Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, pages 33–39,
Sydney, July 2006. c©2006 Association for Computational Linguistics
Semantic Analysis of Chinese Garden-Path Sentences
Yaohong Jin
Institute of Acoustics, Chinese Academy of Sciences, Beijing 100080, China
jinyaohong@hotmail.com
Abstract
This paper presents a semantic model for
Chinese garden-path sentences. Based on
the Sentence Degeneration model of
HNC theory, a garden-path can arise
from two types of ambiguities: SD type
ambiguity and NP allocated ambiguity.
This paper provides an approach to proc-
ess garden-paths, in which ambiguity de-
tection and analysis take the place of re-
vision. The performance of the approach
is evaluated on a small manually anno-
tated test set. The results show that our
algorithm can analyze Chinese gar-
den-path sentences effectively.
1 Introduction
A characteristic of garden-path sentences is that
they contain a temporarily ambiguous verb
structure, in which a participle is confused with
the main verb of the sentence. For example, con-
sider the sentence While Anna dressed the baby
spit up on the bed. Initially the baby is assumed
to be the object of dressed, but when spit up is
encountered, some sort of error arises. This ini-
tial process, then, must be somehow revised.
This paper models the phenomenon of gar-
den-path sentences in Chinese and addresses the
mechanisms of semantic analysis. Let v1 be the
first verb in the sentence and v2 the second verb.
Modeling the garden-path that arising from two
verbs like v1 and v2 will the focus of this paper.
Models of reanalysis, which concern the syn-
tactic relationships between the error signal and
the head of the phrase that has been misanalyzed
(Frazier, 1998), attempt to explain how the revi-
sions take place. However, for Chinese gar-
den-path sentence analysis, the syntactic rela-
tionship is not enough because the same
syntactic relationship can have different semantic
interpretations. For example,咬死猎人的狗
which is temporarily ambiguous in Chinese has
different interpretations in the following two
sentences. In the first sentence狗(dog) is the
subject of咬死(killed), and in the second it is
the object of咬死(killed).
(1)咬死猎人的狗逃跑了(The dog, which
killed the hunter, had run away).
(2)咬死猎人的狗是熊逃跑的唯一出路(It
is the only way for the bear to run away that
killed the hunter’s dog).
So, semantic analysis is important for Chinese
garden-path sentences. In this paper, garden-path
sentences will be modeled using the Sentence
Degeneration model (SD) of the Hierarchical
Network of Concepts theory (HNC) (Huang,
1998; 2004). Furthermore, our analysis algorithm,
in which ambiguity analysis takes the place of a
revision process, is introduced. We evaluated the
model and the algorithm using a small sentence
set with grammatical structures like
NP1+V1+NP2+v2+NP3. The experiment results
show that our algorithm can efficiently process
Chinese garden-path sentences.
In the rest of this paper: Section 2 discusses
previous work. Section 3 gives a detailed defini-
tion of the Sentence Degeneration model. Sec-
tion 4 describes in detail the Semantic Model of
Chinese garden-path sentences. Section 5 de-
scribes the algorithm and section 6 gives evalua-
tion results. Section 7 presents our conclusions
and future work.
2 Previous Work
The phenomenon of garden-path sentences has
attracted a lot of attention in the research com-
33
munities of psycholinguistics and computational
linguistics. The goal of this research is to dis-
cover how people understand garden-path sen-
tences and how to analyze them automatically.
In English, garden-path sentences always in-
volve a subordinate clause and a main clause to-
gether with an NP that attaches initially to the
former but ultimately to the latter (Karl and Fer-
nanda, 2003). This NP is the point of misunder-
standing and the verb after the NP is always the
error signal. Models of reanalysis are aimed at
describing and motivating the mechanisms used
by the sentence comprehension system to detect
errors, deduce useful information about the na-
ture of the necessary repair of those errors, and
ultimately create a successful analysis (Ferreira
and Christianson, 2001). Fodor and Inoue(1998)
proposed the principles of Attach Anyway and
Adjust to explain how reanalysis processes oper-
ate. Ferreira and Christianson(2001) stated that
Reflexive Absolute Transitive (RAT) verbs, such
as wash, bathe, shave, scratch, groom, and so on,
are likely to give rise to garden-paths. Michael J.
