[NP coh-a-ha-nun TV phu-lo-ku-laim-ul]. [VP noh-chye + pe-li-ko + mal-ass-ta.] In intra-chunk dependency parsing, we make the. VP chunk a single node and ...
Resolving Ambiguity in Inter-chunk Dependency Parsing Mi-Young Kim, Sin-Jae Kang, and Jong-Hyeok Lee Div. of Electrical and Computer Engineering Pohang University of Science and Technology (POSTECH) San 31, Hyoja-dong, Nam-gu, Pohang, 790-784, R. of KOREA
{colorful,sjkang,jhlee}@postech.ac.kr
Abstract Recently, dependency grammar has become quite popular in relatively free word-order languages. We encounter many structural ambiguities when parsing a sentence using dependency grammar. We use a chunking procedure to avoid constructing a mistaken dependency structure. Chunking reduces the scope of dependency relations between dependents and governors. This paper presents a method to resolve ambiguity in inter-chunk dependency parsing by using valency information, a structural preference rule and a statistical model. The proposed method is a combination of a rule-based approach and a statistical model. The structural preference rule is an important clue to resolve ambiguity and complements the valency information. The statistical method, using structural, semantic, and lexical information, is applied to resolve ambiguity when selecting the governor of adjuncts. Experimental results show that dependency parsing using this method resolves ambiguity in inter-chunk dependency parsing with 88.03 % accuracy.
1
Introduction
Dependency grammar has been widely used for Korean language parsing because of its ability to handle variable-ordered word and discontinuous constituents, which is compatible with Korean (Na Dong Ryul, 1994). The grammar determines the relation between governors and dependents in a sentence. After parsing, we obtain many dependency trees because the input sentence has many structural ambiguities. We use a chunking
procedure to avoid constructing a mistaken dependency structure. Text chunking divides sentences into non-overlapping segments on the basis of fairly superficial analysis (Abney Steven, 1991). Consider the following Korean sentence.1 Na-nun o-nul cenyek-ey il-ccik cam-I tul-e-se coh-a-ha-nun TV phu-lo-ku-laim-ul noh-chye-pe-li-ko mal-ass-ta. (Since I went to bed early tonight, I missed my
favorite TV program) This sentence can be chunked as follows. Na-nun [NP o-nul cenyek-ey] il-ccik cam-I tul-e-se [NP coh-a-ha-nun TV phu-lo-ku-laim-ul] [VP noh-chye-pe-li-ko mal-ass-ta.]
In this manner, we chunk eojeols (corresponding to word-phrase in English, or bunsetsu in Japanese) to make an NP chunk and a VP chunk. Next, intra-chunk dependency parsing and inter-chunk dependency parsing are performed. After chunking, the scope of dependency is reduced because the dependent within a chunk cannot have a governor out of the scope of the chunk. This reduces parsing results and makes parsing fast and simple. We use the method developed in Kim (2000) to make NP chunking. We make an NP chunk using NP rules and bigram lexical patterns. For VP chunking, we use the following two rules. main verb + auxiliary verb quotative verb + non-adnominal verb After chunking, we decide the intra-chunk dependency relation. The intra-chunk dependency result is as follows. 1
We use Yale Romanization System for Korean Expression.
n o h -c h y e -p e -li-ko + m a l-a ss-ta .
Na-nun [NP o-nul cenyek-ey] il-ccik cam-I tul-e-se
n a -n u n
[NP coh-a-ha-nun TV phu-lo-ku-laim-ul]
tu l-e -se
ce -n y e k-e y
[VP noh-chye + pe-li-ko + mal-ass-ta.]
il-c c ik
o -n u l
In intra-chunk dependency parsing, we make the VP chunk a single node and determine the dependency relation within the NP chunk. Further, in inter-chunk dependency parsing, we determine the dependency relation among eojeols which are not chunked or are the head of chunks. The head of a chunk becomes the last eojeol of the chunk because Korean is a head-final language.
