Automated Generalization of Phrasal Paraphrases from the Web*
Weigang Li 
School of Computer
Science and Tech-
nology, Box 321,
Harbin Institute of
Technology, Harbin,
 P.R. China, 150001
lee@ir.hit.edu
.cn
Ting Liu 
School of Computer
Science and Tech-
nology, Box 321,
Harbin Institute of
Technology, Harbin,
 P.R. China, 150001
tliu@ir.hit.ed
u.cn
Yu Zhang
School of Computer
Science and Tech-
nology, Box 321,
Harbin Institute of
Technology, Harbin,
 P.R. China, 150001
zhangyu@ir.hit
.edu.cn
Sheng Li 
School of Computer
Science and Tech-
nology, Box 321,
Harbin Institute of
Technology, Harbin,
 P.R. China, 150001
lis@ir.hit.edu
.cn
Wei He 
School of Computer
Science and Tech-
nology, Box 321,
Harbin Institute of
Technology, Harbin,
 P.R. China, 150001
truman@ir.hit.
edu.cn
Abstract
Rather than creating and storing thou-
sands of paraphrase examples, para-
phrase templates have strong 
representation capacity and can be used
to generate many paraphrase examples.
This paper describes a new template
representation and generalization
method. Combing a semantic diction-
ary, it uses multiple semantic codes to
represent a paraphrase template. Using
an existing search engine to extend the
word clusters and generalize the exam-
ples.  We also design three metrics to
measure our generalized templates. The 
experimental results show that the rep-
resentation method is reasonable and 
the generalized templates have a higher 
precision and coverage.
1 Introduction
Paraphrases are alternative ways to convey the 
same information (Barzilay and McKeown,
2001) and they have been applied in many fields
of natural language processing. There are many
previous work on paraphrase examples extrac-
tion or combining them with some applications
such as information retrieval and question an-
swering (Agichtein et al., 2001; Florence et al., 
2003; Rinaldi et al., 2003; Tomuro, 2003; Lin
and Pantel, 2001;), information extraction 
(Shinyama et al., 2002; Shinyama and Sekine, 
2003), machine translation (Hiroshi et al., 2003;
Zhang and Yamamoto, 2003), multi-document
(Barzilay et al., 2003).
There is also some other research about 
paraphrase. (Wu and Zhou, 2003) just extract 
the synonymy collocation, such as <turn on, 
OBJ, light> and <switch on, OBJ, light> using
both monolingual corpora and bilingual corpora 
to get an optimal result, but do not generalize
them. (Glickman and Dagan, 2003) detects verb
paraphrases instances within a single corpus
without relying on any priori structure and in-
formation. Generation of paraphrase examples
was also investigated (Barzilay and Lee, 2003;
Quirk et al., 2004).
Rather than creating and storing thousands of 
paraphrases, paraphrase templates have strong 
representation capacity and can be used to gen-
erate many paraphrase examples. As (Hirst, 
2003) said, for each aspect of paraphrase there 
are two main challenges: representation of 
knowledge and acquisition of knowledge. Cor-
responding to the problem of generalization of 
paraphrase templates, there are also two prob-
lems: the first is the representation of paraphrase
templates and the second is acquisition of para-
phrase templates.
There are several methods about paraphrase
templates representation. The first method is 
using the Part-of-Speech (Barzilay and McKe-
own, 2001; Daumé and Marcu, 2003; Zhang and 
Yamamoto, 2003), the second uses name entity 
as the variable (Shinyama et al., 2002; Shinyama
and Sekine, 2003), the third method is similar to 
the second method which is called the inference 
rules extraction (Lin and Pantel, 2001).
A paraphrases template is a pair of natural
language phrases with variables standing in for
certain grammatical constructs in (Daumé and 
*: Supported by the Key Project of National Natural Sci-
ence Foundation of China under Grant No. 60435020
49
Marcu, 2003). He used Part-of-Speech to repre-
sent templates. But for some cases, the POS will 
be very limited and for some other cases will be 
over generalized. For example:
��,�9 � ��?�k            
(In my view/mind ----I feel)
The above pair of phrases is a paraphrase, it 
can be generalized using POS information: 
�  [pronoun] ,�9 
(In [pronoun] view/mind) 
[pronoun] ?�k 
( [pronoun] feel)
But for this template many noun words will
be excluded. From this point of view, the tem-
plate representation capacity is limited. But for 
other examples, the POS information will be 
over generally. For example:
8�p,X���� 
(What's the price for the apples?)