Pazzani(1984) demonstrated how to reanalyze
one type of garden-path sentence that arises from
a passive participle and a main verb conflicting.
However Ferreira and Henderson(2001) demon-
strated that reanalysis is more difficult when the
head of the misanalyzed phrase (baby in the baby
that was small and cute) is distant from the error
signal.
In Chinese, there has been little research that
directly addresses the problem of garden-paths.
Zhiwei Feng(2003) interpreted the temporarily
ambiguous verb structure in a garden-path in two
ways; one is as a subordinate clause (MV), the
other is a Reduced Relative (RR). He defined
Garden Path Degree (GPD) as MV/RR. He stud-
ied some types of temporarily ambiguous verb
structures such as NP1+VP+NP2+de+NP3,
VP+NP1+de+NP2, V+Adj+de+N and
V+V+de+N, and stated that when GPD is larger
than 3, the temporarily ambiguous verb structure
may give rise to a garden-path. Moreover he
used the Earley algorithm to process garden-path
sentences.
3 Sentence Degeneration model (SD)
The Sentence Degeneration model, which is one
model of the Hierarchical Network of Concepts
theory (HNC), focuses on representing the sub-
ordinate clause in a sentence. The theory of the
Hierarchical Network of Concepts (HNC theory),
founded by Prof. Zengyang Huang of the Insti-
tute of Acoustics, Chinese Academy of Sciences,
is a theoretical approach to Natural Language
Processing (NLP). The objective of HNC is to
establish natural language representation patterns
based on the association veins of concepts, which
can simulate the language perception process of
the human brain and can be applied to computa-
tional Natural Language Understanding.
Sentence Degeneration (SD) represents the
semantic patterns of the subordinate clause in a
sentence. There are three types of SD: prototype
SD, key-element SD, and packed SD.
In Prototype SD a subordinate clause wholly
acts as one role of the other sentence without any
alteration. For example,中国加入世界贸易组
织(China joined the WTO) is a complete sen-
tence. However in sentence (3) this sentence acts
as the subject of促进(accelerate). Unlike Eng-
lish, in Chinese there is no relative pronoun, such
as that or which, to indicate that this is a subor-
dinate clause. This phenomenon is named Proto-
type SD.
(3)中国加入世界贸易组织会促进全球经济
的发展(That China joined the WTO will accel-
erate the development of global economics.).
Key-element SD involves an NP which se-
mantically is an attributive clause. For example,
although in sentence (4)加入世界贸易组织的
中国is an NP, it can be transformed from the
sentence中国加入世界贸易组织by moving
the subject中国to the tail and adding the Chi-
nese word的(of or ‘s) in front of it. We look at
this NP as a specific attributive clause1 in Chi-
nese, and look at中国as the core concept of
this clause. Because the core concept of this
clause is the subject, which is the key element,
this phenomenon is called key-element SD. Be-
sides the subject, the object and the verb of the
sentence can be the core of key-element SD. For
example, in sentence (5)中国对世界经济的影
响is one key-element SD transformed from the
sentence中国影响世界经济, and the verb影
响is its core.
(4)加入世界贸易组织的中国将严格遵守贸
易规则(China, which joined WTO, will strictly
confirm the world trade rule.)
(5)这一切体现了中国对世界经济的影响
(Everything of all reflected the influence that
China economics impacts on the world).
Packed SD is also an NP in which the attrib-
1 This NP has to be translated as an attributive clause using
which in English.
