File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/94/c94-2106_metho.xml
Size: 18,092 bytes
Last Modified: 2025-10-06 14:13:42
<?xml version="1.0" standalone="yes"?> <Paper uid="C94-2106"> <Title>A System of Verbal Semantic Attributes Focused on the Syntactic Correspondence between Japanese and English</Title> <Section position="3" start_page="0" end_page="672" type="metho"> <SectionTitle> 2. Semantic Structure of Verbal Patterns </SectionTitle> <Paragraph position="0"> This chapter examines the relationship between the usage of verbs I and the semantic structure of verbs. In machine translation systems, it is well known that the translation pattern pairs of source IALT-J/E's pattern dictionaries include both verbs and adjectives. Japanese adjectives are the equivalent of English 'be Adjective': for example &quot;A-ga utsukushii&quot; => &quot;A is beautiful&quot;. 'verbs' will be used to refer to both verbs and adjectives from now on.</Paragraph> <Paragraph position="1"> hmguage and target language sentence are effective in detmrnining the meaning of verbs.</Paragraph> <Paragraph position="2"> Our machine translation system, AI,T-J/E, uses two types of Japanese to English transfer pattern dictionaries based on verbs: the semantic valence pattern transfer dictionary and the idiomatic expression transfer dictionary (Fig. 1). These dictionaries consist of pairs made of \[The semantic valence Imttern transfer d ictimmryl (1) NI(SUBJECI'S)-ga N2(F(X)D)-wo taheru eat => N1 eatt N2 (2) Nl(*)-ga yomigaeru revive :> NI revive \[The idiomatic expression transfer dictionary l (1) NI(sUBJECTS)-ha se-ga takai back high => NI is tall.</Paragraph> <Paragraph position="3"> Fig. 1 Japanese to English Transfer Pattern</Paragraph> <Section position="1" start_page="672" end_page="672" type="sub_section"> <SectionTitle> Dictionaries </SectionTitle> <Paragraph position="0"> (The semantic constraints are shown in parenthesis, * indicates flmre is no senmntic constraint.) Japanese unit sentence patterns derived from Japanese verbs 2 with semantic constraints to their case elements and English patterns which correspond to thc Japanese expressions, t&quot;or example, pattern(l) in Fig. 1 shows how, if the Japanese verb is &quot;taberu&quot; and the noun phrase with a &quot;ga&quot; particle, which shows a subject, has the semantic attribute SUBJECTS and the noun phrase with a &quot;we&quot; particle, which shows a direct object, has the setnantic attribute FOOl) then the verb should be translated as &quot;eat&quot;. The noun phrase with the &quot;ga&quot; particle is translated as the English subject. The noun phrase with the &quot;we&quot; particle is translated as the English direct object. Here, wc exantine the rehttionship between the usage of verbs and the semantic structure of verbs using verbal patterns that have been entered into the Japanese to English transfer pattern dictionaries of ALT-J/IL Fig. 2 shows an example of entries in the Japanese to English transfer pattern dictionary which indicate the patterns of the 2In the idiomatic expression transfer dictionzu'ies, these arc the core secg)r of idiomatic exp,essions such as &quot;Abura we uru&quot; literally, &quot;to sell oil&quot;, but kliomatically, &quot;to idle away time&quot;.</Paragraph> <Paragraph position="1"> Japanese verb &quot;tsutsumu&quot;. This verb has three patterns.</Paragraph> <Paragraph position="2"> The first example shows a pattern pair indicating that the equiwtlent of the Japanese expression &quot;N 1 (SUBJECTS) ga N2 (CONCRETE OBJECTS or PEOPI,E) we N3(CLOTHtiS or I'APERS) de tsutsumu&quot; is the English expression &quot;N1 wrap N2 in/with N3&quot;. When the Japanese verb &quot;tsutsumu&quot; was used with these cases, this sentence gives the impression that NI really does the wrapping action. So, in this case, this pattern has the w:rb meaning &quot;N1 conducts bcxlily action.&quot;.</Paragraph> <Paragraph position="3"> The second example shows a pattern pair indicating that the equivalent expression of the Japanese expression &quot;NI (FIRE, ATMOSPIIERE or AIR) ga N2(CONCRETE OBJECTS, CULTtlRE or PLACES) we laulsumu&quot; is the English expression &quot;NI envelop N2&quot;. This sentence gives the impression that the state of N2 which isn't usually enveloped by N1 changes to the enveloped state. So, even though the same Japanese verb &quot;ISUL~'UmU&quot; was used with these cases, in this case, the pattern has a verb meaning of &quot;NI changes N2's attributes.