File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/87/p87-1018_metho.xml
Size: 27,334 bytes
Last Modified: 2025-10-06 14:12:00
<?xml version="1.0" standalone="yes"?> <Paper uid="P87-1018"> <Title>SEMANTIC STRUCTURE ANALYSIS OF JAPANESE NOUN PHRASES WITH ADNOMINAL PARTICLES</Title> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 1. Introduction </SectionTitle> <Paragraph position="0"> Japanese has many noun phrase patterns of the type A no B. The noun phrase pattern, which consists of two nouns A and B with an adnominal particle no, and which has at least the same ambiguity as B of A (and some additional ambiguities not found with the equivalent English construction), does not express any explicit semantic relations between the two nouns.</Paragraph> <Paragraph position="1"> Consequently, its interpretation depends mainly on the semantic characteristics of the nouns. Furthermore, phrase patterns NI no N2 no ... no Nn often appear. Because the number of possible dependencies between the constituents is 2 &quot;'I (2n-3)l! / n!, semantic analysis of such phrases is necessary to resolve the ambiguities. To date, there have been no adequate analyses for this linguistic phenomenon, nor have there been any clear methodological proposals for its semantic analysis.</Paragraph> <Paragraph position="2"> This paper describes a) the semantic diversity of A no B, b) the analysis of the semantic structure for A no B by a unification-based method of semantic function application, c) typical semantic structures of A no B, d) the possibility of paraphrasing A no B as a noun phrase with a relative clause by the addition of a verb, and e) the resolution of ambiguities using contextual information from the viewpoint of relation between A no B and its corresponding relative clause.</Paragraph> <Paragraph position="3"> Although A no B is a simple fo~n, it is interesting in two respects. First, A no B represents a general linguistic problem for semantic processing. The reason is that, in some cases, A or B is a noun form derived from a verb or adjective, thus necessitating the semantic processing of verbs and adjectives. Second, A no B can be paraphrased as a noun phrase with a relative clause, in just the same way as some English complex nominals \[3, 5\]. Putting it another way, as information is condensed into a simple expression, there are ambiguities as to the semantic relations between the two nouns. Consequently, contextual analysis plays a crucial part in the resolution of the ambiguities.</Paragraph> </Section> <Section position="4" start_page="0" end_page="127" type="metho"> <SectionTitle> 2. Semantic Diversity of A no B </SectionTitle> <Paragraph position="0"> A no B is frequently found in Japanese sentences.</Paragraph> <Paragraph position="1"> An exsmlnation of scientific and newspaper articles showed that the occurrence of A no B accounts for about half the total number of noun phrases in a text ill\]. The other occurrences are noun phrases with relative clauses, and coordinated noun phrases. In constructions of the type A no B, A or B can represent either a simple noun, as in Taroo no ie (&quot;raro's house&quot;), NP of the same A no B pattern, as in kariforunia no shuto no jinko (&quot;the population of the capital of California&quot;), or NP with a relative clause, as in Watashi ga atta hito no na (&quot;the name of the person who I met&quot;). There is also a fourth pattern involving an additional particle such as kara, made, de and so on, as in Tookyoo kara no densha (&quot;the train from Tokyo&quot;). This paper deals mainly with constructions of the first type, though the method presented here is also applicable recursively to patterns of the second and third types: this is possible because in such constructions, the semantic features of A (i.e. X no Y, or SY) derive from its head (Y). In the fourth type, analysis is slightly less straightforward, because the particle does provide some additional useful information.</Paragraph> <Paragraph position="2"> A no modifies a head B to restrict or clarify the referencetl. 21 of B. In the example Sutanfoodo daigaku no kyooju (&quot;professor at Stanford University&quot;), Sutanfoodo daigaku (&quot;Stanford University&quot;) restricts and clarifies the range of reference for kyooju (&quot;professor&quot;). Such A no B constructions can be classified semantically into five main groups according to the characteristics of A and B, as shown in Table 1. The five main groups can be further classified into a total of about 80 semantic relations. In the study mentioned above \[III, the authors examined about ten thousand examples of A no B occurrences, and checked the semantic relations.