34
uter is a prototype SD or key-element SD. For
example, in sentence (6) and (7) both noun
phrases中国加入世界贸易组织的消息and
中国对世界经济的影响程度are Packed SD’s.
Moreover, the attributer of消息is中国加入世
界贸易组织which is a prototype SD, and the
attributer of程度is中国对世界经济的影响
which is a key-element SD. The words消息
and程度are called packed words.
(6)中国加入世界贸易组织的消息令人激动
(The news that China joined WTO is exciting.)
(7)中国对世界经济的影响程度将越来越大
(The degree of influence that Chinese economics
impacts on the world is deeper and deeper.)
Let ElJ be the semantic structure of the subor-
dinate clause, GBKi be the subject/object, and El
be the verb of the clause. The semantic pattern of
the clause can be given as
ElJ=GBK1+El+GBK2+GBK3, where GBK2 and
GBK3 can be absent and the position of GBKi
can be changed. Suppose ElJ-GBKi stands for
the action of subtracting the GBKi from ElJ,
ElJ-El stands for subtracting the El. The seman-
tic patterns of SD can be given as follows:
1. ElJ. It means that ElJ is a prototype SD.
2. (ElJ-GBKi)+的+GBKi. It means that this
key-element SD can be transformed from the
clause ElJ by moving GBKi to the tail and
adding the Chinese word的in front of
GBKi.
3. (ElJ-El)+的+El. It means that this
key-element SD can be transformed from the
clause ElJ by moving El to the tail and add-
ing the Chinese word的in front of El. Al-
though this El looks like a noun because
there is Chinese word的in front of it, it is
regarded as a verb when restored back to the
ElJ.
4. a prototype SD or key-element SD+{的
}+noun. It means the three patterns above
can serve as the attributer of the packed SD.
Although the key-element SD and packed SD
look like NP’s in Chinese, they need to be trans-
formed back into clauses during semantic analy-
sis. It means that in patterns 2 and 3 the GBKi
and El have to be restored into ElJ. This is why
we named these phenomena Sentence Degenera-
tion. Moreover, in patterns 2 and 3, the Chinese
word的is necessary to indicate the transforma-
tion, and we call it a sign of SD.
Therefore, if an NP or other structure includes
a verb and the Chinese word的, it has to be
analyzed as one type of SD. These semantic pat-
terns of SD are useful for describing the inter-
pretation of temporarily ambiguous verb struc-
tures, such as those in garden-path sentences.
4 Semantic Model of Chinese Gar-
den-Path Sentence
Based on the Sentence Degeneration model,
there are two types of Chinese Garden-Path Sen-
tences: SD type ambiguity garden-paths and NP
allocated ambiguity garden-paths.
A temporarily ambiguous verb structure in a
sentence always has more than one semantic in-
terpretation that can be represented as a type of
SD. This phenomenon we call SD type ambiguity.
If an SD type ambiguity includes a prototype SD,
a garden-path arises. For example, an ambiguous
structure like咬死猎人的狗has two different
interpretations as A and B in sentence (1) and
(2):
A. It is a key-element SD in sentence (1),
where狗(dog) is the subject of咬死(kill), and
猎人(hunter) is the object of咬死(kill).
B. It is a prototype SD in sentence (2), where
狗(dog) is the object of咬死(kill), and猎人
(hunter) is the attributer of狗(dog).
Obviously,咬死猎人的狗has SD type am-
biguity, and one type of SD is prototype SD.
Therefore, sentence (1) and sentence (2) are gar-
den-path sentences.
An NP allocated ambiguity garden-path is a
sentence in which one NP can be both the object
of v1 and the subject of v2. Given the structure
NP1+v1+NP2+v2+NP3, if NP1+v1+NP2 is a
clause, NP2+v2+NP3 is a clause, too; there is an
ambiguity about whether NP2 serves as either
the object of the first clause or the subject of the
second clause. Unlike the garden-path that arises
from an SD type ambiguity, NP allocated ambi-
guity garden-paths confuse the main verb of the
sentence. For example, Sentence (8) has two dif-
ferent interpretations as A and B. The difference
in the two interpretations is the role of the solu-
tion. So, sentence (8) is a garden-path sentence
with an NP allocated ambiguity.