8�p��J��x� 
(How much is the apples per Jin?)
Here, we just generalize one variable “8�p ”.
Then, the template becomes:
[noun] ,X���� 
(What's the price for the [noun]?)
[noun] ��J��x� 
(How much is the [noun] per Jin?)
If there is a sentence “0�A� ,X���
(What's the price for the notebook?)”, its’ para-
phrase will be “0�A� ��J��x (How much 
is the notebook per Jin?)” according to this tem-
plate. Obviously, the result is unreasonable.
(Shinyama et al., 2002) tried to find para-
phrases assuming that two sentences sharing
many Named Entities and a similar structure are 
likely to be paraphrases of each other. But just 
name entities are limited, too. And (Lin and 
Pantel, 2001) present an unsupervised algorithm 
for discovering inference rules from text such as 
“X writes Y” and “X is the author of Y”. This 
generalized method has good ability. But it also
has some limited aspect. For example:
[Jack] writes [his homework].
According to the paraphrase template, the
target sentence will be transformed into “[Jack]
is the author of [his homework]”. It’s obviously
that the generated sentence is not standard.
So how to represent paraphrase templates
and generalize the paraphrase examples is a very 
interesting task. In this paper, we present a novel
approach to represent paraphrase template with 
semantic code of words and using an existing
search engine to get the paraphrase template.
The remainder of this paper is organized as 
follows. In the next section, we give the over-
view of our method. In section 3, we define the 
representation method in details. Section 4 pre-
sents the generalization method. Some experi-
ments and discussions are shown in Section 5. 
Finally, we draw a conclusion of this method
and give some suggestions about future work. 
2 Overview of Generalization Method
The origin input of our system is a seed phrasal
paraphrase example. And the output is the gen-
eralized paraphrase templates from the given 
examples. The overall architecture of our para-
phrase generalization is represented on figure 1. 
A seed phrasal
paraphrase examples
Getting the slot word
Extend the slot word
using Search Engine
on every example
Mapping two word
sets to their semantic
code sets
Intersection operation
on the two semantic
code sets
Generalizing a
template
Figure 1: Sketch Map of Paraphrase example
Generalization
We also use the example (1) to illustrate the 
representation. Here a semantic dictionary called 
“TongYiCiCiLin” (Extension Version)1 is used. 
The pair of phrases is a phrasal paraphrase. At
first, after preprocessing which includes word
segment, POS tagging and word sense disam-
biguation, we get the slot word in the paraphrase.
In this example, the slot word is “� (I)”. Then
we search the web using the context of the slot 
word. Every phrase in the phrasal pair derives a
set of sentences which include the original 
phrase context. A dependency parser on these 
sentences is used to extract the corresponding
word with the slot word. Two word sets can be 
obtained through the two sentence sets. Then,
we map word sets to their semantic code sets
1 TongYiCiCiLin (Extended Version) can be downloaded
from the website of HIT-IRLab (Http://ir.hit.edu.cn). In the 
past section, we abbreviate the TongYiCiCiLin (Extended
Version) to Cilin (EV) 
50
according to Cilin(EV). Then an intersection 
operation is conducted on the two sets. We use 
the intersection set to replace the slot word and 
generate the final paraphrase template. 
In order to verify the validation of the gener-
alized paraphrase template, we also design an 
automatic algorithm to confirm whether the 
template is reasonable using the existing search 
engine.
3 Representation of Template 
In the section of introduction, some representa-
tion methods of paraphrase template have been 
introduced. And we proposed a new method us-
ing word semantic codes to represent the vari-
able in a template. Before we introduce the 
representation method, Firstly, we give some 
general introduction about the semantic diction-
ary of Cilin(EV). 
