File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/96/c96-2146_metho.xml
Size: 20,975 bytes
Last Modified: 2025-10-06 14:14:11
<?xml version="1.0" standalone="yes"?> <Paper uid="C96-2146"> <Title>Analyzing Japanese Double-Subject Construction having an Adjective Predicate</Title> <Section position="3" start_page="0" end_page="865" type="metho"> <SectionTitle> 2 Definition of key terms </SectionTitle> <Paragraph position="0"> Key terms used in this paper me defined as follows: * case-marldng particle In English, each case is marked by its relative posilion in terms of tile predicate or by a preposition. In contrast to lhis, Japanese language marks each case by a certain sort of postposilional parlicle located next to noun phvases sudl as &quot;ga&quot; or &quot;we&quot;. Such particles arc callcd &quot;case-markiu g pm-ticl es&quot;.</Paragraph> <Paragraph position="1"> * adverbial particle &quot;Adverbial particles&quot; give a case an additional function (topicalizalion etc.) by their allachmenl to Ibc cascmarkingparlicle. Some adverbial particles such as &quot;wa&quot; or &quot;me&quot; often stand-in for case-marking parlicles and give the case ;m additional fimclion.</Paragraph> <Paragraph position="2"> * sentence having an adjective predicate In the Japanese I~mguage, adjectives function as predicates in sentences as do verbs, qlmrefore, in this paper, lhe sentence in which m~ adjective acts as a predicate is called a &quot;senlence having an Mieclive predicate&quot;.</Paragraph> <Paragraph position="3"> . double-sul;jeet construction Usually, a simple sentence has only one subjeclivc case. ttowevcr, m~my Japanese adjeclives (and some verbs) can dominate two surface subjective cases within a simple sentence. Such sentence structure is called tile &quot;double-subject constrnction&quot;. This constn~ction includes both adverbial particle &quot;wa&quot; and subjective case-marking p~trticle &quot;ga&quot;. Therefore, the Japanese double-subject construction is also referred to as file &quot;wa&quot;-&quot;ga&quot; construction.</Paragraph> <Paragraph position="4"> * valency structure The sentence structure can be considered as a combination of a predicate and its modifiers. Such sentence stntcturc is called the &quot;valency structure&quot;. The valency structure represents what surface cases the modifiers dominated by a given predicate correspond to.</Paragraph> <Paragraph position="5"> * valency pattern &quot;Valency patterns&quot; are stnlcture patterns that formulate possible valency structures for predicates. A valency pattern is defined for each usage for each predicate. Figure 1 shows an example of a valency pattern for a verb &quot;shoukai sunl (introduce)&quot;. For vcrbs and adjectives, wdcncy patterns formulated as shown in figure 1 are collect ed beforehand in the dictionary called the valency pattern dictionary.</Paragraph> <Paragraph position="6"> Japanese valency pattern tbr &quot;~,~ 9'\]&quot; J-~&quot;</Paragraph> <Paragraph position="8"> Remarks: Ni: Label for a valency element SR: Semantic restriction on a noun</Paragraph> </Section> <Section position="4" start_page="865" end_page="865" type="metho"> <SectionTitle> * valency element </SectionTitle> <Paragraph position="0"> In the valency stn~cture, each relation between a predicate and its modifier is called a &quot;valency element&quot;. As shown in figure I, valency elements are described using both the semantic restriction on the noun including the modifier and the restriction on the case-marking pmaicles including the modifier.</Paragraph> <Paragraph position="1"> * sentence analysis (or analyzing a sentence) &quot;Sentence analysis&quot; is the process that reveals the valency structure of the input sentence. ALT-J/E perfonns sentence analysis by binding the modifiers for the predicate in the input sentence to the valency elements in the wtiency pattern for the predicate in the wdency pattern dictionary.