<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-1019">
  <Title>The Grammatical Function Analysis between Korean Adnoun Clause and Noun Phrase by Using Support Vector Machines</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Korean Adnoun Clauses and their
</SectionTitle>
    <Paragraph position="0"> analysis problems Adnoun clauses are very frequent in Korean sentences. In a corpus, for example, they appear as often as 18,264 times in 11,932 sentences (see section 4, for details). It means that effective analyses of adnoun clauses will directly lead to improved performance of lexical, morphological and syntactic processing by machine.</Paragraph>
    <Paragraph position="1"> In order to indicate the difficulties of the adnoun clause analysis, we need to have some basic knowledge on the structure of Korean adnoun clause formation. Thus, we will briefly illustrate the types of Korean adnoun clauses.</Paragraph>
    <Paragraph position="2"> Then, what makes the analysis tricky will be made clear.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 Two types of adnoun clauses
</SectionTitle>
      <Paragraph position="0"> There are two types of adnoun clauses in Korean : relative adnoun clause and appositive adnoun clause. The former is a more general form of adnoun clause and its formation can be exemplified as follows :  simple sentences 1.b and 1.c in terms of adnoun clause formation. The functional morpheme 'eul', which represents the object relation between 'chaeg' and 'sseoss-da' in 1.c, does not appear in 1.a but 'chaeg' is the functional object of 'sseu-n' in 1.a. This adnoun clause is called a relative adnoun clause whose complement moves to the NP modified by the adnoun clause and the NP modified by a relative adnoun clause is called a head NP. In 1.a 'geu-ga sseun' is a relative adnoun clause and 'chaeg' is its head noun (or NP).</Paragraph>
      <Paragraph position="1"> Let us consider another example of an adnoun clause.</Paragraph>
      <Paragraph position="3"> an-da(know).</Paragraph>
      <Paragraph position="4"> (Everybody knows the fact that he is honest.) The adnoun clause in 2 is a complete sentence which has all necessary syntactic constituents in itself. This type of adnoun clause is called an appositive adnoun clause. And the head NP modified by the appositive adnoun clause is called a complement noun (Lee, 1986; Chang 1995). In 2, 'geu-ga jeongjig-han' is an appositive adnoun clause and 'sasil' is a complement noun. Generally, such words as &amp;quot;iyu(reason), gyeong-u(case), jangmyeon(scene), il(work), cheoji(condition), anghwang(situation), saggeon(happening), naemsae(smell), somun(rumor) and geos(thing)&amp;quot; are typical examples of the complement noun (Chang, 1995; Lee, 1986).</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 The problems
</SectionTitle>
      <Paragraph position="0"> The first problem we are faced with when analyzing grammatical functions of Korean adnoun clauses is obviously the disappearance of the functional morphemes which carry important information, as shown in the previous subsection (2.1).</Paragraph>
      <Paragraph position="1"> Apart from the morpheme-ommission problem, there is another reason for the difficulty. As it is directly related to a language particular syntactic characteristic of Korean, we need first to understand a unique procedure of Korean relativization. Unlike English, in which relative pronouns (e.g., who, whom, whose, which and that) are used for relativization and they themselves bear crucial information for identifying grammatical function of the head noun in relative clauses (see example 1.a, in section 1), there is no such relative pronouns in Korean. Instead, an adnominal verb ending is attached to the verb stem and plays a grammatical role of modifying its head noun.</Paragraph>
      <Paragraph position="2"> However, the problem is that these verb ending morphemes do not provide any information about the grammatical function associated with the relevant head noun.</Paragraph>
      <Paragraph position="3"> Take 3.a-c for examples.</Paragraph>
      <Paragraph position="5"> meog-eun(ate) geu(he).</Paragraph>