</Paragraph> <Paragraph position="3"> The appendix shows the semantic relations together with examples. It is necessary to analyze these seman- null 1. B functions as a predicate semantically, and A is its argument.</Paragraph> <Paragraph position="4"> /care no renhi (:0~ C/) ~, ~is love&quot;) B: ren'ai (.~,~, &quot;love ~) ... action, A: kate (~. &quot;he&quot;) ... agent of the action 2. B functions as a case role such as location, and is restricted relatively by A.</Paragraph> <Paragraph position="5"> gakkoo no nine (.-----~ C/) ~. &quot;front of a school') B: nine (~, &quot;front&quot;/&quot;oefore'} ... location/time, A: gakkoo (~. &quot;school') ... object 3. B is an attribute of A.</Paragraph> <Paragraph position="6"> hako no omosa (;U C/3 t ~ , &quot;weight of a box') B: omosa(lt ~, &quot;weight&quot;} ... attribute, A: kako(R, 'q0ox-) ... object 4. B is an argument of a predicate functioned semantically by A. sanpo no him (~ C/) ~,, &quot;man who strolls') B: Aim (/~, &quot;man') ... agent, A: sanpo (\['~. &quot;strolls&quot;) ... action 5. A is a kind of an attribute value orB.</Paragraph> <Paragraph position="7"> kooennoki(~--~o)YK, &quot;tree in a park') B: ki (~, &quot;Wee') ... object, A: kooen (~ \[\], &quot;park&quot;) ... value of an object's attribute location tic relations in such detail in order to produce good quality machine translation from Japanese into Eng null Lish among other tasks. To date, linguistic processing has not entailed such a detailed classification.</Paragraph> <Paragraph position="8"> The semantic structure of A no B is generally a function of the meanings of A and B, but the processing is not just a simple computation based on the semantic contents of A and B. For instance, when B functions as a predicate semantically, there is a case relation between A and B. However, there are no syntactic clues such as a case particle, unlike in full sentences. Hence, it is necessary to consider the semantic characteristics of A and B in order to analyze the semantic structure.</Paragraph> <Paragraph position="9"> Processing of context \[12\] is generally necessary to determine the correct semantic structure of A no B uniquely, as A no B is often ambiguous if considered out of context. For instance, in the case of Ft~ransujin no hanashi (&quot;speech of a Frenchman&quot;), there are two possible semantic relations for Furansujin (&quot;Frenchman&quot;): i.e. as agent or content of hanashi (&quot;speech&quot;). 3. Semantic Structure Analysis of A no B</Paragraph> <Section position="1" start_page="123" end_page="124" type="sub_section"> <SectionTitle> 3.1 Analysis by Function Application </SectionTitle> <Paragraph position="0"> The semantic structure of A no B is generally analyzed from A and B by &quot;semantic function application&quot;, which is similar to the idea of function application in the CUG framework (categorial unification grammar) 14. za\], viewing either A or B as a functor, and the other as its argument.</Paragraph> <Paragraph position="2"> From a different viewpoint, this is a generalization of the method of case frame analysis in which the analysis of the semantic structure of a verb-plus-noun phrase is based on the case-frame of the verb. That is, when a verb as a functor is applied to a noun phrase as its argument, if the noun phrase and a slot of the case-frame unify, the semantic structure is obtained as a result of assigning the relevant information from the noun phrase to the slot- So, the analysis is a kind of semantic treatment using the unification-based method. In this view, the case frames correspond to subcategorization frames, and the analysis corresponds to unifications applied to a subcategorization frame Is, s\] Characteristics of the function-based analysis are mainly to express input-output relations clearly, and to put stress on a lexical-based method.