</Paragraph> <Paragraph position="2"> * semantic category A &quot;semantic category&quot; is a class for dividing nouns into concepts according to their meaning. For example, both &quot;man&quot; and &quot;woman&quot; belong to the same semantic category \[(human)\]. In the below, the words surrounded by 1( )\] are semantic categories.</Paragraph> </Section> <Section position="5" start_page="865" end_page="866" type="metho"> <SectionTitle> 3 Japanese Sentence Analysis based on </SectionTitle> <Paragraph position="0"> the valency structure Figure 2 outlines the translation process in ALT-J/E (Hayashi, 1987). Morphological analysis segments the input Japanese sentence into its component words such as predicates and nouns. Dependency analysis determines the dependency structure to indicate the association between words. From the association between a predicate and its modifiers in the dependency structure, sentence analysis tries to determine the wflency strncture, i.e., it determines, for the valency pattern for the predicate, which valency element ead~ modifier corresponds to. This valency stnlcture is then converted into the equivalent English sentence structure. Finally, the outpnt English sentence is generated from the stnlcture.</Paragraph> <Paragraph position="1"> The following describes how ordintu'y sentences are analyzed by ALT-J/E using the valency structure for the predicate of file input sentence.</Paragraph> <Paragraph position="2"> Sentence analysis is the process that converts the dependency structure into file valency structure by referring to wdency patterns. All valency patterns for each usage with each predicate are prepared beforehand and held in the valency pattern dictionary.. As shown in figure 1, for each predicate, both file semantic restriction on a noun including its modifier and file restrictions on case-marking particles including its modifier are described for each valency element. When the modifier in the input sentence satisfies both restrictions for a certain wllency element in the valency pattern for the predicate, the modifier is botmd to that valency element. The valency smlctnre is determined by binding all predicate modifiers to valency elements. An adverbial particle can correspond to more than one case-marking pataicle, for example, &quot;wa&quot; is a possible proxy for &quot;ga&quot;, &quot;wo&quot;, &quot;hi&quot; and so on. Therefore, sentence analysis first tries to bind modifiers that have case-marking particles, each of which represents which case is unambiguously marked by particle spelling like &quot;ga&quot; Japanese input sentence: ~Lt:kg.L~L'~L:/)))~;~{7'~ b-l-<o (he introduced his sister to me) C/ morphological analysis and dependency analysis dependency structure .~ refer to valency pattern sentence analysis as ordinary construction</Paragraph> <Paragraph position="4"> marks subjective case, to wflency elements in tile wflcncy pattern lbr the prcdicate. Thc analysis thcn tries to bind modifiers that have ml adverbial particle to the non-bolmd valency elements.</Paragraph> <Paragraph position="5"> Figure 3 shows an example of this type of Japanese sentence mmlysis for the sentcncc &quot;kare wa watashi ni kate no imouto we shoukai shita (he theeduced his sister to me)&quot;. First, the mmlysis tries to bind modifiers with case-tam'king parlicles &quot;we&quot; and &quot;ni&quot; to tile wdency elements in the wdency pattern for the predicate &quot;shoukai sent (introduce)&quot; which is oblaincd IYom lhc valency pattem dictionary. As modifiers satisfy bolh restrictions on the noun memfing and case-marking parlicles, they ,are bound to N2 (objective case 1) mid N3 (objective case 2), respectively. The ~malysis then lfies to bind lhe modifier with adverbial particle &quot;wa&quot; to a non-bound valency element. So far, N2 and N3 in the wflcncy paltem have already been bound.</Paragraph> <Paragraph position="6"> Therefore, as shown in figure 3, the only non-bound valency element is N1 which is a subjective case. The adverbial particle &quot;wa&quot; c~m stand-in for case-marking p~wticlc &quot;ga&quot;, which is the non-bound wfiency clcment, mid the noun &quot;km-e (he)&quot; satisfies the scmantic restriction on the subjcctive case N1 i.e. \[(agent)\].