      <Paragraph position="6"> (He who ate a rice in a restaurant.)  adnominal ending 'eun', the grammatical function of each relative noun is different. The grammatical function of the head noun in 3.a is subject, in 3.b, object and in 3.c, adverbial. The word order gives little information because Korean is a partly free word-order language and some complements of a verb may be frequently omitted. For example, in sentence 4, the verb of relative clause 'sigdang-eseo meog-eun(who ate in the restaurant or which one ate in the restaurant)' have two omitted complements which are subject and object. So 'bab' can be identified as either of subject or object in the relative clause.</Paragraph>
      <Paragraph position="8"> (I saw the rice which (one) ate in a restaurant.) Korean appositive adnoun clauses have the same syntactic structure of relative adnoun clauses as in example 2 in section 2.</Paragraph>
      <Paragraph position="9"> Yoon et al. (1997) classified adnoun clauses into relative adnoun clauses and appositive adnoun clauses based on a complement noun dictionary which was manually constructed, and then tries to find the grammatical function of a relative noun using lexical co-occurrence information. But as shown in example 5, a complement noun can be used as a relative noun, so Yoon et al. (1997)'s method using the dictionary has some limits.</Paragraph>
      <Paragraph position="11"> (He talked about the truth which he discovered.) Li et al. (1998) described a method using conceptual co-occurrence patterns and syntactic role distribution of relative nouns. Linguistic information is extracted from corpus and thesaurus. However, he did not take into account appositive adnoun clauses but only considered relative adnoun clauses.</Paragraph>
      <Paragraph position="12"> Lee et al. (2001) classified adnoun clauses into appositive clauses and one of relative clauses. He proposed a stochastic method based on a maximum likelihood estimation and adopted the backed-off model in estimating the probability P(r|v,e,n) to handle sparse data problem (the symbols r, v, e and n represent the grammatical relation, the verb of the adnoun clause, the adnominal verb ending, and the head noun modified by an adnoun clause, respectively).</Paragraph>
      <Paragraph position="13"> The backed-off model handles unknown words effectively but it may not be used with all the backed-off stages in real field problems where higher accuracy is needed.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="2" type="metho">
    <SectionTitle>
3 Support Vector Machines
</SectionTitle>
    <Paragraph position="0"> The technique of Support Vector Machines(SVM) is a learning approach for solving two-class pattern recognition problems introduced by Vapnik (1995). It is based on the Structural Risk Minimization principle for which error-bound analysis has been theoretically motivated (Vapnik, 1995). The problem is to find a decision surface that separates the data points in two classes optimally. A decision surface by SVM for linearly separable space is a hyperplane H : y = w[?]x - b = 0 and two hyperplanes parallel to it and with equal distances to it,  examples are called support vectors because they only participate in the definition of the separating hyperplane, and other examples can be removed and/or moved around as long as they do not cross  . In order to maximize the distance, we should minimize ||w ||with the condition that there are no data points between</Paragraph>
    <Paragraph position="2"> The SVM problem is to find such w and b that satisfy the above constraints. It can be solved using quadratic programming techniques(Vapnik, 1995). The algorithms for solving linearly separable cases can be extended so that they can solve linearly non-separable cases as well by either introducing soft margin hyperplanes, or by mapping the original data vectors to a higher dimensional space where the new features contain interaction terms of the original features, and the data points in the new space become linearly separable (Vapnik, 1995). We use</Paragraph>
    <Paragraph position="4"> system for our experiment (Joachimes, 1998).</Paragraph>
    <Paragraph position="5"> SVM performance is governed by the features. We use the verb of each adnoun clause, the adnominal verb ending and the head noun of the noun phrase. To reflect context of sentence, we use the previous noun phrase, which is located right before the verb, and its functional morpheme. The previous noun phrase is the surface level word list not the previous argument for the verb in adnoun clause. Part of speech(POS) tags of all lexical item are also used as feature. For example, in sentence 'Igeos-eun geu-ga sseu-n chaeg-ida.', 'geu' is a previos noun pharse feature, 'ga' is its functional morpheme feature, 'sseu' is a verb feature, 'n' is a verb ending feature, 'chaeg' is a head noun feature and all POS tags of lexical items are features.</Paragraph>