(8)这个学生忘记答案在书的背后(The stu-
dent forgot the solution was in the back of the
book.)
A. the solution is the subject of was, the main
verb is forgot; the solution was in the back of the
book, which is a prototype SD, is the object of
35
forgot.
B. the solution is the object of forgot, the main
verb is was, the student forgot the solution, which
is a prototype SD, is the subject of was.
We can see that it is necessary for both types
of garden-path that NP1+v1+NP2 be a clause. If
there is an NP allocated ambiguity garden-path,
NP1+v1+NP2 is a clause together with
NP2+v2+NP3 as a clause. If there is an SD type
ambiguity garden-path, NP1+v1+NP2 has to be
a prototype SD together with one of other two
types of SD (Key-element SD or packed SD).
Thus, this clause, NP1+v1+NP2, is called a gar-
den-path detecting signal.
Therefore, in our model the garden-path is
represented as one of two types of ambiguity: the
SD type ambiguity and NP allocated ambiguity.
Garden-path processing involves detecting and
analyzing these two types of ambiguities.
5 Algorithm for processing Chinese Gar-
den-Path Sentences
A Chinese Garden-Path Sentence is processed
in four steps:
(1) Initially, v1 is analyzed as the main verb.
(2) When v2 is encountered, if there is a clause
before v2, this is a garden-path detecting
signal. It is necessary to detect and analyze
the garden-path in this sentence.
(3) Detect if v1 and v2 can give rise to a gar-
den-path (see section 5.1).
(4) Determine the main verb of the sentence and
the semantic interpretation of the gar-
den-path sentence (see section 5.2).
5.1 Garden-path detection
Given an input string S, suppose its gram-
matical structure is NP1+v1+NP2+v2+NP3,
where NP1 and NP3 can be absent. Therefore, a
garden-path detecting signal means that
NP1+v1+NP2 is a clause.
The garden-path can be detected in two steps
as follows:
Step 1: test if there is SD type ambiguity in
NP1+v1+NP2.
We can look at the clause NP1+v1+NP2 as a
prototype SD without any change. If this proto-
type SD can be analyzed as another type of SD,
such as key-element or packed SD, an SD type
ambiguity is found, and the input S is a gar-
den-path sentence. Otherwise, if there is no SD
type ambiguity, the input S is a non garden-path
sentence.
As mentioned above, sentence (1) has an am-
biguity between a prototype and a key-element
SD, and it is a garden-path sentence. Consider
another sentence (9), with grammatical structure
NP1+v1+NP2+v2. Because the Chinese word
的(of) in NP2 is a sign of SD, the structure can
be rewritten as NP1+v1+NP21+的+NP22+v2.
(9)小王研究鲁迅的文章发表了(The paper
which Mr. wang research on Luxun is published.)
The structure NP1+v1+NP21+的+NP22 can be
analyzed in two different ways as follows. Ob-
viously there is an ambiguity between prototype
SD and packed SD, and sentence (9) is a gar-
den-path sentence.
A. It is a prototype SD, where文章(paper) is
the object of研究(research), and文章(paper)
was written by鲁迅(Luxun).
B. It is a packed SD, where鲁迅(Luxun) is
the object of研究(research), and文章(paper)
was written by小王(Mr. wang).
Although the structure v1+NP2+v2 in sen-
tence (1) and the structure NP1+v1+NP2+v2 in
sentence (9) can give rise to garden-paths, not all
the instances of these two structures are like this.