</Paragraph> <Paragraph position="7"> Therefore, file modifier with &quot;wa&quot; is bonnd to tim subjective case N1. Finally, lhc valency structure as shown in figure 3 is obtained.</Paragraph> </Section> <Section position="6" start_page="866" end_page="869" type="metho"> <SectionTitle> 4 Processing the Japanese double-subject </SectionTitle> <Paragraph position="0"> construction Many Japmmse adjective predicates domhmtc two subjective cases and so form the double-subject construction. The double-subject constmctiou having an Miective predicate actually has several wlrimlts, so no one approach c,'m be uscd to mlalyze it. Accordingly, this section classifies the four types mid the characteristics of eadl type ~'e described.</Paragraph> <Section position="1" start_page="866" end_page="867" type="sub_section"> <SectionTitle> 4.1 Classification of the Japanese dnuble-subjcet </SectionTitle> <Paragraph position="0"> construction Tile Japanese doublc-subjcct consmlction (also calicO the &quot;wa&quot;-&quot;ga&quot; conslruction) can be classified into lhe following four types based on a previous study (lshigami, 1977): * type-I In this vafi~mt, adverbial particle &quot;wa&quot; is a proxy for a case-marking particle such as &quot;ni&quot; other than subjective case-ram-king &quot;ga&quot;. Example-1 shows &quot;wa&quot; as a proxy for case-marking particle &quot;hi&quot; in tile sentence &quot;wamshi no ie wa gakkou ga chikai (file school is near my house)&quot;.</Paragraph> <Paragraph position="1"> Example- 1 (myhouse) (adverbial (school) (case-marking (nea 0 pro'title &quot;wa&quot;) particle &quot;ga&quot;) ~1C/ &quot;wa&quot; is \[br &quot;ni&quot;. proxy (myhousc) (destination (school) (subjective (hem') case-marking case-marking particle &quot;ni&quot;)</Paragraph> <Paragraph position="3"> in this wu'iant, adverbial parlicle &quot;wa&quot; is a proxy for case-umrking panicle &quot;no&quot; representing a noun modifier (pre-nominal). Example-2 shows &quot;wa&quot; as a proxy for pre-nominal case-marking p~ulicle &quot;no&quot; in the sentence &quot;zou wa ham ga nagai (elcplumls have long trunks)&quot;. In this wu-itmI, the case of&quot;wa&quot; modifying a predicate must be analyzed as &quot;no&quot; modifying the Japanese input sentence: 7~7~h~~: t, ~o (elephants have long trunks) Japanese valency pattern for &quot;~ 1~ v, dependency structure ~ (subjective) JR: 7~ /F @@~ Ni: Label ~br a valency element refer to valency pattern SR: Semantic restriction on a noun on case-marking pazticles .' men&quot; ana y~ i: sentence analysis as ordinary construction as double-subject construction Therefore, this analysis involves re-formation of the valency structure.</Paragraph> <Paragraph position="5"> In this variant, the case with case-marking particle &quot;ga&quot; sometimes represents ~m objective case. Allhough an objective case is usually marked by case-marking particle &quot;wo&quot;, some adjective predicates have an oNective case marked by case-marking particle &quot;ga&quot;.</Paragraph> <Paragraph position="6"> Example-3 shows that &quot;kanojo (she)&quot; with &quot;ga&quot; is an objective case and &quot;kare (he)&quot; with &quot;wa&quot; is a subjective case in the sentence &quot;kare wa kanojo ga snkida (he likes her)&quot;. As case-marking particle &quot;ga&quot; normally indicates the subjective case, binding &quot;ga&quot; to the subjective case leads to incorrect mmlysis if only surface spelling is considered.</Paragraph> <Paragraph position="8"> In this variant, the case with adverbial particle &quot;wa&quot; acts as all adverbial phrase representing time and is actually a special form of type-1. Representing time is optional for most predicates. Moreover, Japm~ese time expressions are often translated into English adverbial phi'ases. Therefore, type-4 is separated from type-1 in this classification from viewpoint of enghmering.