</Paragraph> <Paragraph position="3"> As the meaning of A no B depends on the individual A and B, it follows that each lexical entry must have information regarding its &quot;functionality&quot;. This is also the method adopted in CUG. Furthermore, these functors, arguments, and resulting semantic structures are represented as sets of at1~ribute-value pairs, again as in CUG. This is also similar to frame representations found in AI. The set of attribute-value pairs associated with a functor noun and an argument noun are generally represented as in Figure 1, and will be called a &quot;semantic structure&quot;. The characteristics of these structures are described in Section 3.3. In the representation, the attributes left and right indicate an argument for a functor word and a position (direction), and the values represent conditions imposed on the argument. Syncat, semcat and sense indicate syntactic, semantic and head word meaning respectively.</Paragraph> <Paragraph position="4"> Marker indicates the case particle found as a post-position with the noun phrase. Pred gives semantic conditions which restrict and clarify the relation between A syncat: < syntactic.features :> semcat~ < semarUic-feoJures :> sense: < word.senae > marker. < c~e-partic\[e :> leR: NONE right: syncat: <~ syntactic.features > semcat: < semantic-features > sense: \[\] pred: < cuae.rmme >: syncat: (syncat) semcat: (semcac) sense: (sense) case: < syntactic -cc~se-name :> marker: (marker) result: syncat: <: syntactic-\[eatures ~> semcat: <semarutic-features > sense: < word-senae marker: \[\] pred: (right pred) Figure la Format for a functor noun having an argument at its right and B. Result shows sets of attribute-value pairs obtained by the semantic function application. In the representation, words in parentheses such as (syncat) and (right pred) are path notations and are used to point to a value in the manner of an index notation Isl.</Paragraph> </Section> <Section position="2" start_page="124" end_page="124" type="sub_section"> <SectionTitle> 3.2 Semantic Structure Analysis of A no B </SectionTitle> <Paragraph position="0"> The noun phrase A no B is regarded as a composi- null tion of A no and B. Therefore, A no B is composed of A no and B by the function role of either A no or B. Which of A no or B has a function role depends on syntactic and semantic characteristic as described in section 3.3. Then A no is regarded as being constructed from A and no. Accordingly, the semantic structure of A no B is analyzed as follows: First, the functor no gets argument A, and makes a noun phrase A no with the semantic characteristics inherited from A. Secondly, the functor A no or B gets an argument B or A no respectively and makes a noun phrase A no B with the semantic characteristics inherited from B. The analysis process is shown as follows.</Paragraph> <Paragraph position="1"> (1) functor: no, argument: A, result: Ano (2) functor: Ano, argument: B, result: AnoB, or functor: B, argument: Ano, result: AnoB In the case of A p no B (where p is an additional particle), A and p are combined first. The semantic structure of A p is almost the same as that of A no except for the additional information derived from the marker p. After this, the final semantic structure is composed in the same way as for A no B. This paper focuses mainly on the analysis process after constituents of A no B have been found, and does not pay specific attention to the method of how constituents are found, for which purpose the active chart parsing method is used.</Paragraph> <Paragraph position="2"> With regard to the composition of A no, we take the choice giving no the functor role from the viewpoint of generality, although it is possible to view A as having this role. No has a functor role that shifts characteristics and functions of A to the semantic structure of A no, and adds a marker feature to the semantic structure of A no. The representation of no is shown in Figure 2. In the analysis of A no B, the semantic characteristics and functions of A and B weigh heavily, because although there is an adnominal case particle no, it is semantically rather neutral compared with other case particles. To put it another way, case particles usually function as explicit indicators of the preferred semantic interpretation. This fact suggests the significance of studying the method of analysis of A no B.</Paragraph> <Paragraph position="3"> When A no has a functor role, the functor must get B as its argument and extract a semantic relation between A and B. For example, in guruupu no shuukai (&quot;meeting of a group&quot;), guruupu no modifies an action nominal and makes a result semantic structure indicating the semantic relation (agent) as in Figure 3. In the representation >pred indicates a constraint that an argument must have a pred feature.