Na-nun cenyek-ey il-ccik cam-I tul-e-se phu-lo-ku-laim-ul noh-chye-pe-li-ko mal-ass-ta. (Since I went to bed early night, I missed the
program) The dependency tree after inter-chunk dependency parsing is shown in Figure 1. noh-chye-pe-li-ko+ m al-ass-ta. na-nun ce-nyek-ey
tul-e-se il-ccik
phu-lo-ku-laim -ul cam -i
Figure 1: The inter-chunk dependency tree
Figure 2 shows the dependency tree of the full sentence. Considering all these steps, we conclude the overall parsing procedure, as in Figure 3. Although we reduce the scope of governors of some dependents by chunking, both intra-chunk dependency and inter-chunk dependency have many structural ambiguities. This paper presents a method to resolve ambiguity in inter-chunk dependency parsing.
c a m -i c o h -a -h a -n u n
TV
Figure 2: The dependency tree of the full sentence
POS-tagged sentence Chunking NP Chunking VP Chunking
Inter-chunk dependency Determine dependency relation 1) Between complements and predicates Check cross links
Na-nun [NP o-nul cenyek-ey] il-ccik cam-I tul-e-se [NP coh-a-ha-nun TV phu-lo-ku-laim-ul] [VP noh-chye +pe-li-ko + mal-ass-ta].
Each underlined eojeol in the NP chunk is the head of the corresponding chunk. After chunking, this sentence can be simplified as follows. We only consider eojeols which are not chunked or are the head of chunks.
p h u -lo -k u -la im -u l
Intra-chunk Dependency
2) Between adjuncts and predicates Generate dependency tree
Figure 3: The overall parsing procedure
2
Ambiguity in Inter-chunk Dependency Parsing
Consider the following sentence. Na-nun o-nul cenyek-ey il-ccik cam-I tul-e-se coh-a-ha-nun TV phu-lo-ku-laim-ul noh-chye-pe-li-ko mal-ass-ta. (Since I went to bed early tonight, I missed my
favorite TV program) This sentence has many ambiguities. ‘na-nun’, ‘ce-nyek-ey’, ‘il-ccik’, ‘cam-i’ have two governor candidates like ‘tul-e-se’ and ‘noh-chye + pe-li-ko+ mal-ass-ta’. To resolve ambiguity, we suggest a method as follows.
2.1
The Suggested Method to Resolve Ambiguity
Early parsing work investigated rule-based approaches. Grammars are hand-crafted, often with a large amount of lexically-specific information in the form of subcategorization information. Ambiguity is resolved only through selectional restrictions (Allen James, 1987). Selectional restrictions, however, have many difficulties, and structural preferences also come into play during disambiguation (Collins Michael,
1999). As a result, researchers began to investigate machine-learning approaches to the problem, primarily through statistical methods. This paper proposes a hybrid method to resolve ambiguity. The proposed method is a combination of a rule-based approach and a statistical method. Using valency information, we can distinguish between complements and adjuncts. Therefore, we devide the disambiguation method into two steps. The first step is to determine the dependency relation between complements and predicates and second is between adjuncts and predicates. From the next section, we describe these two steps.
3
Resolving Ambiguity between Complements and Predicates
Figure 4 shows the disambiguation procedure for determining the dependency relation between complements and predicates. We use valency information to resolve ambiguity between complements and predicates. The dictionary describes valency information, including numerical valence, syntactic form, and selectional restrictions. However, in some cases, several governor candidates can share the same dependent as their complements. In these cases, structural preference provides a clue to resolve ambiguity. Therefore, we suggest some structural preference rule to complement valency information.
3.1
Structural Preference Rule
Consider the following example. Ku-nye-ka ce-nyek-ul mek-ko wun-tong-ul haiss-ta. (She had dinner and took exercise)
Dependent : Ku-nye-ka (subject) Governor candidates: mek-ko (valency : {subj. obj.}) haiss-ta (valency: { subj. obj.}) In this example, the governor candidates of the dependent ‘Ku-nye-ka’ are ‘mek-ko’ and ‘haiss-ta’. These two governor candidates require the subject as their complements.
Dependent is complement Use valency information Disambiguate ?
yes
no Use structural preference rule Select governor
Figure 4: The disambiguation procedure for determining the dependency relation between complements and predicates
NP ‘Ku-nye-ka’ is the subject of the verb phrase ‘ce-nyek-ul mek-ko wun-tong-ul haiss-ta’. So we choose the governor of ‘Ku-nye-ka’ as ‘haiss-ta’ which is the head of the verb phrase. From this example, we induce structural preference rule Ⅰas in Figure 5. If several governor candidates require the same dependent as a complement, Î We consider all these governor candidaes share the same complement (= dependent). The governor of the dependent is the last candidate Figure 5 : structural preference rule Ⅰ
4
Resolving Ambiguity between Adjuncts and Predicates
It is impossible to decide the governor of an adjunct using valency information. Thus, it needs structural preference by intuition or a statistical method using a parsed corpus. If a dependent is a predicate, structural preference is a primary clue to determine the governor rather than the statistical method and semantic, or syntactic information. If a dependent is an adverbial, the lexical information of the adverbial is important to decide its governor. In this case, we use a statistical method, using structural, semantic, and lexical information. The overall flowchart is given in Figure 6. In the next subsection, we describe structural preference rule Ⅱ and statistical method to resolve ambiguity.