</Paragraph> <Paragraph position="9"> Example-4 shows that time expression &quot;6-gatsu wa (in Jmle)&quot; acts as an adverbial phrase in the sentence &quot;6-gatsu wa ame ga tot (it has much rain in June)&quot;.</Paragraph> <Paragraph position="11"> ~(t particle &quot;wa&quot;) particlc &quot;ga&quot;) ime expression with &quot;wa&quot; acts as an adverbial phrase (June) (rain) (subjective (much) (adverbial case-marking phrase) particle &quot;ga&quot;)</Paragraph> </Section> <Section position="2" start_page="867" end_page="869" type="sub_section"> <SectionTitle> 4.2 Problems in processing the Japanese double- </SectionTitle> <Paragraph position="0"> subject construction Type-I and type-4 cases can be analyzed using the processing flow described in section 3 because adverbial particle &quot;wa&quot; simply acts as a case-marking particle. However, the following problems ,arise when processing type-2 and type-3 cases in the normal way. * Problem with type-2 cases Fignre 4 shows the m~alysis of the type-2 example, &quot;zou wa hana ga nagai (elephanls have long thinks)&quot;. The predicate &quot;nagai (long)&quot; has only one valency element N1 with &quot;ga&quot;. According to ordinary sentence analysis, the modifier &quot;hana ga&quot; is bound to the valency element NI, which means that the other modifier &quot;zou wa&quot; is left unbound. That is, sentence analysis camlot be completed as shown in the left bottom of figure 4.</Paragraph> <Paragraph position="1"> This complicates the accurate Wanslation of this modifier into English.</Paragraph> <Paragraph position="2"> Figure 5 shows the analysis of sentence analysis of the type-3 example, &quot;kare wa k~mqio ga sukida (he likes her)&quot;. If modifiers with a case-marking parlicle are preferentially processed as described in seclion 3, then the modifier with &quot;ga&quot; binds to subjective case N1 which is wrong in this example. This leads us lo lhe wrong interpretation as shown in lhe left bottom of figure 5.</Paragraph> <Paragraph position="3"> 5 Proposed method for analyzing Japanese double-subject construction In order to overcome lhe problems described in the previous section, this section proposes a method for analyzing a Japanese double-subject conslmction having m~ adjective predicate. The mefl~od has fllrec processing phases. The first determines whether lhc input sentence has double-subject construction or not. The second determines which of tim four wtri~mls file sentcnce is. qlm last processes the sentence according Io its type.</Paragraph> <Paragraph position="4"> Figure 6 shows lhc processing flow. The input sentence has already undergone morphological analysis and dependency mmlysis, i.e., it has been tdready determhled whirl nouns modi\[y the adjective predicate wilh what sort of poslpositional p~nticle expressions. Earl processing phase is described below in detail with reference to figure 6.</Paragraph> <Paragraph position="5"> * Judgment of douMe-subject construction If the input sentence contains two modifiers that have adverbial particle &quot;wa&quot; mid case-tam'king particle &quot;ga&quot;, it is determined as a double-subject constnmtion.</Paragraph> <Paragraph position="6"> Sentences olher them double-subject construction, as sentence with an adjective predicate</Paragraph> <Paragraph position="8"> ~ tt\[i~-~nd each modifier to a valency 1 I\[ element i!} the valency paltern .... l / ~ yus / ~->--1 h~o bind each m~lilier I ~ro-ce~ing as type-l~ d, \[ \[ff~t wdency element \[ t~ ~~ tin the wdency pattern \] I ~,rocessing as tyl,e-2\] t'rocessing V as ordinmy sentence end Figure 6. Processing flow for analyzing a Japanese douMe-suhjeet construction.</Paragraph> <Paragraph position="9"> well as sentences with verb predicates, are processed normally as described ill section 3.