</Paragraph> <Paragraph position="4"> The main semantic category of A no B is generally taken from the head B of A no&quot; B. However, in some cases the semantics of B are different from those of A no B, and it is necessary to change the semantic catesyncat: p sense: no(c), no) left: syncat: {n np} semcat: \[\] sense: \[\] marker: no left: NONE right: \[\] result: \[\] right: NONE result: syncat: np semcat: (left semcat) sense: (left sense) marker: no left: NONE right: (left right) result: (left result) ('meeting of a group') gories. For example, heita,&quot; (&quot;soldier&quot;) is animate, but oraocka no heitai (&quot;toy soldier&quot;) is not. Therefore omocAa no has the function of changing the semantic category of the head which it modifies. Such a function is obtained by a kind of overwriting unification 19!</Paragraph> </Section> <Section position="3" start_page="124" end_page="126" type="sub_section"> <SectionTitle> 3.3 Semantic Structures in Five Main Groups </SectionTitle> <Paragraph position="0"> The characteristics of the semantic structures in the f~ve ma/n groups are as follows.</Paragraph> <Paragraph position="1"> \[Case 1\] In this case, B, which is the nominal form of a predicate (a verb or an adjective), functions as an ar~ument~ and A, which is a semantic case argument of B, functions as a functor. Notice that when B functions semantically as a predicate, there are two alternatives for the assignment of the functor role. The first is that the predicate word functions as the functor. The second is the reverse L41. This paper adopts the latter way mainly because of the characteristic of free word order in a Japanese sentence.</Paragraph> <Paragraph position="2"> The semantic structure of A and A rw is almost the same except for a marker feature, and has the following functor role: when A no is an obligatory case (argument) of the predicate B, A no unifies with the argument feature of&quot; B. When A no is an optional case (adjunct), the semantic structure of A no is added to that of B as an optional case by unification. The functor role is added to A by a kind of lexical rule. Ez~mples are shown in Figures 3 and 4.</Paragraph> <Paragraph position="3"> \[Case 2 and Case 3\] In these cases, B represents a kind of case role or attribute respectively, which functions as a predicate. So, functionality is given to A in the same way as described above. Examples are shown in Figures 5 and 6.</Paragraph> <Paragraph position="4"> \[Case 4\] The reverse case of Case 1, that is, A is the nominal Form of&quot; a predicate, and B is the semantic case element of the predicate. So B is a functor and A no is its argument in the reverse way. The example is shown in Figure 7. Kooen (&quot;park&quot;) in the example gets an argument in the opposite direction to that of example 4.</Paragraph> <Paragraph position="5"> The phrase in this case corresponds to a noun phrase with a relative clause. So, a feature embedded is used in the representation, that is, it means that the pred feature is introduced from the complement.</Paragraph> <Paragraph position="6"> \[Case 5\] Semantic relations in this case are classified mainly into three types : a) relational restriction such as a human relation, b) attributive restrict/on such as a kind relation and c) situational restriction such as a location relation.</Paragraph> <Paragraph position="7"> (a) relational restr/ction: This case includes the relationships between humans, organizations, and whole-part relations. Generally a predicate role is given to B and a functor role is given to A in the same way as Case 1. An example is shown in Figure 8. In the example, sensei (&quot;teacher&quot;) has a pred feature and is an argument of the functor watasA~ ('I&quot;).</Paragraph> <Paragraph position="8"> (b) attributive restriction: A has attributive character- null istics such as quantity, kind, degree, and property, and B is generally a thing. As A functions as a kind of pred- &quot; icate, a predicate feature is assigned to A. An example is shown in Figure 9 with kooshifima r~o n,,no Ccheckered-pattern cloth&quot;), where kooshijurna has a pred feature and is an argument of the functor ~,,no (&quot;cloth&quot;). (c) situational restriction: A has situational meanings such as location, time, source, destination, purpose, and method, and restricts B by the situation. Like the relational restriction case, B is assigned a predicate feature, and A a functor role as shown in Figure 10. In the example, doozoo (&quot;oronze statue&quot;) has a pred feature and is an argument of the functor kooen (&quot;park&quot;).