A Knowledge Based Approach to Identification of Serial Verb Construction in 
Chinese-to-Korean Machine Translation System 
 
Dong-il Kim
∗
, Zheng-Cui, Jinji-Li
∗∗
, Jong-Hyeok Lee 
Department Computer Science and Engineering, Electrical and Computer Engineering Division, 
Pohang University of Science and Technology (POSTECH)   
and Advanced Information Technology Research Center (AlTrc) 
San 31 Hyoja Dong, Pohang, 790-784, Korea 
E-mail: {dongil, cuizheng, ljj,jhlee}@postech.ac.kr 
 
                                                      
∗
 Also an assistant professor at Yanbian University of Science  
& Technology (YUST) Yanji, Jilin, China. 
∗∗
 Also a lecturer at YUST 
Abstraction 
In Chinese language processing, the 
recognition and analysis for serial verb 
constructions (SVCs) is a fascinating research 
topic. Chinese language researchers each may 
have a different definition and interpretation of 
SVC since the structure of SVC makes Chinese 
unique to other languages and contains complex 
semantic and pragmatic information. This paper 
proposes a formal definition of SVC and a 
knowledge based approach for the recognition of 
SVCs, which is adopted in TOTAL-CK, a 
transfer-based MT system from Chinese to 
Korean. The recognition process is carried out in 
two stages: the analysis stage classifies SVCs 
into general categories, and the transfer stage 
performs further classification for Korean 
transfer. Some evaluation result for each stage 
was also given with statistics of each category of 
SVCs 
Introduction 
Many Chinese language researchers have paid 
special attention to the so-called ‘serial verb 
constructions (SVCs)’, where two or more 
semantically or pragmatically related verb 
phrases or clauses are juxtaposed together 
without any functional marker. Because of 
different definitions and interpretations of SVCs 
among researchers, their categorizations differ 
according to researchers’ viewpoints. 
In a Chinese to Korean machine translation, 
the hidden relation of the serial verbs should be 
expressed with some function words from the 
target language viewpoint. Moreover, the 
conceptual scope of these function words is 
different from the scope of SVC categorizations 
that are classified based on the viewpoint of the 
Chinese language itself. 
 In this paper, we proposes a different 
categorization of SVCs defined by the 
contrastive analysis of the two languages, and 
also an SVC identification method that is 
adopted in a Chinese-to-Korean MT system, 
TOTAL-CK. The TOTAL (Translator Of Three 
Asian Languages: Chinese, Korean and 
Japanese) project has been conducted under a 
hybrid strategy with transfer-based and 
example-based engines.  
1. Language Characteristics between 
Chinese and Korean 
In this section, some contrastive analyses of 
the two languages are introduced for better 
understanding of an SVC sentence. Since 
Chinese is an isolating language, morphological 
or syntactic markers rarely appear in a sentence, 
while in an agglutinative language such as 
Korean, these functional markers are not an 
optional unit but an obligatory unit in a sentence.  
An example is given in (1). Notice that the 
Korean alphabets are written with Yale 
Romanization in this paper. 
 
(1) 他  开门   进去。 (ta kai-men jin-qu) 
Ku-nun  mwun-ul  yel-ko   duleka-nta. 
He-NOM  door-ACC open-CON get in-PRENT-DEC. 
He opens the door and enters (the room). 
In the Korean sentence, ko is a connective 
particle, and also nun, ul, and nta denote a topic 
auxiliary particle, an object case particle and 
declarative terminative ending, respectively. All 
these functional markers should be decided in 
the Korean transfer stage. Specifically, we 
require a process to select one from the possible 
conjugational markers when a Chinese SVC 
sentence is transferred to its Korean counterpart. 
2. Related Works 
A SVC is studied among several researchers 
as different names. But the general syntactic 
form is (NP) V1 (NP) V2 (NP)
1
. The variance of 
definition for SVC comes from the different 
scope of interpretation for the sentence pattern. 
 We will introduce three typical researches to 
clearly outline our definition of SVC.  The 
narrowest view of scope is suggested in (Lü, 
1953). In his interpretation, V1 and V2 have the 
same subject and should be not coordinative, but 
it is difficult to decide which one is main or 
additional.  Zhu (Zhu, 1981) includes all cases 
of Lü’s and the possibility of adding an adjective 
to substitute for the second verb position. He 
also includes the case where an additional verb 
and a main verb are used, such as  V+ 着  
expression in V1 position, which indicates that 
V1 is additional and V2 is main. The broadest 
scope is proposed in (Li & Thompson, 1981). 
According to his interpretation, an SVC includes 
not only all the patterns noted above but also a 
pivot construction, a subject/object clause, and a 
coordinate clause, but excludes the pattern with 
an adjective in the V2 position. In this paper, the 
scope of SVC is almost same as Li’s but the 
classification of SVCs differs slightly, detailing 
the categorization in chapter 4.  
 A few computational solutions to identifying 
SVCs have been proposed by some researchers. 
A formal description is shown in  (Chan, 1998) 
using time lapse notation and the related 
definition. However, her method makes it 
difficult to computationally detect SVCs without 
the resources containing the deep level of 
analysis of each lexical, which is not obtainable 
in the current stage of language processing. 
                                                      