</Paragraph> <Paragraph position="10"> * Type determination First, lype-4 is set if the modifier with adverbial parlicle &quot;wa&quot; represents a time expression or not. Tiffs is efficient because type-4 is the most specific type. This judgment is performed by checking whether the semantic calegory of the noun included in the modifier with &quot;wa&quot; is associated with \[(time)\] or not. Second, type-3 is set by the valency paltem for the predicate in lhC/ input. The predicate in a type-3 case has to cover both subjective and objective cases in its usage. The valency pattern for each usage for each predicate is defined in the valency pattern dictionary. Therefore, this judgment is performed by checking whether the valency pattern for the predicate includes both a subjcctive case and an objective case. Finally, type-1 and type-2 are differentiated according to the result of binding betwccn the modifiers in the input senteuce and thevalency elements inthe valency pattern. Type-2 is set ff the modifier with case-marking particle &quot;ga&quot; is bound to the subjective case and the modifier with adverbi~d particle &quot;wa&quot; is not. The remaining cases are judged as type-1. Here, modifiers with &quot;wa&quot; and &quot;ga&quot; ,are often bound in type-1. Notice that the binding process resolves what case-marking parlicle the adverbial particle &quot;wa&quot; stands-in for.</Paragraph> <Paragraph position="11"> * Determining the valency structure In the type-4 preprocessing step in figure 6, the modifier with &quot;wa&quot; (time reference) is considered as an adverbial plwase. Next, binding the other modifiers with the valency elements in the valency pattern for the prcdicate is attcmpted. This facilitates the translation of Japanese time expressions into English.</Paragraph> <Paragraph position="12"> In the type-3 prcprocessing step in figure 6, case-marking particle &quot;ga&quot; has to be converted into the case-marking particle &quot;wo&quot; before binding the former which represents an objective case. This conversion allows us to correctly bind each modifier to its appropriate valency element, becanse predicates in type-3 cases have both a subjective case with &quot;ga&quot; and an objective case with &quot;wo&quot;. See the example in the right bottom of fignre 5.</Paragraph> <Paragraph position="13"> In type-2 processing, adverbial particle &quot;wa&quot; is a proxy for pre-nominal case-marking particle &quot;no&quot; and the modifier with adverbial particle &quot;wa&quot; must be analyzed as the phrase which modifies the snbjective case with &quot;ga&quot;. As a result, type-2 mmlysis involves changing the valency structure. In figure 4, the noun phrase &quot;zou no hana&quot; is fornled from both &quot;zou (elcphant)&quot; with &quot;wa&quot; and &quot;hzma 0lose)&quot; with &quot;ga&quot; by converting &quot;wa&quot; into &quot;no&quot;, and &quot;zou no hana&quot; is bound to the subjective case for the predicate &quot;nagai (long)&quot;. See the example in the right bottom of figure 4.</Paragraph> <Paragraph position="14"> Moreover, type-2 processing tries to determine the semantic relation between the noun with &quot;ga&quot; and the noun with &quot;no&quot; (originally &quot;wa&quot;). Determining the semmltic relation helps us translate the Japanese double-subject construction into the appropriate English construction and expression. Although several semantic relations are known, at the present time ALT-J/E can resolve only two of them based on semantic categories: has-a relation and is-a relation.</Paragraph> <Paragraph position="15"> As adverbial particle &quot;wa&quot; in type-I cases is a proxy for a case-marking particle such as &quot;ni&quot;, &quot;de&quot; and so on, type-I cases can be processed in the way described in section 3.</Paragraph> <Paragraph position="16"> The above approach con'ectly determines die input sentence's valency structure which allows the maddne translation system to produce more accurate output.</Paragraph> <Paragraph position="17"> As a result, the method proposed here improves the translation accuracy of ALT-J/E.</Paragraph> </Section> </Section> class="xml-element"></Paper>