    <Paragraph position="6"> Because we found that the kernel of SVM does not strongly affect the performance of our problem through many experiments, we concluded that our problem is linearly separable. Thus we will use the linear kernel only.</Paragraph>
    <Paragraph position="7"> As the SVMs is a binary class classifier, we construct four classifiers, one for each class. Each classifier constructs a hyperplane between one class and other classes. We select the classifier which has the maximal distance from the margin for each test data point.</Paragraph>
  </Section>
  <Section position="6" start_page="2" end_page="2" type="metho">
    <SectionTitle>
4 Experimental Results
</SectionTitle>
    <Paragraph position="0"> We use the tree tagged corpus of Korean Information Base which is annotated as a form of phrase structured tree (Lee, 1996). It consists of 11,932 sentences, which corresponds to 145,630 eojeols. Eojeol is a syntactic unit composed of one lexical morpheme with multiple functional morphemes optionally attached to it. We extract the verb of an adnoun clause and the noun phrase which is modified by the adnoun clause. We regard an eojeol consisting of a main verb and auxiliary-verbs as a single main-verb eojeol. In case of a complex verb, we only take into account the first part of it. Every verb which has adnominal morphemes and the head word of a noun phrase which is modified by adnoun clause, were extracted. Because Korean is head-fiinal  The SVMlight system is available at http://ais.gmd.de/~thorsten/svm_light/.</Paragraph>
    <Paragraph position="1"> language, we regard the last noun of a noun phrase as the head word of the noun phrase.</Paragraph>
    <Paragraph position="2"> The total number of extracted pairs of verb and noun is 18,264. The grammatical function of each pair is manually tagged. To experiment, the data was subdivided into a learning data set from 10,739 sentences and a test data set from 1,193 sentences. We use 16,413 training data points and 1,851 test data points in all experiments.</Paragraph>
    <Paragraph position="3"> Table 1 shows an accuracy at each of the grammatical categories between an adnoun clause and a noun phrase with SVMs, compared with the backed-off method which is proposed by (Lee, 2001).</Paragraph>
    <Paragraph position="4"> Table 1. the acuracy of SVM and Backed-off model at each of the grammatical categories subj obj adv app total  It should be noted that SVM outperforms Backed-off model in Table 1. By using context information we acquire an improvement of overall 2.1%.</Paragraph>
    <Paragraph position="5"> Table 2 represents the accuracies of the proposed model compared with the Li's model.</Paragraph>
    <Paragraph position="6"> The category 'appositive' is not taken into account for fair comparison. It should be noted that Li et al. (1998)'s results are drawn from most frequent 100 verbs while ours, from 4,684 verbs all of which are in the training corpus.</Paragraph>
    <Paragraph position="7">  It is shown that our proposed model shows the better overall result in determining the grammatical function between an adnoun clause and its modifying head noun.</Paragraph>
    <Paragraph position="8"> Most errors are caued by lack of lexical information. Actually, lexical information in 19% of the test data has not occurred in the training data. The other errors are caused by the characteristics that some verbs in adnoun clauses can have dual subjects which we did not consider in the problem. Take 6 for an example.</Paragraph>
  </Section>
  <Section position="7" start_page="2" end_page="2" type="metho">
    <SectionTitle>
6. Nun-i(eyes) keu-n(be big) Cheolsu
</SectionTitle>
    <Paragraph position="0"> (Cheolsu who has big eyes) In example 6, the context NP is 'nun' and its functional word is 'i' which may represent that it is subject of 'keu-da', thus system may wrongly determine that 'Cheolsu' is not a subject of 'keu-da' because the subject of 'keu-da' has been made with 'nun'. However, both 'Cheolsu' and 'nun' are the subjects of 'keu-da'.</Paragraph>
  </Section>
class="xml-element"></Paper>