&quot;. The third example shows a pattern pair indicating that the equiwdent of the Japanese expression &quot;N 1 (FOG) g(l N 2(C()NCRETF~ OBJECI'S or PLACES)we tsutxumu&quot; is the English expression &quot;N1 veil N2&quot;. In this case, this sentence gives the impression that a natural phenomenon, 'fog', has occurred. So, this pattern has the meaning &quot;Natural Phenomena have arisen&quot;.</Paragraph> <Paragraph position="4"> As shown in these examples, maintaining expressions in pairs which indicate both the common meaning and their usage between the Japanese and English, enables us to eliminate many conceptual ambiguities and makes it possible to give detailed and accurate attribute values to the Japanese verb &quot;tsutsumu&quot;. As in the case of the Japanese verb &quot;tsutsumu&quot;, one verb normally has several kinds of conceptual structures. But one verbal pattern which indicates common word meanings and their use between the Japanese and English (which differ so vastly in syntactic structure) corresponds to one conceptual structure. So, it is possible to eliminate the conceptual ambiguity of verbs by selecting verbal patterns in syntactic semantic analysis. In Japanese to English machine translation, we estimate there are tens of thousands of verbal patterns which need to be defined. If the usage of these patterns can be expressed by a small number of verbal semantic attributes, it is possible to track the semantic relationships of verbs easily. When giving verbal semantic attributes to a pair of individual Japanese and English patterns, it is possible to refer to the meaning of verbs not only in Japanese but also in English.</Paragraph> <Paragraph position="5"> 3. System of Verbal Semantic</Paragraph> </Section> <Section position="2" start_page="672" end_page="672" type="sub_section"> <SectionTitle> Attributes 3.1 Classification Standards for Verbal Semantic Attributes </SectionTitle> <Paragraph position="0"> Regarding the classification of verbs for use in machine translation, Nishida et al. (1980) proposed a system of verbal classification. This system of classification was introduced to resolve syntactic and semantic ambiguities of English in English to Japanese machine translation. To this system, they added the semantic attributes of verbs to the patterns of English verbs proposed by Hornby (1975) and determined the case structures depending on the combination of these two kinds of information.</Paragraph> <Paragraph position="1"> This system of verbal semantic attributes was introduced on the condition that the features of syntactic structures are expressed by Hornby's patterns of English verbs. So, this system of classification focused only on word meaning.</Paragraph> <Paragraph position="2"> Therefore this system can not be applied as such to the classification of Japanese verbs because Hornby's patterns can't be applied directly to Japanese verbs. No one has yet to propose exhaustive patterns like Hornby's for Japanese verbs.</Paragraph> <Paragraph position="3"> We expanded our system based on the discussions in section 2, using the following two factors.</Paragraph> </Section> </Section> <Section position="4" start_page="672" end_page="676" type="metho"> <SectionTitle> * Dynamic Characteristics of verbs </SectionTitle> <Paragraph position="0"> Classification based on a verb's meaning and its effects on the discourse: This classification is based on the types of action that can be understcx)d to have occurred when a verb is expressed and what situations have been brought about.</Paragraph> <Paragraph position="1"> Ex. &quot;motsu&quot;(to have) -- Possession &quot;kaihatsusuru&quot;(to develop) -- Production The verb &quot;motsu&quot; indicates that there is an act of possession within the context. In contrast, the verb &quot;kaihatsusuru&quot; indicates that there is something being produced within the context.</Paragraph> <Paragraph position="2"> * Relationslfip of Verbs to Cases Classification based on the role which the cases play with the verbs that govern them: This classification is based on the roles played by the case elements governed by the verb expressed.</Paragraph> <Paragraph position="3"> &quot;kanseisuru&quot; and &quot;kaihatsusuru&quot; &quot;are both verbs which indicate acts of production. But whereas &quot;kanseisuru&quot; indicates that the SUBJ is being produced, &quot;kaihatsusuru&quot; indicates that the SUBJ produces the OBJ.