1
 V1 : first verb, V2 : second verb, NP : noun phrase.  
3. Overview of TOTAL-CK System 
Architecture 
As a typical transfer system, TOTAL-CK 
consists of three parts: Chinese analysis, 
dependency tree transfer, and Korean generation.  
The system architecture of TOTAL-CK is shown 
in figure 1. The design principles and the detail 
descriptions are given in (Kim et al., 2002). 
 
 
 
 
 
 
 
 
 
 
 
 
Figure 1: TOTAL-CK system architecture 
4. Classification of Serial Verb 
Construction 
In the previous chapter, we mentioned the 
syntactic format of SVCs which is NP V1 (NP) 
V2 (NP) and the different scope of definition of 
SVCs by the Chinese language researches. To 
outline the scope of SVCs, we define SVCs in 
terms of dependency relation such that V1 is the 
head of V2, or V2 is the head of V1. It is 
formally defined as follows: 
Definition 1 
Let N represent a set of nodes in a dependency 
tree, and W a set of words. Further Let V be a 
set of verbs, and P be a set of all parts of speech 
in Chinese. Then the functions: head, nw, and 
npos, are defined as below: 
head(n) =hn where n∈N and hn is the head of n 
nw(n) = w where n ∈N and  w ∈ W 
npos(n) = np where n ∈N and  np ∈P  
A definition of SVC is: 
Given a node n such that npos(n) ∈  V and 
Head(n) = hn, 
If and only if npos(hn) ∈ V and hn is the head of 
a given sentence then the sentence is a SVC.  
The three sentences from the top of table 1 
satisfy the given condition. Also the head of 
the node is the sentence head, thus these must 
be SVCs. For the  last sentence, nw(n) is 接
入, and nw(Head(n)) is 总数  whose the POS  
is not verb and also whose the node is not the 
sentence head. Thus it is not an SVC. 
Sentence nw nwh SH SVC
他开门进去。 进  开  Yes Yes
我没想到你住在北京。  住  想  Yes Yes
在这里停车犯法。  停  犯  Yes Yes
各接入网络的总数已
经超过1000个。  
接入  总数  No No
Table 1: Example of Testing SVC 
Where nw: nw(n); nwh: nw(head(n)); SH :testing if 
head(n) is the sentence head ; n is a given node. 
 
Our definition is employed to recognize a 
SVC in the Chinese analysis stage. First we 
describe the classification that is used in the 
Chinese analysis stage.  
4.1 Categories in Chinese Analysis Stage 
All dependency relations, which are detected 
by the above definition, are classified into five 
categories: separate events, object, subject, 
pivotal construction and descriptive clauses, 
based on the classification of Li (Li & 
Thompson, 1981). 
4.1.1 Separate Events  
The serial verb patterns classified by most 
researchers belong to this group where switching 
V1 to V2 provides us a different meaning. In 
addition, we add the case where transposing V1 
to V2 provides us the same meaning in this 
group. 
4.1.2 Object 
If V2 is the main verb in an object clause or a 
object phrase then it belongs to this group. 
4.1.3 Subject 
If V1 is the main verb in the subject clause or 
subject phrase, it is assigned to this group. 
4.1.4 Pivot 
If the noun phrase between V1 and V2 is the 
object of V1 and the subject of V2, then it is a 
pivot construction. 
4.1.5 Descriptive  
 If V2 describes the noun phrase between V1 
and V2, then it is a descriptive SVC. 
 All categories of SVCs are shown in Table 2 
The corresponding Chinese dependency 
relations to object, subject and pivot 
constructions also appear in the some research in 
Chinese language processing (Zhou & Huang, 
1994) but the other two are not shown due to 
their different viewpoints. 
The descriptive construction is directly able to 
be one-to-one mapped to the Korean counterpart. 
However the separate event SVCs should be 
further classified for Korean transfer since the 
separate event SVC is possibly mapped into 
sentences with several different Korean 
conjunctional particles.  Thus, it is touched in 
the transfer stage. 
Category Example 
Separate Event
我买票进去; 他在饭店吃饭喝茶。 
Object 
我没想到你住在北京。 
Subject 
在这里停车犯法。 
Pivot 
我们让他去北京 。  
Descriptive 我有一个姐姐喜欢看书。 
Table 2: Examples of SVC category 
4.2 Subcategories in Transfer Stage 
The separate event SVC for each sense of 
Korean conjunctional particles is classified into 
the following subcategories: restrictive, 
quasi-coordinative, simultaneous, transitional, 
and circumstantial by the Korean language 
viewpoint. 
4.2.1 Restrictive
 