</Paragraph> </Section> <Section position="4" start_page="126" end_page="127" type="sub_section"> <SectionTitle> 3.4 Organization of Lexical Information </SectionTitle> <Paragraph position="0"> To assign an appropriate semantic structure to a noun, the following characteristics must be considered: a) A or B which works as a predicate in some cases works as a modifier (argument or adjunct) of a predicate in the other cases, as with kenkyuu (&quot;research&quot;, &quot;study&quot;) in the example gengo no kenkyuu (&quot;study of language&quot;) and kenkyuu no kaishi (&quot;start of the research&quot;). Therefore, A or B generally has both roles of a predicate and a modifier.</Paragraph> <Paragraph position="1"> b) When there are several no's in a noun phase such as A no B no C, there are several possibilit/es as to the word dependency structure. There are two principal Semantic structure of kooen no doozoo ('bronze statue in a park') possibilities: ((Ano B) no C) as in, for example, jiyuu no raegami no shashin (&quot;photograph of the Statue of Liberty&quot;), and (Ano (Brm C)) as Kariforunia.san no jooshitsu no kome (&quot;rice of fine qaulaity from California&quot;). Thus, the middle noun (B) may relate to the words on either side (A and C), or to only the right-hand word (C). In the ~rst case, the middle noun may be an argument of the predicate on both sides. In the latter case, the right,hOSt word C may be an argtunent of each predicate to the left, the number of which is not in general restricted.</Paragraph> <Paragraph position="2"> c) There are two cases of (A no (B no C)). When C is a nominal predicate, A and B might be separate arg~nents as in Kinoo no Taroo no Sanpo (&quot;raro's walk of yesterday&quot;). When C is an ordinary noun, however, the analysis is further complicated by the fact that implicit predicates such as location, possession, attribution etc., are involved, For example, in Tookyoo no NTT no biru ('~rrr's building in Tokyo&quot;), the inner predicate structure for NTT no bits (&quot;NTT has a building&quot;) is attached to the appropriate argument of the outer predicate Tookyoo no biru Cbuilding is in Tokyo&quot;).</Paragraph> <Paragraph position="3"> From the characteristics described above and the method for assigning a functor role to an axg~nent of a predicate, we adopt the method that a funcmr role is added to a constituent by a kind of lexical rule before function application. In general, several candidate constituents are made by ~he feature structure preformation. For example, at the stage ofAnoB - Ano B, when B is a functor and has a meaning such as location, time and so on, two solutions for B are offered as candidates: one as an argument of Ano, which works as a predicate, and the other as an adjunct.</Paragraph> </Section> </Section> <Section position="5" start_page="127" end_page="128" type="metho"> <SectionTitle> 4 Correspondence between A no B and </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="127" end_page="127" type="sub_section"> <SectionTitle> the Sentence 4.1 Paraphrase of A no B as a Noun Phrase </SectionTitle> <Paragraph position="0"> with a Relative Clause The expression A no B can be paraphrased into A p V B or A' B, adding an appropriate particle p and verb / adjective V, or reforming A to a verbal form A' if appropriate. Both A p V and A' are relative clauses. The paraphrased expression is more informative and some of the ambiguity is resolved. Paraphrases of A no B in Case 1 - Case 4 are rather easy, as added verbs/adjectives do not depend so much on context as compared with Case 5. Noun phrases with a relative clause for each case in the A no B classification are shown in Table 2.</Paragraph> <Paragraph position="1"> Such paraphrases are obtained by a change from a verb-centered to a noun-centered view. A no B is generally related to some event or state in a discourse, and the event or state is represented by an appropriate predicate: pred(A, B). By taking a noun-centered view, the representation is transferred into a representation A \[pred(A(*), B)\], that is, A in pred(A, B).