</Paragraph> <Section position="1" start_page="672" end_page="676" type="sub_section"> <SectionTitle> 3.2 Semantic Attribute System considering </SectionTitle> <Paragraph position="0"> the Semantic Relationship between Verbs We created a system of verbal semantic attributes as explained above. The semantic attribute values were determined using the usage patterns of typical Japanese verbs. First we classified verbs focussing on their dynamic characteristics. Next, we classified each group again focussing on the relationships of verbs to their cases. The top levels of the created system of verbal semantic attributes are shown in Fig. 3. The left side of this figure lists classifications as based on the dynamic characteristics of the verbs (their meanings). The right side lists the classifications based on the relationship of verbs with their cases (their usage). On the basis of these classification criteria, 97 verbal semantic attributes have been established.</Paragraph> <Paragraph position="1"> We evaluated the coverage of the verbal semantic attributes shown in chapter 3 by examining the verbal semantic attributes for each Japanese to English pair (about 15,000 pairs) in the Japanese to English transfer pattern dictionaries 3.</Paragraph> <Paragraph position="2"> Fig.4 shows how many transfer patterns were created for each verb in the semantic that were counted for each different verb. &quot;File percentage of patterns that came from verbs with more than one pattern was 73.4%. In these verbs that have multiple patterns, the percentage that had different kinds of verbal semantic attributes added to the patterns were 70.1%. This result shows that it is possible to classify semantic attributes for each verb by adding verbal semantic attributes to Japanese and English transfer pairings.</Paragraph> <Paragraph position="3"> Next we counted the number of verbal semantic attribute values givett for each pattern.</Paragraph> <Paragraph position="4"> Fig. 5 shows how many verbal semantic attributes 3Attribute values from a general noun attribute system classified into some 2,800 types have been i)rovidcd as selnantic constraints to the case elements of these patterns (lkchara ct al. 1991) enabling accurate selections of syntactic structures.</Paragraph> <Paragraph position="5"> were used by how many patterns. About 90% of patterns can be described by just one attribute value. This result shows that by giving the verbal semantic attributes proposed in this paper to each pattern in ALT-J/E, even in instances where multiple meanings may exist for a given Japanese verb, meanings can be selectively limited when verbs are viewed in terms of pattern pairings. The verbal semantic attributes which were given in each pattern have the potential to become an important key to tracking semantic relationships between sentences as is shown in chapter 5.</Paragraph> <Paragraph position="6"> Fig.6 shows the most frequent ten verbal semantic attributes for all the patterns. In these verbal semantic attributes, the patterns that ATTRIBUTE was added can almost all be described by only one attribute value (26.4% out of 27%). By contrast, the many patterns semantic attributes to each pattern No. 1 :ATTRIBUTE, Coverage: 27.0% Number of added VSA: 1:26.4%, 2 or more:0.6% described by BODILY ACTION or ATTRIBUTE TRANSFER was added can't be described by one attribute value (2.8% out of 12.7% and 2.1% out of 7.9 %, respectively). These 2 kinds of attribute values indicate the SUBJECT'S Physical Action, and it tends to be difficult to resolve the semantic ambiguities for these patterns. As shown in Fig.6, a few verbal semantic attributes cover a large proportion of patterns. For example, the sum of the coverage of the most frequent attribute value, A'KFRIBUTE, and the second most fi'equent attribute value, BODILY ACTION, cover 39.7 % of all patterns. For these attributes, even if there are several patterns for a given verb, sometimes the same attribute value was given to all the patterns. So the system of verbal semantic attributes is not sufficient to resolve the semantic ambiguities. For such attributes, we need more detail. We are plalming to subdivide these attribute values in the future.</Paragraph> </Section> </Section> <Section position="5" start_page="676" end_page="677" type="metho"> <SectionTitle> 5. Applications for Context Processing </SectionTitle> <Paragraph position="0"> in this chapter, we show examples of applications in context processing.