Dependency relation
Dependent is adjunct Dependent is predicate ?
no
man-ci-cak-ke-li-myen-se
Dependent is adverbial
yes
subj.(ku)
Use structural preference rule
iss-ten
Remove unsuitable candidates by checking cross links Use statistical method by structural, semantic and lexical information
Figure
4.1
6: The disambiguation procedure for determining dependency relation between adjuncts and predicates
Resolving ambiguity between Predicates and Predicates
Suppose a sentence has many predicates. When we determine the dependency relation among predicates in this sentence, structural information is more important than lexical and semantic information. Let us consider the following example.
cwung-el-ke-li-ca
tut-ko -nun
po-myen -so
mal-haiss -ta
subj.(pak-cin-seng)
Figure 7: An example of structural preference
Na-nun ku sa-sil-ul tut-ko-se yeph-ey iss-ten a-i-lul han-pen an-a-cwun hwu cip-u-lo tal-lye-kass-ci-man cip-ey-nun a-mu-to eps-ess-ta. (I hugged the child beside me after I heard the
news. Then I rushed home, but there was no one home.) Dependent : tut-ko-se Governor candidates : iss-ten (valency :subj(a-i), loc-adverb.) an-a-cwun (valency :subj(na),obj) tal-lye-kass-ci-man
Ku-ka kha-tu-lul man-ci-cak-ke-li-myen-se hon-cas-mal-lo cwung-el-ke-li-ca yeph-ey iss-ten pak-cin-seng-i tut-ko-nun si-gyey-lul po-myen-se mal-haiss-ta.
(While he murmured to himself fidgetting with cards, Park sitting beside him heard the noise and said to him looking at the watch.) In this example, there are 5 governor candidates of the dependent ‘man-ci-cak-ke-li-myen-se’. man-ci-cak-ke-li-myen-se Æ cwung-el-ke-li-ca man-ci-cak-ke-li-myen-se Æ iss-tun man-ci-cak-ke-li-myen-se Æ tut-ko-nun man-ci-cak-ke-li-myen-se Æ po-myen-se man-ci-cak-ke-li-myen-se Æ mal-haiss-ta
The true governor of ‘man-ci-cak-ke-li-myen-se’ is ‘cung-el-ke-li-ca’. We can determine this governor using structural information, not semantic or lexical information. ‘man-ci-cak-ke-li-myen-se’ and ‘cung-el-ke-li-ca’ share the same subject, but the subject of the other 3 candidates is different from that of ‘man-ci-cak-ke-li-myen-se’. Therefore, we determine the governor of ‘man-ci-cak-ke-li-myen-se’ is ‘cwung-el-ke-li-ca’. Let us consider another example.
(valency :subj(na),loc-adverb.) eps-ess-ta (valency : subj(a-mwu)) Let us decide the governor of ‘tut-ko-se’. There are 4 governor candidates. After dependency parsing between complements and predicates, we obtain the subjects of these 4 candidates. ‘tut-ko-se’ shares the same subject ‘na’ with 2 governor candidates ‘an-a-cwun’ and ‘tal-lye-kass-ci-man’. We consider ‘tut-ko-se’ as the dependent of the verb phrase ‘an-a-cwun ... tal-lye-kass-ci-man’. (the concrete verb phrase is ‘a-i-lul han-pen an-a-cwun hwu tal-lye-kass-ci-man’). So we determine the governor of ‘tut-ko-se’ is the head of the verb phrase, namely ‘tal-lye-kass-ci-man’.
The structural preference rule concerning this is in Figure 8.