For example, in sentence group (10)年轻人
(younger) and刀(knife) disfavor being objects
of热爱(love) and削(peel), so热爱祖国的年
轻人is only analyzed as a key-element SD, and
小王削苹果的刀is only analyzed as a packed
SD. There is no garden-path detecting signal, so
these sentences are non garden-path sentences.
(10)热爱祖国的年轻人回国了(The
younger who love his country go back.)
小王削苹果的刀不见了(The knife with
which Mr. wang peeled the apple is lost.)
Furthermore, in sentence group (11), v1+NP2
is a clause, so there is a garden-path detecting
signal. However,皮(fruit skin) disfavors being
the subject of削(peel), and门(door) is not a
packed word, so削苹果的皮and小王推开房
间的门are only analyzed as prototype SD.
There is no SD type ambiguity, so these sen-
tences are non garden-path sentences.
(11)削苹果的皮要小心(Peeling the apple
need to be careful.)
小王推开房间的门走了(Mr. wang opened
the door and went away.)
Step 2: test if NP2+v2+NP3 is a clause.
If NP2+v2+NP3 is not a clause, definitely
there is no NP allocated ambiguity, and the sen-
tence is not a garden-path sentence. For example,
36
in sentence (12)伊拉克起因于能源危机(Iraq
is due to the crisis of energy) is not a clause, so
sentence (12) is a non garden-path sentence.
(12)美国打击伊拉克起因于能源危机(That
USA attacked Iraq is due to the crisis of energy)
If NP2+v2+NP3 is a clause, there are two in-
terpretations for v1 and v2.
First, v1 and v2 are serial verbs, and the sen-
tence can be divided into two separate sentences;
one is NP1+v1+NP2, the other is NP2+v2+NP3
and the subject of v2 is NP2. For example, sen-
tence (13) can be divided into sentences (14) and
(15). This phenomenon can be interpreted as
sentence (15) sharing大会(conference) with
sentence (14), which is not NP allocated ambigu-
ity, so the sentence (13) is a non garden-path
sentence.
(13)文件将提交给大会讨论(The file will
be given to the conference to discuss.)
(14)文件将提交给大会(The file will be
given to the conference.)
(15)大会讨论这个文件(The conference
will discuss the file.)
Second, one of v1 and v2 is the main verb of
the sentence, and NP2 has to be in NP1+v1+NP2
or NP2+v2+NP3, and cannot be shared. For
example, in sentence (8), the solution cannot be
shared by forgot and was. Absolutely this is an
NP allocated ambiguity, and the sentence is a
garden-path sentence.
The difference between a serial verb interpre-
tation and an NP allocated ambiguity interpreta-
tion is the semantic information of the two verbs.
Suppose VS(pro) is the set of all verbs whose
subject can be a prototype SD, VO(pro) is the set
of all verbs whose object can be a prototype SD.
Verbs about mental acts, emotions or other hu-
man feelings, such as forget, worry, cry, belong
to the VS(pro). Verbs about propositions, causes
and results, such as be, result in, be due to, be-
long to both VO(pro) and VS(pro).
If NP2+v2+NP3 is a clause, and if v1 is not one
of VO(pro) and v2 is not one of VS(pro), the sen-
tence is a non garden-path sentence and these
two verbs are serial verbs. Otherwise, the sen-
tence is a garden-path sentence.
5.2 Garden-path analysis
A garden-path is always affected by the selec-
tion of the main verb of a sentence. In the gar-
den-path caused by SD ambiguity, v1 is regarded
as the main verb initially, however, in the end, v2
is the real main verb. In the garden-path caused
by NP allocated ambiguity, both v1 and v2 can be
the main verb. So, the garden-path analysis in-
cludes two steps: the first step is determining the
main verb of the sentence; the second step is
disambiguating the SD type or the NP allocated
ambiguity, and determining the semantic struc-
ture of the sentence.