 
 The action of V2 is performed under the 
restriction given by V1.  There are different 
types of restriction, such as space, group-related, 
causal, and instrumental. The examples are 
presented in table 3. 
Sentence V1 V2 R 
type 
我的妹妹今天离开北京前往
汉城
2
。  
离开  前往  space 
他代表山西省出席了座谈会。  代表  出席  group
我投票赞成第一个人。  投票  赞成  causal
这个图书馆利用计算机进行
图书借阅管理。  
利用  进行  tool 
Table 3: Examples of Restrictive Separate Events 
 
                                                      
2
 The sentence can also be interpreted as purposive 
separate events. But it is included into a restrictive separate 
event SVC because it is impossible to detect the differences 
between restrictive and purposive, as this requires 
pragmatic level information 
 
4.2.2 Quasi-Coordinative 
In quasi-coordinative, two different cases 
exist. First, transposing V1 to V2 never causes a 
meaning shift of the sentence, named alternative. 
The other is that V1 and V2 are only 
sequentially related, called consecutive. 
4.2.3. Simultaneous 
In a simultaneous case, V1 and V2 occur at the 
same time. 
4.2.4 Transitional 
 If the action of V1 is interrupted by the action 
V2, then it is transitional. 
4.2.5 Circumstantial 
When V2 occurs on the condition of the action 
of V1, then it is classified as a circumstantial 
case.  
The examples for rests of the separate event 
are given in Table 4. 
Type Example 
Q-Coordinative 他在饭店吃饭喝茶。 (alternative)  
他买票进去。 (consecutive) 
Simultaneous  nullnullnullnull歌 。  
Transitional 我的弟弟开车出事了。  
Circumstantial 不买别进。  
Table 4: Examples of Separate Events 
 
In restrictive, quasi-coordinate, simultaneous, 
transitional, and circumstantial separate event 
SVC Chinese sentences, all the above verbs are 
mapped into the corresponding Korean verb 
followed by the Korean conjunctional particle 
‘se’, ‘ko’, ‘un-chay-lo’, ‘taka’ and ‘myen’, 
respectively. 
5. Identification of SVCs 
 To recognize SVCs, we divide the identifying 
process into two stages. The general categories 
of SVCs are able to be found at the analysis 
stage and the subcategories of a separated event 
SVC are detected in the transfer stage.  
5.1 Analysis Stage 
To recognize the five general categories of 
SVCs, two resources are used: one is the 
Grammatical Knowledge Base of Contemporary 
Chinese (GKBCC) and the other is a verb list 
with valency information (VLVI) (Zhu et al., 
1995).  Checking a verb in GKBCC allows us 
to simply detect a pivot SVC. The remainders of 
the other types of SVCs should be carefully 
handled. There are two possible ambiguous 
structures of SVCs 
Case 1 : NP V1 V2 (NP2) 
Case 2 : NP V1 NP1 V2 (NP2) 
Where NP, NP1 and NP2 are noun phrases. 
The algorithm for each case is illustrated in 
figure 2 and figure 3. In Figure 3, the test ‘V1 
takes NP & VP’  means that the verb 偷听  can 
have a noun phrase or an object clause as an 
object. The test, ‘satisfy valency’  denotes that  
the second verb  喜欢  takes a human subject, 
and  外国人  can be the subject of the verb 喜欢 , 
thus it is classified as an object case. For the 
other sentence, since 公园  cannot be the subject 
of the verb 锻炼 , it is determined as a subject 
case. 
 