3.1 TongYiCiCiLin (Extended Version) 
Cilin (EV) is derived from original TongY-
iCiCilin in which word senses are decomposed 
to 12 large categories, 94 middle categories, 
1,428 small categories. Cilin (EV) removes 
some outdated words and updates many new 
words. More fine-grained categories are added 
on the base of original classification system to 
satisfy the more complex natural language ap-
plications. The encoding criterion is shown in 
the table 1:
Table 1 Encoding table of dictionary
Encoding
bit 1 2 3 4 5 6 7 8
Example D a 1 5 B 0 2 =
Attribute Big Middle Small groups Atom groups
Layer 1 2 3 4 5
The encoding bits are arranged from left to 
right. The first three layers are same with Cilin. 
The fourth layer is represented by capital letters 
and the fifth layer is two-bit decimal digit. The 
last bit is some more detailed information about 
the atom groups. 
3.2 An Example of a Paraphrase Template 
For simplicity, we just select one slot word in 
every paraphrase. And we stipulated that only 
content word can be slot word. We also use the 
above paraphrase example (1). 
��,�9 � ��?�k         
(In my view/mind ----I feel)
Here, we get the slot word “�(I)”. Through 
the Word Sense Disambiguation processing, we 
get its semantic code “Aa02A01=” according to 
the fifth layer in Cilin(EV). If we just use the 
semantic code of the slot word, we can get a 
simple paraphrase template as follows:  
� [Aa02A01=] ,�9 
(In [Aa02A01=]  view/mind)
[Aa02A01=] ?�k 
([Aa02A01=]  feel)
But it is obviously that the template is very 
limited. Its’ representation ability is also limited. 
So how to extend the ability of a paraphrase 
template is a challenging work.  
3.3 Extending the Template Abstract Ability 
According to the feature of Cilin(EV) architec-
ture, we can use the higher layer’s semantic 
code instead of the slot word to generalize the 
paraphrase template naturally. Of course it’s a 
very simple method to extend the template abil-
ity, but it also brings more redundancy of a 
paraphrase template and it will be proven in the 
later section. 
So we use multiple semantic codes of the dif-
ferent layer instead of only one semantic code of 
slot word in Cilin (EV). The later experimental 
results prove this representation has a good per-
formance with a good precision and coverage. 
4 Generalizing to Templates 
As mentioned above, we can use multiple se-
mantic codes to generalize paraphrase examples. 
So the problem of how to generalize paraphrase 
examples is transformed into the problem of 
how to get the multiple semantic codes set. We 
proposed a new method which uses the existing 
search engine to reach the target.  
4.1 Getting the Candidate Sentences 
After we removed the slot word in the para-
phrase examples, two phrasal contexts of the 
original paraphrase phrases were obtained. Each 
phrase without slot word is used as a search 
query for an existing search engine and achiev-
ing many sentences which include the query 
word. For this example, the two queries are “� 
,�9 (in…view)” and “?�k (feel)”. Each query 
gets one sentence set respectively. Part of the 
two result sentence sets are shown in figure 2 
and figure 3: 
51
Figure 2. Sentence Set 1 
Figure 3. Sentence Set 2 
From the above two sentence sets, we can
find that there is some noisy information in the 
sentences. In order to extend the correspondent
words of the slot word, it is not enough that we 
just use the position information or POS tagging
information of the slot word. Even if we extract
these words, many of them can’t be found in the
dictionary because they are not simple words.
Benefiting from the idea of (Lin and Pantel, 
2001), we use a dependency parser to determine
the correspondent extended words. 
4.2 Dependency Parser 
In this paper, we use a dependency parser (Ma et 
al., 2004) to extract the candidate slot word. For 
example, the dependency parsing result of the 
phrase of “��,�9 ” is shown in figure 4. 
Figure 4. Dependency parsing result 
The arcs in the figure represent dependency
relationships. The direction of an arc is from the 
head to the modifier in the relationship. Labels 
associated with the arcs represent types of de-
pendency relations. Table 2 lists a subset of the
dependency relations in the HIT-IRLab depend-
ency parser2.
