitle> <Paragraph position="0"/> <Paragraph position="2"> In EFR level, lambda variable x is explicitly used as a place holder for the gap.</Paragraph> <Paragraph position="3"> A functor &quot;which&quot; dominates both the EFR for the embedded sentence and that for the head noun. A CPSF assigned to the functor &quot;which&quot; sends conceptual information of the head noun to the gap as follows: firstly it creates a null NF out of the head noun, then the null NP is substituted into the lambda variable for the gap.</Paragraph> <Paragraph position="4"> In word choice or semantic based translation in general, various kinds of transformations are carried out on target language structure. For example, her arrival makes him happy, (38) must be paraphrased into: he becomes happy because she has arrived (39) since inanimate agent is unnatural in Japanese. In order to retrieve appropriate lexical item of target language for transformation, mutual relations among lexlcal items are organized using network formalism (lexical net). The node represents a lexicel item and a link represents an association with specification of what operation causes that link t() be passed through. \[t also contains description of case ~ransformation needed Ln order co map case structure appropriately. The below illustrate s part of Lexical net: We have constructed a prototype system.</Paragraph> <Paragraph position="5"> It is slmplified then practical system in: - it has only limited vocabulary, - interactive disembiguation is done instead of automatic disambiguaCion, and - word choice mmchenism is limited to typical cases since overall definition of rules have not yet been completed.</Paragraph> <Paragraph position="6"> Sample texts are taken from real computer manuals or abstracts of computer journals.</Paragraph> <Paragraph position="7"> Initially, four sample texts (40 sentences) are chosen. Currently it is extended to I0 texts (72 sentences).</Paragraph> <Paragraph position="8"> Additional features are introduced Ln order to make the system more practical.</Paragraph> <Paragraph position="9"> a) Parser: declarative rules are inefficient for dealing with sentences in real cexts. The parser uses production type rules each of which is classified according to its invocation condition. Declarative rules are manually converted into this rule type.</Paragraph> <Paragraph position="10"> b) Automatic postedicor: transfer process defined so far concentrates on local processings. Even if certain kinds of ambiguities are resolved in this phase, there still remains a possibility that new ambiguity is introduced in generation phase. Instead of incorporating into the transfer-generation phase a sophisticated mechanism for filtering out ambiguities, we attach a postprocessor which will &quot;reform&quot; a phrase structure yielding ambiguous output. Treetree transformation rules are utilized here. Current result of our machine cransLacion system is shown in Appendix.</Paragraph> </Section> class="xml-element"></Paper>