Given a garden-path sentence with grammati-
cal structure NP1+v1+NP2+v2+NP3, the analy-
sis process is as follows:
First, if an SD type ambiguity is detected, it
means NP1+v1+NP2 can be a prototype SD and
key-element or packed SD, and v2 always is the
main verb of the sentence. The ambiguity can be
processed as in Figure 1. For example, In sen-
tences (1) and (2),是(is) is one of VS(pro) and
逃跑(run away) is not, so咬死猎人的狗is
processed as a key-element SD in sentence (1)
and a prototype SD in sentence (2).
Second, if an NP allocated ambiguity is de-
tected, it means that both NP1+v1+NP2 and
NP2+v2+NP3 can be clauses. The main verb can
be determined in Figure 2. The NP allocated am-
biguity can be processed as in Figure 3.
(16)张先生看见李小姐正在跳舞(Mr.
Zhang saw Miss. Li dancing.)
The result of garden-path analysis is a seman-
tic structure for the sentence. In Figures 1 and 3,
a flag of prototype SD, key-element SD and
packed SD, which indicates the semantic inter-
pretation, is added to the grammatical structure
of the sentence. Therefore, the main verb, which
is always outside the flag, and the semantic
structure of the sentence are both represented.
v2 is NP1+v1+NP2 is Sentence semantic structure Example
one of VS(pro)
not one of VS(pro)
not one of VS(pro)
a prototype SD
a key-element SD
a packed SD
(NP1+v1+NP2)+v2+NP3
<NP1+v1+NP2>+v2+NP3
{NP1+v1+NP2}+v2+NP3
Sentence (2)
Sentence (1)
Sentence (9)
Figure 1: The process of SD type ambiguity in garden-path sentences. Where ( ) is the flag
37
indicating that the content in it is prototype SD, < > is the flag of key-element SD, and { } is the
flag of packed SD.
V1 is v2 is the main verb is Example
one of VO(pro)
one of VO(pro)
not one of VO(pro)
not one of VS(pro)
one of VS(pro)
one of VS(pro)
v1
prior to be v1
v2
Sentence (16)
Sentence (8)
Sentence (12)
Figure 2: The main verb determining in garden-path sentences. Here prior to means that if v1 is
one of VO(pro) and v2 is one of VS(pro), the main verb is v1 in most cases. In some cases it depends
on the meaning of v1 and v2, whether v1 is the main verb. These cases are out of the scope of consid-
eration of this paper.
Main
verb is
NP2 is Sentence semantic
structure
Example Comment
v1 The subject
of v2
NP1+v1+(NP2+v2
+NP3)
Sentence
(8),(16)
NP2+v2+NP3 is a prototype SD,
this SD is the object of v1.
v2 The object
of v1
(NP1+v1+NP2)+v2
+NP3
Sentence (12) NP1+v1+NP2 is a prototype SD,
this SD is the subject of v2.
Figure 3: The process of NP allocated ambiguity in garden-path sentence. Where ( ) is the flag
indicating that the content in it is prototype SD.
6. Evaluation and Discussion
To conduct a reliable evaluation, a test sentence
set and a simple knowledge base were developed.
The test set includes 100 manually annotated
Chinese garden-path sentences and 100 non gar-
den-path sentences with grammatical structure
NP1+v1+NP2+v2+NP3. The knowledge base
includes two aspects: one is if the verb is one of
VS(pro) or VO(pro), the other is the concepts
which the subject/object of the verb favor. And
there are about 800 verbs in our knowledge base.
Next, two experiments have been conducted.
The first one is designed to test if our model can
detect garden-paths effectively. The second one
is designed to evaluate if garden-path sentences
can be correctly analyzed. The results of the ex-
periments are shown in Tables 1 and 2.
Total Num Detected Correct P(%) R(%) F(%)
100 85 79 92.9 79 85.4
Table 1: Performance of detection algorithm.
Total of
Detected
SD type ambiguity
Analysis Correct
NP allocated ambiguity
Analysis Correct
Total of
Correct
P(%)
85 53 24 77 90.6
Table 2: Performance of analysis algorithm.