Table 2. A subset of the dependency relations 
Relation Description
ATT nG2� (attribute)
HEDm(head)
SBJ A� (subject)
ADV (�4�X (adverbial)
VOB |�G2� (verb-object)
� �C
�G�4�)� ,� 9 Y
�
 N��L$ 
� C{L
4{l� ,� 9  	�?UC �C.J�P�H    
(� � �� ,� 9 TT�o9&
�,X 
 , 7 AxA� � �6,� ,� 9 ��G�?U "    
� �� ,� 9  !�4�x5b�,X�,�
2 More information about the dependency parser can be got
from http://ir.hit.edu.cn/cuphelp.htm
4.3 Extracting the extended words 
We just use a very simple method to get the ex-
tended words from the parsed sentences. At first, 
we record the relations of the original parsed 
phrasal examples. And then we use these rela-
tions to matched similar part in the candidate 
parsed sentence except slot word. And we omit
these unseen relations and content words which
don’t appear in the original parsed phrasal ex-
amples. Then we can get the extended words. 
Fw��{+= ?� k 4\	�( 
��Z 4! ?� k 5b  
�� ?� k 
��,��E�	��$ eK�
�    
#\C
5� ?� k D�V/� 
D?S ?� k 7��CDk3!OZ  B 7 2 0 -6�
Figure 5. Dependency parsing result 
Figure 5 shows the dependency parsing result 
of the phrase of “��C
�G�4�)�,�9 ”(In for-
eign capital fund manager view). We can easily
find that the extended word of the slot word
“ � ”(I) is “ 4�)� ”(manager). Two extended
word sets can be extracted from two sentence
sets. Then we map each word to their semantic
code to get two semantic code sets. Intersection
operation is conducted on these two semantic
code sets to obtain their intersection set. Finally, 
we use the semantic code set instead of the slot 
word to generate the paraphrase template.
4.4 Some tricks 
Because the precision of the current dependency
parser on Chinese is not very high, we just ex-
tract a part of the candidate sentences to parse. 
There are three patterns to segment the long 
candidate sentences according to position of slot 
word in paraphrase examples. They are called
FRONT, MIDDLE and BACK. Here we use an
example to illustrate it as shown in table 3: 
Table 3 Examples of sentence segmentation
Pattern Origin Phrase Segment examples
FRONT (SW)?��Fw � � { + = ?� k
4\	�( 
MIDDLE�(SW)"��� � C 
� G� 4� )� ,�
9 Y
�
 N�
�L$ 
The bold section in the sentence will be ex-
tracted to parse. Pattern type can be decided by 
52
the position relation between slot word and con-
text words. And these patterns can reduce the 
relative error rate of the dependency parser. That 
is to say, if the original phrase is parsed wrongly, 
the extracted segments may be parsed wrongly 
with the similar error. But according to our 
method, this kind of parser error has little influ-
ence on the final extracting result. 
5 Experiments and Discussions 
5.1 Setting 
We extract about 510 valid paraphrase examples 
from a Chinese paraphrase corpus (Li et al., 
2004). For simplicity, we just select those 
phrasal paraphrase examples which own same 
word. And we stipulate only content word can 
be as slot word. We just use four seed phrasal 
paraphrases as the original paraphrases in this 
paper. And the generalized paraphrase templates 
represented by semantic codes of the fifth layer 
in Cinlin (EV) are also shown in the Table 4: 
Table 4: Examples of the generalized template 
Origin 
Phrases 
Generalized Paraphrase  
templates 
�?�k [Aa01A01=,Aa01A05=,   
Aa01C03=,Aa02A01=,  …]?�k1
� � ,�
9
� [Aa01A01=,Aa01A05=, 
  Aa01C03=,Aa02A01=,...  ],�9
�&�Z �  [Ac03A01=,Ah04A01=, 
Ah05A01=,Am03D01@,…]Z2
&�k\ [Ac03A01=,Ah04A01=, 
Ah05A01=,Am03D01@,…]k\
F� &� 
Z
F� [Fb01A01=,Gb07B01=, 
Hb06A01=,He15B01=,… ]Z3
< &� `
Z
< [Fb01A01=,Gb07B01=, 
Hb06A01=,He15B01=,…  ]`Z
8� p ,X
�  
��
[Aa03A01=,Ac03A01=, 
Ba05A10#,Bb02A01=,…] ,X � 
��4
8� p �
�J�x
[Aa03A01=,Ac03A01=,Ba05A10#,
Bb02A01=,…]��J��x
5.2 Evaluation on Templates 
The goal of the evaluations is to confirm how 
reasonable this kind of representation method of 
paraphrase templates is and how well the tem-
plate is. We evaluated the generalized para-
phrase template in three ways. They are listed in 
the following three categories: 1) Reasonability; 
2) Precision; 3) Coverage. 