</Paragraph> <Paragraph position="2"> The expression that gives the corresponding predicate is taken from the value of the pred attribute in the semantic structure. A noun phrase paraphrased with a relative clause is generally constructed as follows: 1) the head B is put first, 2) a verb is chosen based on the rel attribute, and put to the left orB, 3) a noun phrase corresponding to the appropriate case role as given by the argument structure of the predicate, is constructed from A and the particle indicated by a default-marker.</Paragraph> <Paragraph position="3"> and put to the left of the verb. For instance, in zoo no omosa (&quot;weight of an elephant&quot;), first, the head omosa is taken; second, verb rnotsu (&quot;nave&quot;) is taken from a value of rel, and put to the left of omosa; third, the agent zoo ga (&quot;elephant&quot;) is put to the left of omosa. In this way, the desired complex noun phrase zoo ga motsu omosa (&quot;weight that an elephant has&quot;) is arrived at.</Paragraph> </Section> <Section position="2" start_page="127" end_page="128" type="sub_section"> <SectionTitle> 4.2 On Disambiguation by Contextual Information </SectionTitle> <Paragraph position="0"> Although A no B is semantically ambiguous, it can generally be disambiguated by contextual information.</Paragraph> <Paragraph position="1"> Although inferences including association and analogy are generally necessary, this paper briefly mentions the possibility of the disambiguation method by unification</Paragraph> <Paragraph position="3"> p: ga / o / de / ni (case particles), V: suru (&quot;do&quot;) I ohonau (&quot;do&quot;) / okoru (&quot;happen&quot;) hare no hehhon (&quot;his marriage&quot;) -~ bare ga suru kehkon ('marriage that he performs&quot;) \[Case2\] --, A p V B p: ga / o (case particles), V: aru (&quot;be') / suru ('do&quot;) / shita ('done&quot;) /e no ma~ ('front of a house&quot;) --* iegaaru mae (&quot;front of a place where a house is') p: n.i I ga I ham / no tame ni (particles), V: aru (&quot;be in&quot;) / motsu ('have') / tsuhurareru (&quot;be made') / ohosu (&quot;cause&quot;) ~oen no doozoo (&quot;statue in a park&quot;) -b hoo~n ni aru doozoo (&quot;statue which is in a park') between a predicate structure in A no B semantic structure and the related event structure in the discourse. A sequence of related events is described in a discourse. On the other hand, the semantic structure is represented by an appropriate predicate feature. From these, the correct structure can be obtained by unifying an event semantic structure with a predicate feature in A no B as follows.</Paragraph> <Paragraph position="4"> event-semantic-structure-in-context - pred-structure-in-semantic-structure-of-A no B Here, &quot;-&quot; means that the left hand side unifies with the right hand side.</Paragraph> <Paragraph position="5"> Ambiguities of A no B may result from amibiguities regarding the predicates that could be added, ambiguities in the words themselves, or ambiguous case relations. The disambiguation process is illustrated below using an example in which the added predicates are ambiguous. Generally, a verb-centered semantic structure is extracted from a sentence. For the sentence, (sl) Hanako wa kyonen e o k.aita.</Paragraph> <Paragraph position="6"> ('~Hanako painted a picture last year.&quot;) the following semantic structure is obtained. This representation is simplified, showing only the information needed for the explanation.</Paragraph> <Paragraph position="7"> pred: \[reh paint agent: Hanako object: picture\] This semantic structure can be obtained also from the noun-centered semantic structure as follows.</Paragraph> <Paragraph position="8"> picture \[pred: reh paint agent: Hanako object: picture(*)\] Next, let us assume that the sentence (s2) occurs in the context of (sl).</Paragraph> <Paragraph position="9"> (s2) Hanako no e wa tenrankai de yuushoo shita.</Paragraph> <Paragraph position="10"> (&quot;The picture of Hanako/Hanako's picture won the first prize in an exhibition.&quot;) Hanako no e (&quot;the picture of Hanako&quot; or &quot;Hanako's picture&quot;) is ambiguous when taken out of context, with a range of possible semantic relations including possession, purchase, producer, and content.</Paragraph> <Paragraph position="11"> However, the ambiguity is resolved by unifying the semantic structure of the previous sentence with each of the semantic structures representing the possible semantic relations: the only semantic structure which can be successfully unified has the producer relation.</Paragraph> </Section> </Section> class="xml-element"></Paper>