 
 
SVC sentence 
 
 
 
Y 
Object SVC 
Ex) 他希望去北京 。  
 
 an object exists?
 
N 
 
Separate Event SVC
Ex) 我去问刘厂长。  
 
 
 
Figure 2: Algorithm of Detecting SVC for Case 1 
 
 
 
 
Subject SVC 
Ex) 厂长 偷听 私 人电话 违反
了国家法律。  
 
SVC sentence 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   Figure 3: Algorithm of Detecting SVC for Case 2 
Object SVC 
Ex) 厂长偷听属下打电话 。
Y V1 takes 
NP &VP 
N 
N 
Y 
 
V2 takes VP
as Sub. ?
N 
Separate Events SVC 
Ex) 我的爷爷来公园锻炼 身 。
Y 
Descriptive SVC 
Ex) 他碰到了一个外国人  
喜欢中国。  
Subject SVC 
Ex) 克林顿总统访问  
中国推动 双边的 关系的发 展 。
 
V2 takes VP  
as Sub. ? 
N
Satisfy 
Valency ? 
Y 
 
5.2 Transfer Stage 
The simultaneous separate events is easily 
recognized by the lexical (着 ) attached to the first 
verb. Also, we use a simple heuristic to detect 
the circumstantial separate events with the 
lexical pattern information.  
The resource used in this stage is a Chinese 
thesaurus called Tongyi-ci-cilin (Mei, 1983). 
With the thesaurus the remainders of separate 
event SVCs are processed with great care. If V2 
is related to the interrupt concept then the 
transitional separate events are assigned. The 
most difficult and frequently occurring cases are 
the restrictive separate events and 
quasi-coordinative separate event.  
The key idea of using the thesaurus is based 
on the observation that the verb V2, if restricted 
by V1 makes it possible that the concept of V2 
will also be restricted by the concept of V1. To 
complete the solution, we first define the 
relations: RSTV, RSTL and RSTM as follows:  
Definition 2 
We define the relations: RSTV, RSTL, and 
RSTM, as follows: 
RSTV= {(V1,V2)  where V1 and V2 are the 
first verb and second verb in a given SVC 
sentence and V2 is semantically restricted by 
V1 : (V1,V2)  (V2,V1)} ≠
RSTL= {(CL1,CL2) where CL1 and CL2 are 
the low level concept of the first verb and the 
low level concept 
3
 of second verb in the 
Chinese thesaurus, respectively, and CL2 is 
semantically restricted by CL1 : (CL1,CL2) ≠  
(CL2,CL1)} 
RSTM= {(CM1, CM2) where CM1 and CM2 
are the middle level concept of the first verb and 
the middle level concept of second verb in the 
Chinese thesaurus, respectively, and ML2 is 
semantically restricted by ML1 : (ML1,ML2) ≠  
(ML2, ML1)} 
The relations RSTV, RSTL, and RSTM are 
not symmetric and not reflexive. Based on the 
definition we derive the following heuristics: 
 if (V1,V2)∈ RSRV then (CL1,CL2) ∈RSTL 
But if (V1,V2) ∈ RSTV then not always 
(CM1,CM2) ∈ RSTM. 
                                                      
3
 The thesaurus consists of three levels of hierarchy. For 
example, H, Hj, and Hj20 correspond to the one of highest 
concept, the next narrow term called middle-level concept 
and the narrowest term called low-level concept, 
respectively. 
All three examples from the top of table 5 
satisfy the condition that, if (V1, V2)  RSTV 
then (CL1,CL2) 
∈
∈RSTL and (CM1,CM2) ∈ 
RSTM. If the condition is always true, then we 
use the middle-level concept relation for 
detecting a restrictive separate event in order to 
increase the applicability of our rules. Also, the 
data structure of RSTM is easily represented 
with an adjacent matrix with the size of 21*21
 4
 
(Sahni, 1998) where the matrix M is a square 
matrix, whose column and row are the 
middle-level concept, and if M(i,j) = 1 then 
concept j is semantically restricted by concept i, 
otherwise (i,j) ∉RSTM. 
RSTV RSTL RSTM 
Example 
V1 V2 CL1 CL2 CM1 CM2
他参加政府会议 公
开批评了 我的 谈
话。
 