1) Reasonability 
The reasonability of a paraphrase template aims 
to measure the reasonable extent of the presenta-
tion method with multiple semantic codes. For 
example, if we use POS to generalize a para-
phrase template, its reasonability is very lower; 
that is to say, POS is not suitable to represent 
paraphrase template in some extent.  
We use an existing search engine to calcu-
late the reasonability of every paraphrase tem-
plate. Firstly, we instantiate all paraphrase 
examples from a template. Then all these exam-
ples are as the queries of the search engine. If 
two phrases in one paraphrase can be matched 
completely from the search engine, it also means 
that one or more examples are found on the Web 
via search engine, we then consider this para-
phrase is reasonable. Using this method we can 
get the approximate evaluation of all the exam-
ples. We define two metrics: 
Strict_Reasonability = S / N 
Loose_Reasonability = (L + S) / N 
Where N is the total number of the instanti-
ated examples; S is the number of the para-
phrase examples which two phrases in it can be 
matched all; L is the number of paraphrase ex-
amples only one phrase in a paraphrase can be 
matched.
2) Precision 
Every template is correspondent to the examples 
number with the semantic code of different layer 
in Cilin (EV) as shown in table 5.  
Table 5 Templates and their correspond exam-
ples number 
Instantiated examples 
numberNumber of Paraphrase
templates Cilin3 Cilin4 Cilin5
1 2696 1815 478
2 13032 6354 3011
3 1057 587 177
4 3004 2229 429
From the above table, we can find that every 
template can instantiate many examples. If 
manually judging all of these examples will 
spend plenty of time. So we just sample part of 
all instantiate examples, 200 paraphrase exam-
ples for each template in this paper. For each 
53
phrase in a sample paraphrase example, it is as 
search query to get the first two matched sen-
tences. Evaluators would be asked whether it is
semantically okay to replace the query in the
sentence by the correspondent phrase in a para-
phrase. They were given only two options: Yes
or No. If search query have no matched results, 
we consider that this phrase cannot be replace 
with its correspondent paraphrase. According to 
the above regulations, we know that every para-
phrase examples correspondent to 4 sentences. If 
we sample n examples from a template, the pre-
cision of a paraphrase template can be calculated
by:
Precision = R / (4 * n) 
Where, R is the number of sentences which
is considered to be correct by the evaluator.
3) Coverage 
Evaluating directly the coverage of a paraphrase 
template is difficult because humans can’t enu-
merate all the words to be suitable to the tem-
plate. We use an approximate method to get the
coverage of a template. At first we use another 
search engine to get candidate sentences with 
similar method for generalization of a para-
phrase template. From these retrieved sentences
we can get many different words with the 
known generalized words because more than
85% of search results from different search en-
gine are different. Evaluators extract every sen-
tence which can be replaced with the 
correspondent phrase in a paraphrase and the
new sentences retain the origin meaning. We 
know each sentence is correspondent to a word. 
Then we define two metrics: 
Surface_Coverage = M / NS
Semantic_Coverage =
Map(K) / (Map(NS-M) + Map(K)) 
Where, NS is the number of all manually
tagged right words, M is the number of words 
which can be instantiated from a paraphrase
template, K is the number of all the words that 
generalized the template at the front. Map(X) is 
the total word number of the word clusters 
which derived from X word in the semantic dic-
tionary of Cilin(EV).