</Paragraph> <Section position="1" start_page="676" end_page="676" type="sub_section"> <SectionTitle> 5.1 Analysis of Anaphoric Reference of Japanese Zero Pronouns </SectionTitle> <Paragraph position="0"> Using verbal semantic attributes to analyze anaphoric referents of zero pronouns appearing in Japanese texts is one applicati(m that has been considered (Nakaiwa et al. 1992). This technique pays attention to verbal semantic attributes and the relationship between the semantic attributes of tim verbs which govern zero pronouns and the semantic attributes of ttle verbs which govern case element candidates which may be anaphoricatly referred to. The contexts are carefully examined to determine anaphoric reference elements.</Paragraph> <Paragraph position="1"> This method has been realized in the machine translation systmn AIA'-J/E. The enhanced ALTJ/E was assessed by processing common Japanese newspaper articles. It was found that 95% of the Japanesc zero pronouns requiring anaphoral resolution in the 102 sentences from 30 newspapcr articles' lead paragraphs tlad their referents determined correctly using rules tuned for the 102 sentences(window test). In tile case of a blind test, the rate of success in anaphora resolution in which the zero pronoun referent exists within the sentence in another 98 sentences from newspaper articles was about 83% using tile rules. To demonstrate the effectiveness of this method, we evaluated the performance of the method proposed by Walker et.al. (1990) using the 98 sentences. Its rate of success in anaphora resolution where the zero pronoun referents existed within the sentence was about 74%. This result shows that our method is more effective than Walker's method, and that the rules used in our method determine universal relationships between verbs. If a few rules appropriate ff)r tile 98 sentences are added, tile rate increases to 95%. This result shows that the load imposed by rule customization is low.</Paragraph> <Paragraph position="2"> Even in the case of sentences in machine translation systems for which target meas cannot be constrained, this method allows the construction of rules independent of the translation target areas by means of verbal semantic attribute pairings. Using the verbal semantic attributes, anaphoric reference resolution of zero pronouns can be conducted with a limited volume of knowledge.</Paragraph> </Section> <Section position="2" start_page="676" end_page="677" type="sub_section"> <SectionTitle> 5.2 Supplementation of Elements Outside </SectionTitle> <Paragraph position="0"> Sentences against Elliptical Case Elements Verbal semantic attributes can be used with elliptical case elelnents in Japanese texts to supplement case elements whose referents do not appear within tim texts. To analyze such elliptical phenomena, it is possible to use case elements' semantic constraint conditions to estimate supplementary elements. Semantic information used to estimate supplementing elelnents is a constraint on cases for selecting the transfer f)attcrn. With this xnethod, therelbre, the majority of the constraints involve abstract semantic information, fi'equently posing difficulties in pinpointing elements to be supplemented. For example, if in Fig. 1(2), &quot;Ni(*)--ga yomigaeru(revive)&quot;, N l were to be omitted, ttle case element N I has no seman|ic constraint, and supplementary elements to the case can't be determined. In this case, it is effective Io complete the case element corresponding to S|JBJECT using tilt&quot; verbal semantic attributes of the pattern, &quot;N i's b(xtily state is transfcrled&quot;. Thus if a method presuming supplementary elements of elliptical case elements corresponding to the verbal semantic attributes is used, the deduction of more accurate supplementary elements would be possible.</Paragraph> </Section> <Section position="3" start_page="677" end_page="677" type="sub_section"> <SectionTitle> 5.3 Application for Other Context Processings </SectionTitle> <Paragraph position="0"> The verbal semantic attributes can be applied to other context processing problems. Estimating the relationship between verbs by pairing of the verbal semantic attributes, analysis of the tenses relationship of events as indicated by certain sentences and events indicated by another, together with sentence abridgment can be considered.</Paragraph> </Section> </Section> class="xml-element"></Paper>