4.2
Resolving Ambiguity between Adverbials and Predicates
We use a statistical method to resolve ambiguity between adverbials and predicates. Early work investigated the use of Probabilisitc Context Free Grammars (PCFG), but with rather disappointing results. A simple
If a dependent is a predicate Î If there are the candidates which have the same subject Î The governor of the dependent is the last candidate that has the same subject. Else Î The governor of the dependent is the last candidate of all the governor candidates Figure 8 : The structural preference rule Ⅱ
PCFG’s failing is its lack of sensitivity to lexical information and structural preferences. For example, suppose that a right-branching structure is preferable. Yet two different structures (one is left-branching and the other is right-branching) can have the same context free grammar rules. Research then moved to probabilistic versions of lexicalized grammars (Magerman 95) (Collins Michael 1999). Collins suggested a head-driven approach (1999). He makes a lexicalized tree by modifying the phrase structure tree to include lexical information. He also uses 4 kinds of associated sets of probabilities. We modify and adopt the lexical dependency method among these 4 kinds of probabilities. The characteristics of the statistical method that we propose are as in Figure 9. 1. It is used only to resolve the ambiguity in determining the governor of adverbials. 2. It considers structural information by using a 1-depth dependency tree (considering the parent node and sister nodes of the dependent) and sister nodes include only predicates which indicate distance information. 3. It uses semantic information of the governor. 4-1. If there is some particle in a dependent, use ‘adverbial particle- driven method’ 4-2. Else ‘lexicon-driven method’ Figure 9 : The characteristics of the statistical method
This paper proposes two methods, whether a dependent has some adverbial particles or not. If there is some particle in a dependent, we use ‘adverbial particle-driven method’. If not, we apply ‘lexicon-driven method’ First, we describe the ‘adverbial particle-driven method’. 1) Adverbial particle- driven method
Let us consider the following example. Ku cwung pwu-san-ey sel-lip-toyn kong-cang-i saing-san-cey-phwum-to ta-yang-ha-ko ka-cang khu-ta.
(The factory built in Busan makes a variety of products and is the biggest of its kind) Let’s perform inter-chunk dependency parsing of this example sentence. After determining complements- predicates dependency and adjuncts (predicates) -predicates dependency, we obtain the following imperfect dependency tree. khu-ta kong-chang-i sel-lip-toyn
cwung
ta-yang-ha-ko
ka-cang
saing-sancey-phwum-to pwu-san-ey
ku Figure 10 : The imperfect tree of an example
The nodes that do not have their governors are ‘cwung’ and ‘pwu-san-ey’. They are adverbials. Let us determine the governor of ‘pwu-san-ey’. Dependent : pwu-san-ey Governor candidates : sel-lip-toyn ta-yang-ha-ko khu-ta
Before deciding the governor of adverbials, we first check the cross links to reduce the governor candidates of adverbials. In the case of ‘pwu-san-ey’, however, no governor candidates are unsuitable because no cross links occur between ‘pwu-san-ey’ and governor candidates. The governor candidates of ‘pwu-san-ey’ are ‘sel-lip-toyn’, ‘ta-yang-ha-ko’ and ‘khu-ta’. The dependent ‘pwu-san-ey’ has an adverbial particle ‘ey’. We consider that ‘ey’ has more information than ‘pwu-san’ because ‘ey’ is a locative adveribial particle and some predicates have a tendency to require a certain adverbial particle. We use only ‘ey’ information about the dependent ‘pwu-san-ey’. We construct a 1-depth dependency tree, considering the parent node and sister nodes of the dependent. Sister nodes
include only predicates which indicate the distance information. However, we have sparse data problems. The count(governor ,dependent) may be too low to give a reasonable estimate. Therefore, we use the semantic information of governor candidates for smoothing. Here, semantic information refers to the Kadokawa thesaurus concept number. We can retrieve the semantic information of the governor from the dictionary. Figure 11 shows 1-depth dependency trees of each governor candidate after applying adverbial particle-driven method. The probability of 1-depth dependency tree to give a ‘smooth’ estimate is as follows. governor(sem A) dependent (adverbial particle B)
the sum of λ 1, λ 2, λ 3 is 1, and they are weights that should be chosen by experiment (λ 1 >λ 2 >λ 3). The governor of ‘pwu-san-ey’ becomes ‘sel-lip-toyn’ according to the probability result. 2) Lexicon-driven method Let us determine the governor of the dependent ‘cwung’. ‘cwung’ is adverbial which has no adverbial particles. So we cannot use adverbial particle information of the dependent. Instead, we use lexicon of the dependent. Dependent : cwung Governor candidates : sel-lip-toyn ta-yang-ha-ko khu-ta