5.3 Result 
In order to exhibit the merit of our method, we 
conduct four groups of experiment. They are
POS-Tag, Cilin3, Cilin4 and Ciln5, respectively.
Especially, we just randomly select 400 words 
to satisfy the POS information.
Table 6: Experiment Results 
Reasonability
(%)
Coverage
(%)
St_R Lo_R Su_C Se_C
Preci-
sion
(%)
POS 10.50 17.00 90.00 ---- 11.75
Cilin3 45.57 84.50 27.55 38.71 45.75
Cilin4 46.89 84.54 23.87 44.48 64.13
Cilin5 46.24 83.12 20.39 39.47 69.88
Every value in table 6 is a average value of 
four values correspondent to four templates.
From the table we can find that the reasonability
of the Cilin-based representation template
changes little, and that of POS-based representa-
tion is very lower. We find that the longer origi-
nal phrases are, the lower the coverage of the
generalized template is. Although the average 
coverage of generalized template is relatively
low, we can draw a conclusion that using multi-
ple semantic codes to generalize phrasal para-
phrase examples is reasonable.
The column of the coverage shows that the 
coverage rates of Cilin-based templates are all
not more than 50%. And the POS-based tem-
plate has a very high coverage rate. And we 
know that the extended information is not
enough only depending on one search engine. 
We will combine several different search en-
gines with together to solve this problem in the 
future work. 
1.0 1.5 2.0 2.5 3.0 3.5 4.0
0
10
20
30
40
50
60
70
80
90
100
 strict_Reasonability  loose_Reasonability
 surface_Coverage  semantic_Coverage
 Precision
Va
lue
s o
f P
erc
en
t
Different Template Representation Method
Figure 6. Experimental Results
The numbers from one to four on the X-axis
are correspondent to POS, Cilin3, Cilin4 and 
Cilin5 in figure 6. We can see the features
clearly of different representation methods of 
template from the figure 6. We can find that
54
Cilin5-based template has the highest precision, 
but its coverage is lower. And Cilin3-based 
template has opposite feature. This is because 
that one semantic code of Cilin3 includes more 
words than that of Cilin5. At the same time, 
more words bring more redundant information. 
And Cilin4-based template has a good tradeoff 
between coverage and precision. So we con-
clude that the semantic code of fourth layer in 
Cilin (EV) is more suitable to represent para-
phrase template.  
Some additional information can be extracted 
from the generalized template. Such as, the col-
location information between the slot word and 
the context words can be extract. For example, 
in the fourth template, we can get the informa-
tion about which words can be collocated with 
“x (Jin)”.
Although this kind of representation of para-
phrase template has a good performance, it is 
weak for those words or structures that don’t 
exist in dictionary. Also, this method is not suit-
able to the named entities representation. 
6 Conclusion
In this paper, a novel method for automated 
generalization of paraphrase examples is pro-
posed. This method is not dependent on the tra-
ditional limited texts instead it is based on the 
richness of the Web. It uses the multiple seman-
tic codes to generalize a paraphrase example 
combing a semantic dictionary (Cilin (EV)). The 
experimental results proved that this representa-
tion method is reasonable and the generalized 
templates have a good precision and coverage.  
But this is just the beginning of the para-
phrase examples generalization. And we sim-
plify the problem in some aspects, such as we 
limited the number of the slot word in a para-
phrase example, and we stipulate only the same 
word can be slot word. Also, we find that our 
templates are weak for those words or structures 
that don’t exist in dictionary. Some methods in 
information extraction about named entities 
generalization can be used for reference in the 
future. Moreover, how to combine the semantic 
code with other representation forms together is 
also an interesting work. 
References
[1] Chris Quirk, Chris Brockett, and William Dolan. 
Monolingual Machine Translation for Para-
phrase Generation. editors, Dekang Lin and 
Dekai Wu, In Proceedings of EMNLP 2004, 
Barcelona, pages 142-149  
[2] Dekang Lin and Patrick Pantel. 2001. Discovery 
of Inference Rules for Question Answering. 