 
参
加
批
评
Hj20 Hi21 Hj Hi
国家主席江泽民
出席讲话。  
出
席
讲
话
Hj20 Hj12 Hj Hi
他给华大使带来
一个大花篮表示 祝
贺。  
带
来
表
示
Hj36 Hj14 Hj Hi
 
 
他代表山西省出
席了座谈会。  
代
表
出
席
Hi17 Hj20 Hi Hj
Table 5: Example of RSTV, RSTL and RSTM 
 
However, the last example reveals that the 
condition is not always true since we have the 
result, both (Hi,Hj) and (Hj,Hi) ∈  RSTM.    
Thus, it violates the definition of RSTM. Hence, 
we may not directly use the middle-level concept 
adjacent matrix and the size of the low-level 
concept matrix is too large to be used.
5
  
 We come up with a solution of a frame with 
multi level concepts. The frame consists of three 
parts: the middle-level concept adjacent matrix, 
the low-level concept adjacent lists and the 
collocation serial verb list for detecting a serial 
verb that always appears together. 
Our solution is that the exceptional cases are 
covered by either the collocation verb lists or the 
low-level concept adjacent list. The remaining 
frequently occurring cases are captured by the 
middle-level adjacent matrix. This leads to the 
sparse matrix of the low-level concept which 
                                                      
4
 The number of verbs related middle-level concept in the 
Chinese thesaurus is 21. 
5
 The number of verbs related low-level concepts in the 
Chinese thesaurus is about 500. 
causes the adaptation of adjacent lists rather than 
an adjacent matrix for the low-level concepts. 
The order of searching the frame is the 
collocation list, the low-level concept list and the 
middle-level concept matrix. In the collocation 
list, if V1 and V2 belongs to the collocation list 
of the restrictive separate events, such as 捉拿归
案  or the one of quasi-coordinative, such as 立案
侦察  then the sentence is assigned to a restrictive 
case or a quasi-coordinative case, respectively. 
In the low-level concept lists and the 
middle-level concept matrix, if matching 
succeeds, which means that V2 is semantically 
restricted by V1, then a restrictive case is 
assigned; otherwise, a quasi-coordinate case is 
detected
6
. The detailed process for identifying 
the subcategories of separate events is shown in 
figure 4. 
6. Evaluation 
We randomly selected 1000 SVC sentences 
from 1998 people’s daily newspapers. The 
number of verbs in the sentence is two since our 
dependency parser is still being improved to 
detect the sentences with multiple embedding 
clauses. In table 6, the distribution of each type 
of SVC and the precision are shown. 
Type Frequency Percentage 
Separate events 402 40.2 
Object 479 47.9 
Subject 31 3.1
Pivot 39 3.9 
Descriptive 1 0.1 
Error 56 5.6
(Presicion:94.4%)
Total 1000 100 
Table 6: Distribution of Categories of SVC 
 
The precision is 94.4% and some of the errors 
occur from the tagger, thus some sentences are 
not SVCs. The rest of the errors result from 
missing information in the knowledge bases:  
                                                      
6
 For a sentence 国 家 主 席江泽 民出席 讲话  where the 
relation (Hj20,Hj12) is not in the low-level adjacent list, but 
(Hj,Hi) is 1 in the middle-level matrix, it is assigned to the 
restrictive case, while for the sentence 他代表山 西省出席 了
座谈会  where (Hi17,Hj20) is in the low-level adjacent list, 
thus searching is stopped, it is assigned as a restrictive case. 
A sententence 他在饭店吃饭喝茶  do not satisfy all 
conditions, thus it is detected as Quasi-Coordinate. 
 
 Separate Event SVC
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Figure 4: Detection Algorithm of Subcategory of 
Separate Event SVC. 
 
GKBCC and VLVI. We need the complete list 
of verbs, which has a clause as a subject. These 
verbs in the list will be gradually collected in 
future works. 
The evaluation table for the separate event 
SVCs is provided in Table 7. 
Type Frequency Percentage 
Restrictive 153 38.5 
Quasi-coordinative 184 45.7
Simultaneous 33 8.2 
Transitional 3 0.7
Circumstantial 12 2.9 
Error 19 4.7
(Precision:95.3%)
Total 402 100 
Table 7: Result of Separate Event 
 
N 
N
 
V2 is Interrupt 
Concept ? 
 