First, we check cross links to reduce the governor …… candidates of ‘cwung’. In this example, the governor candidates are not reduced because no cross links occur between ‘cwung’ and governor P ( gov[ semA ], sister1.., sister n | dep[ adv − particleB ]) candidates. The governor candidates of ‘cwung’ count ( dep[ adv − parti .B ], gov[ semA ], sister1.., sister n) are ‘sel-lip-toyn’, ‘ta-yang-ha-ko’ and ‘khu-ta’. = λ1 × count ( dep[ adv − parti .B ]) This method uses lexical information of the count ( dep[ adv − parti .B ], gov[ semA ]) dependent and adopts the distance between + λ2 × count ( dep[ adv − parti .B ]) dependent and governor in the sentence(Collins count ( dep[ adv − parti .B ], gov , sister 1.., sister n) Michael, 1997). The lexicon-driven method + λ3 × count ( dep[ adv − parti .B ]) proposes a new concept of distance, which means the number of predicates between dependent and 1) governor: sel-lip-toyn 2) governor: ta-yang-ha-ko governor candidates. In adverbial particle driven sel-lip-toyn ta-yang-ha-ko method, a sister node is used for distance pwu-san-ey saing-saninformation. However, 1-depth dependency tree pwu-san-ey cey-phwum-to using the lexicon-driven method contains no 3) governor: khu-ta sister nodes because distance information is khu-ta presented separately. Distance also includes the last Part-of-Speech tag of predicates. If the pwu-san-ey ka-cang kong-cang-i predicate of which the last Part-of-Speech tag is ta-yang-ha-ko subordinate conjunctive ending, adnominal After applying adverbial particle driven method ending, or quotative particle, it is part of an embedded sentence. We can ignore the predicate 1) governor: sel-lip-toyn 2) governor: ta-yang-ha-ko sem[sel-lip-toyn] sem[ta-yang-ha-ko] which is part of embedded sentence to obtain a ‘smooth’ estimate. ~ey(adverbial ~ey(adverbial particle) particle) Figure 12 shows 1-depth dependency tree by the lexicon-driven method. The procedure to obtain 3) governor: khu-ta sem[khu-ta] probability of 1-depth dependency tree is as in Figure 13. It gives a ‘smooth’ estimate by ~ey(adverbial Predicate particle) (conjunctive) ignoring predicates which are parts of the embedded sentence. . Figure 11 : 1-depth dependency tree sister 1
sister n
by adverbial particle-driven method
1) governor: sel-lip-toyn
2) governor: ta-yang-ha-ko
sel-lip-toyn chwng
ta-yang-ha-ko
pwu-san -ey
saing-sancey-phwum-to
chwng
structure tree to the dependency tree semi-automatically. The performance of our parser is evaluated using the following measures. Dependency Precision =
3) governor: khu-ta
number of correct dependency relations
khu-ta
number of dependency relatios in proposed parse
chwng kong-cang-i
ka-cang ta-yang-ha-ko
After applying lexicon-driven method 1) governor: sel-lip-toyn governor(dist:0)
2) governor: ta-yang-ha-ko governor(dist:1) chwng
chwng
Dist 1: adnominal
3) governor: khu-ta governor(dist:2) : END Dist 1: adnominal Dist 2: coordinate
chwng
Figure 12 : 1-depth dependency tree by lexicon- driven method
Governor (Dist:n) Dependent (lexicon A) Considering only predicates which are not part of embedded sentence Dist 1: POS C …. Dist m: POS D
Dist 1: POS P …. Dist n: POS Q After applying lexicon-driven method Remove the predicates of which the last POS tag is one of {subordinate conjunctive ending, adnominal ending, quotative particle
Figure 13 : The procedure to obtain probability of 1-depth dependency tree by lexicon- driven method
P( gov[dist : n] | dep[lexicon : A]) count( gov[dist : n], dep[lexicon : A], dist1[POS : P].., dist n[POS : Q]) = λ1 × count(dep[lexicon : A)]) count( gov[dist : m], dep[lexicon : A], dist1[POS : C ].., dist m[POS : D]) + λ2 × count(dep[lexicon : A])
5
Experimental Results
We use the Korean Information Base 97 (KIBS97) tree corpus for our statistical method. This corpus has been parsed using a phrase structure grammar. We converted this phrase
The performance was evaluated by the dependency precision. Because the dependency tree produced has no phrase structure label, the accuracy rate is sufficient for performance estimation (Yoon Juntae 1999). Our parser was compared with the original parser which use no methods to resolve ambiguity in inter-chunk dependency parsing. We assume that the original parser determines the governor of dependent is nearest to the dependent. Since Korean is a head-final language, the last word in a sentence is always a root. Therefore, the number of the dependency relation is (n-1) in a sentence consisting of n eojeols(Yoon Juntae 1999). The number of dependency relation in test data is 1508. After chunking, the number of inter-chunk dependency relation is 972. It proves that determining dependency relation after chunking is simplified. Table 4 shows the experimental result using the method proposed in this paper. The accuracy is much higher than that of the original parser. The overall dependency precision is 88.03 % and the average number of eojeols in a sentence is 8.54 in test data. It needs to experiment using the test data which have so many predicates. To improve precision, we should complement structural preference rule Ⅰ and make a stronger statistical method for adverbials.