Natural Language Engineering 7(4):343-360 
[3] Dekang Lin and Patrick Pantel. Discovery of 
inference rules for question answering. Natural 
Language Engineering, 1, 2001.  
[4] E. Agichtein, S. Lawrence, and L. Gravano. 
Learning search engine specific query transfor-
mations for question answering. In Proceedings 
of the 10th International World-Wide Web Con-
ference (WWW10), 2001 
[5] Fabio Rinaldi, James Dowdall, Kaarel Kalju-
rand, Michael Hess, Diego Molla. 2003. Ex-
ploiting Paraphrases in a Question Answering 
System. The Second International Workshop on 
Paraphrasing: Paraphrase Acquisition and Ap-
plications 
[6] Florence Duclaye France. Learning paraphrases 
to improve a question-answering system. In 
EACL Natural Language Processing for Ques-
tion Answering, 2003 
[7] Graeme Hirst. Paraphrasing Paraphrased. In 
Proceedings of the Second International Work-
shop on Paraphrasing, 2003 
[8] Hal Daumé III and Daniel Marcu. Acquiring 
paraphrase templates from document/abstract 
pairs. In NL Seminar in ISI, 2003 
[9] Hua Wu, Ming Zhou. Optimizing Synonym 
Extraction Using Monolingual and Bilingual 
Resources. In Proceedings of the Second Inter-
national Workshop on Paraphrasing, 2003 
[10] Hua Wu, Ming Zhou. Synonymous Collocation 
Extraction Using Translation Information. In 
Proceedings of the 41st Annual Meeting of the 
Association for Computational Linguistics, 
2003 
[11] Jinshan Ma, Yu Zhang, Ting Liu, and Sheng Li. 
A Statistical Dependency Parser of Chinese un-
der Small Training Data. Workshop: Beyond 
shallow analyses - Formalisms and statistical 
modeling for deep analyses, IJCNLP-04, 4 2004. 
[12] Noriko Tomuro. 2003. Interrogative Reformula-
tion Patterns and Acquisition of Question Para-
phrases. The Second International Workshop on 
Paraphrasing: Paraphrase Acquisition and Ap-
plications 
[13] Oren Glickman and Ido Dagan. Identifying 
lexical paraphrases from a single corpus: A case 
study for verbs. In Proceedings of Recent Ad-
vantages in Natural Language Processing, Sep-
tember 2003 
55
[14] Regina Barzilay and Kathleen McKeown. Ex-
tracting paraphrases from a parallel corpus. In 
Proceedings of the ACL/EACL, Toulouse, 2001 
[15] Regina Barzilay and Lillian Lee. Learning to 
Paraphrase: An Unsupervised Approach Using 
Multiple-Sequence Alignment. In Proceedings 
of HLT-NAACL 2003, pages 16-23  
[16] Regina Barzilay, Noemie Elhadad, Kathleen R. 
McKeown. 2003. Inferring Strategies for Sen-
tence Ordering in Multidocument News Sum-
marization. The Second International Workshop 
on Paraphrasing: Paraphrase Acquisition and 
Applications 
[17] Weigang Li, Ting Liu, Sheng Li. Combining 
Sentence Length with Location Information to 
Align Monolingual Parallel Texts. AIRS, 2004, 
pages 71-77 
[18] Yusuke Shinyama and Satoshi Sekine. Para-
phrase acquisition for information extraction. 
editors, Kentaro Inui and Ulf Hermjakob, In 
Proceedings of the Second International Work-
shop on Paraphrasing, 2003, pages 65-71 
[19] Yusuke Shinyama, Satoshi Sekine, Kiyoshi 
Sudo, and Ralph Grishman. Automatic para-
phrase acquisition from news articles, In Pro-
ceedings of Human Language Technology 
Conference (HLT2002), San Diego, USA, Mar. 
15, 2002 
[20] Zhang Yujie, Kazuhide Yamamoto. Automatic 
Paraphrasing of Chinese Utterances. Journal of 
Chinese Information Processing. Vol. 117 No. 
16: 31-38(Chinese) 
56