V1 & V2 in middle 
-level concept? 
 
N 
Restrictive 
Separate Events
Quasi-Coordinate Separate
Events 
N 
 
V1 & V2 in low-  
level concept? 
N 
N 
 
In 
Restrictive ?
Y 
Y 
Y 
Y 
V1 & V2 in 
Collocation list? 
 
In Quasi 
Coordinate?
Y 
Y 
Simultaneous 
Separate Events
 V1 with 着  ?
N 
Circumstantial
Separate Events
Y 
 
Match lexical
Pattern ? 
N 
Transitional 
Separate Events
Y 
The precision of identifying the category of 
separate event is 95.3%. The errors resulted from 
a circumstantial case since our heuristics is too 
restrictive to detect all cases, thus, it might be 
revised further, and since the low-level concept 
lists are not completed.  The low-level concept 
lists will be continuously updated for increasing 
coverage in the tuning stage of the machine 
translation system. 
Table 8 shows the distribution of the 
subcategory of restrictive separate events for 
Korean transfer. 
Type Frequency Percentage
Space 86 56.2 
Group-related 38 24.8 
Causal 17 11.1
Instrumental 12 7.9 
Total 153 100
Table 8: Category of Restricted Separated Event 
 
In table 9, the frequency for each type of 
accessed resource is listed. Notice that most 
restrictive separate event SVCs are recognized in 
the middle-level matrix. The two cases in 
collocation are all the case of quasi-coordinative, 
thus, the total number is greater than 153. 
Type of accessed 
resource 
Frequency Percentage
Middle-level matrix 121 78.0 
Low-level list 32 20.6 
Collocation list 2 1.29 
Total 155 100 
Table 9: Access Frequencies for Resource Type 
 
 
 
 
 
 
 
 
Figure 5: Demo system of TOTAL-CK 
 
In figure 5, a demo system of TOTAL-CK is 
illustrated. For a given Chinese SVC sentence 
displayed in the top position of the right-most 
window, the corresponding Korean sentence is 
followed in the next row.  The tagged results, 
the segment of chunking, and the Chinese 
dependency tree with indentation are shown in 
each window from left to right.  
Conclusion and Future work 
In this paper, we formally define serial verb 
constructions, and classified the SVC into 
several categories. These categories are related 
to the analysis stage and the transfer stage of 
TALK-CK. We provided a resolution algorithm 
detecting SVCs in each step. Finally, at each 
stage, a promising experimental result is shown.  
Further research must help to better resolve 
the conditional separate event SVC and 
purposive separate event SVC.  
Acknowledgements 
This work was supported by the Korea 
Science and Engineering Foundation 
(KOSEF)  through the Advanced Information 
Technology Research Center(AITrc). 

References  
Chan  Y. W. (1998) Formal Criteria for interpreting 
Chinese Serial Verb Constructions. 
Communications of COLIPS 8(1) ,pp.13-29. 
Kim D.I., Cui Z, Li J.J. and Lee J.H. (2002). 
Resolving Structural Transfer Ambiguity in 
Chinese-to Korean Machine Translation. 2002 
International Conference on Chinese Language 
Computing, Taichung, Taiwan. 
Li C  N. , Thompson S A. (1981), Mandarin 
Chinese: A functional reference grammar. 
University of California Press, USA. 
Lü S. X.(1953) Yufa Xuexi (The Study of Chinese 
Grmmar). Bejing, Zhungguo Qingnian  Press. 
Mei J.J.(1983) Chinese Thesaurus 
(Tong-Yi-Ci-Ci-lin). Shanghai Cishu Press. 1983. 
Sahni, S.(1998) Data structures, algorithms, and 
applications in C++. Boston McGraw-Hill, USA.  
Zhou M and Huang C. (1994) Approach to the 
Chinese Dependency Formalism for the Tagging of 
Corpus. Journal of Chinese Information Processing, 
8/3, pp. 35-52, 1994 
Zhu D.X.(1981), Yufa Jiangyi (Lectures on Chinese 
Grammar),Bejing, Xiangwu Press. 
Zhu X. F, Yu S. W and Wang H. (1995) The 
Development of Contemporary Chinese 
Grammatical Knowledge Base and its Applications. 
Communications of COLIPS, 5/1-2 