6
Conclusion
In this paper, we presented a method to resolve ambiguity in inter-chunk dependency parsing. This method is a combination of a rule-based approach and a statistical method. We use valency information in the dictionary to determine the dependency relation between complements and predicates. We also use structural preference rule to complement the valency information. Structural preference is an important clue to resolve ambiguity in parsing.
The process to disambigute the dependency relation between adjuncts and predicates is divided into two steps. The first disambiguates the dependency between predicates and predicates, and the second is between adverbials and predicates. Table 1 : Training data
number of sentences (number of eojeols)/(a sentence)
22483 11
Table 2 : Test data
number of sentences (number of eojeols)/(a sentence) Number of dependency relation before chunking Number of inter-chunk dependency relation after chunking
200 8.54 1508 972
Table 3 : Remaining ambiguities after applying valency information
Number of inter-chunk dependency 259 relation that has ambiguity. Using Using Using preference preference statistical Total method rule Ⅰ rule Ⅱ 81 53 125 259 Table 4 : Resolving ambiguity using our parser
Using structural preference rule Ⅰ Correct Wrong Total Accuracy 66 15 81 81.48 % Using structural preference rule Ⅱ Correct Wrong Total Accuracy 49 4 53 92.45 % Using statistical method Correct Wrong Total Accuracy 113 12 125 90.4 %
Table 5 : Resolving ambiguity using original parser
determine the governor of dependent as the nearest the dependent Correct Wrong Total Accuracy 156 103 259 60.23 % Table 6 :Dependency precision
Dependency precision
88.03 %
The statistical method that this paper proposes has the following characteristics. First, it uses structural information by considering a 1-depth tree. Second, it uses semantic information of the governor for smoothing. Third, it uses lexical information or adverbial particle information. The experimental results show that the parser using this hybrid method performs with 85.10 % accuracy. In future work, we should construct a more parsed tree corpus. There are adverbials of which the governor is a noun or adverbial, not a predicate. A new method is required to identify the governor of these adverbials.
References Abney, Steven. 1991. Parsing by chunks. In Berwick, Abney, and Tenny, editors, Principle-Based Parsing. Boston:Kluwer Academic Publishers. Allen, James. 1987. Natural Language Understanding, The Benjamin/Cummings Publishing Company, Menlo Park. Collins, Michael. 1997. Three Generative, Lexicalised Models for Statistical Parsing. In Proceedings of the 35th Annual Meeting of the ACL, and 8th Conference of the EACL, Madrid, Spain. ACL. pp 16~23. Collins, Michael. 1999. Head driven statistical models for natural language parsing. Ph. D. thesis, University of Pennsylvania. Kim Miyoung. 2000. Text Chunking by Rule and Lexical Information. In proceedings of the 12th Hangul and Korean Information Processing Conference, Chonju, Korea. pp 103~109. Magerman, David. 1995. Natural Language Parsing as Statistical Pattern Recognition. Ph.D. thesis, Stanford University. Na Dong Ryul. 1994. Investigationon parsing Korean. Korea Information Science Society Review, 12(8). Yoon Juntae. 1999. Three Types of Chunking in Korean and Dependency Analysis Based on Lexical Association. In Proceedings of the 18th international conference on computer processing of oriental languages. pp 59~66.