Where, P is precision ratio, R is recall ratio, and
F is F-measure (Fβ=1, which is defined as
2PR/(P+R)).
We can see that on this small test set, our al-
gorithm achieves good performance in detection
and analysis of Chinese garden-path sentences.
We also conducted an error analysis, showing
that two main factors lead to detection errors.
The first is that attributer processing of NP2’s
is not considered. For example, in管理朋友的
38
公司(manage the friend’s company),朋友的公
司(the friend’s company) is an NP which cannot
be divided to NP21(朋友friend)+的+NP22(公
司company) and cannot be detected if there is
an SD type ambiguity.
The second is coordination ambiguity inter-
acting with NP allocated ambiguity, as in Sandra
bumped into the busboy and the waiter told her
to be careful, which has not been considered.
Furthermore, there are two sentences correctly
detected as Chinese garden-path sentences, but
there are neither SD type ambiguities nor NP al-
located ambiguities in them. This is why there
are 79 correct detections in Table 1, but only 77
correct analyses in Table 2. One of these sen-
tences is sentence (17), in which我是县长(I am
the mayor) looks like a prototype SD, however “
是…的”is used to emphasized县长(the mayor)
in Chinese.
(17)我是县长派来的（It is the mayor that
instructed me to come here）
7. Conclusions and Future Work
The contributions of this paper are three-fold.
First, the Sentence Degeneration model is in-
troduced which can represent the differences in
interpretation of the same grammatical structure.
Second, we represent garden-paths as SD type
ambiguity and NP allocated ambiguity. These
two ambiguities come from semantics but not
grammar.
Third, we present a unified approach to proc-
essing garden-paths, in which ambiguity detec-
tion and analysis take the place of revision. The
result of our approach is the semantic structure of
a garden-path sentence.
The results of two experiments we conducted
show that our model and algorithm can analyze
Chinese garden-path sentences effectively. In our
future work, we will build a complex knowledge
base for verbs to support our semantic analysis.
We will also develop attributer processing and
coordination disambiguation to improve the per-
formance of our algorithm.
Moreover, we will extend our algorithm to
detect and analyze garden-paths caused by sen-
tences which have no verbs. This phenomenon is
a typical ambiguity in Chinese sentences, such as.
Sentence (18):
(18)她穿裙子很漂亮(She is beautiful dress-
ing skirt).
Acknowledgments
I thank Prof. Zengyang Huang and Dr. Chuanji-
ang Miao for their valuable comments on an
early draft of this paper. I also thank Zheng Wu
for his wonderful work in algorithm develop-
ment.

References
Frazier, L., & Clifton, C., Jr. 1998. Sentence reanaly-
sis and visibility. In J. D. Fodor & F.Ferreira (Eds.),
Reanalysis in sentence processing. Dordrecht,
Kluwer.
Fernanda Ferreira,etc. 2001. Misinterpretations of
Garden-Path Sentences: Implications for Models of
Sentence Processing and Reanalysis. Journal of
Psycholinguistic Research, 30(1).
Michael J. Pazzani. 1984. Conceptual Analysis of
Garden-Path Sentences. Proceedings of the 10th in-
ternational conference on Computational linguis-
tics.
Karl G.D. Bailey, etc. 2003. Disfluencies affect the
parsing of garden-path sentences. Journal of Mem-
ory and Language, 49:183–200.
Zengyang Huang. 1998. The theory of Hierarchical
Network of Concepts (in Chinese). Tsinghua Uni-
versity Press, Beijing, China.
Zengyang Huang. 2004. The Basic Concepts and
expression of the Concepts space (in Chinese).
Ocean Press, Beijing, China.
Zhiwei Feng. 2003. The automatic parsing algo-
rithm for garden-path sentence (in Chinese).
Contemporary Linguistics, 5(4).
