Paraphrasing of Chinese Utterances
Yujie Zhang
∗
Communications Research Laboratory
2-2-2, Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0289 Japan
yujie@crl.go.jp
Kazuhide Yamamoto
ATR Spoken Language Translation Research Laboratories
2-2-2, Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0288 Japan
yamamoto@fw.ipsj.or.jp
Abstract
One of the key issues in spoken language trans-
lation is how to deal with unrestricted expres-
sions in spontaneous utterances. This research
is centered on the development of a Chinese
paraphraser that automatically paraphrases ut-
terances prior to transfer in Chinese-Japanese
spoken language translation. In this paper, a
pattern-based approach to paraphrasing is pro-
posed for which only morphological analysis is
required. In addition, a pattern construction
method is described through which paraphras-
ing patterns can be eﬃciently learned from a
paraphrase corpus and human experience. Us-
ing the implemented paraphraser and the ob-
tained patterns, a paraphrasing experiment was
conducted and the results were evaluated.
1 Introduction
In spoken language translation one of the key
issues is how to deal with unrestricted expres-
sions in spontaneous utterances. To resolve this
problem, we have proposed a paraphrasing ap-
proach in which the utterances are automati-
cally paraphrased prior to transfer (Yamamoto
et al., 2001; Yamamoto, 2002). The paraphras-
ing process aims to bridge the gap between the
unrestricted expressions in the input and the
limited expressions that the transfer can trans-
late. In fact, paraphrasing actions are often seen
in daily communication. When a listener can-
not understand what a speaker said, the speaker
usually says it again using other words, i.e., he
paraphrases. In a Chinese-Japanese spoken lan-
guage translation system, the pre-processing of
Chinese utterances is involved and we attempt
to apply a paraphrasing approach. This paper
∗
This work was done when the author stayed at ATR
Spoken Language Translation Research Laboratories.
is focused on the paraphrasing of Chinese utter-
ances.
Some cases of paraphrasing research with cer-
tain targets have been reported. For example,
there has been work on rewriting the source
language in machine translation with a focus
on reducing syntactic ambiguities (Shirai et al.,
1993), research on paraphrasing paper titles
with a focus on transforming syntactic struc-
tures to achieve readability (Sato, 1999), and
research on paraphrasing Japanese in summa-
rization with a focus on transforming a noun
modifier into a noun phrase (Kataoka et al.,
1999). We have reported some research on
Chinese paraphrasing (Zhang and Yamamoto,
2001; Zhang et al., 2001; Zong et al., 2001). The
techniques of paraphrasing natural language
can be applied not only to the pre-processing
of machine translation but also to information
retrieval and summarization.
2 Goals and Approach
In the pre-processing stage of translation, Chi-
nese paraphrasing focuses on
(1) transforming the expressions of spoken lan-
guage into formal expressions,
(2) reducing syntactic and semantic ambigui-
ties,
(3) generating as many diﬀerent expressions as
possible in order to include expressions that
can be translated by the transfer, and
(4) paraphrasing the main constituents of the
utterance in case the paraphrasing of the
whole utterance has no eﬀect.
The aim of paraphrasing types (1), (2) and (4)
is to simplify the expressions of utterances, and
that of paraphrasing type (3) is to increase the
variations of utterances. At present, we focus
on paraphrasing types (1), (2) and (3).
Paraphrasing is a process that automatically
generates new expressions that have the same
meaning as the input sentence. At first glance
one would think that the problem could be re-
solved by separating it into two processes: the
parsing process that analyzes the input sentence
and obtains its meaning, and the generation
process that generates sentences from the ob-
tained meaning. However, this solution is not
practicable for the following reasons.
• At present, the techniques of parsing and
semantics analysis of the Chinese language
are far below the level needed for appli-
cation. When studying spoken language,
research on parsing and research on se-
mantics analysis are major themes them-
selves. For automatic paraphrasing, we
should first determine what kind of anal-
ysis is required and then start to develop a
parser or a semantics analyzer.
• Even if meanings can be obtained, goal (3)
cannot be achieved if only one sentence is
generated. Here, the demand that para-
phrasing should generate multiple expres-
sions is the most important. This focus is
diﬀerent from that of conventional sentence
generation.
In fact, the paraphrasing can be conducted
at many diﬀerent levels, for instance, words,
phrases, or larger constituents. Although the
paraphrasing of such constituents is probably
related to context, it is not true that paraphras-
ing is impossible without being able to under-
stand the whole sentence (Kataoka et al., 1999).
The paraphrasing process encounters the fol-
lowing problems. (i) How to identify objects,
i.e., which components of an input sentence will
be paraphrased, (ii) how to generate new sen-
tences, and (iii) how to ensure that the gener-
ated sentences have the same meaning as the
input sentence. In order to avoid the large cost
of syntax and semantics analysis, we propose
a pattern-based approach to paraphrasing in
which only morphological analysis is required.
The focus is placed on how to generate as many
diﬀerent expressions as possible and how to get
paraphrasing patterns from a paraphrasing cor-
pus.
Table 1. Part of the part-of-speech tag set of
the Penn Chinese Treebank
Symbol Explanation
NN common noun
NR proper noun
PN pronoun
DT determiner
DEC 69 in a relative-clause
DEG associative 69
M measure word
JJ other noun-modifier
VA predicative adjective
VC 48
VE 37 as the main verb
VV other verb
AD adverb
P preposition excl. 74 and 76
LC localizer
CD cardinal number
OD ordinal number
SP sentence-final particle
BA 76 in ba-construction
CC coordinating conjunction
3 Paraphrasing Pattern
The paraphrase corpus of the spoken Chinese
language consists of 20,000 original sentences
and 44,480 paraphrases, one original sentence
having at least two paraphrases (Zhang et al.,
2001). The paraphrases were obtained by the
manual rephrasing of the original sentences:
words may be reordered, some words may be
substituted with synonyms, or the syntactic
structures may be changed. Such a paraphrase
corpus contains the knowledge of how to gener-
ate paraphrases for one sentence. We intend to
get paraphrasing patterns from the corpus. By
pairing each paraphrase with its corresponding
original sentence, 44,480 pairs were obtained.
Hereafter, we call such pairs paraphrase pairs.
Word segmentation and part-of-speech tagging
were carried out on the paraphrase pairs. The
part-of-speech tagger accepted the Penn Chi-
nese Treebank tag set, which comprises 33
parts-of-speech (Xia, 2000). A part of the Penn
Chinese Treebank tag set is shown in Table 1.
3.1 Extraction of Instances
For one paraphrase pair, the paraphrase may
diﬀer from its original sentence in one of the
following paraphrasing phenomena: (1) word
order, (2) substitution of synonyms, and (3)
change of syntactic structure. For most para-
phrase pairs, the paraphrases contain a mixture
of the above phenomena. We need to classify
the paraphrasing phenomena and learn the rela-
tive paraphrasing patterns. In this way, we can
restrict the paraphrasing process to some lan-
guage phenomena and summarize the changes
in the information of the resultant paraphrases.
The following paraphrasing phenomena were
considered and related paraphrase pairs were
extracted.
3.1.1 Word Order
Word order in the spoken Chinese is com-
paratively free. In the paraphrase corpus,
quite a large proportion of the paraphrases
is created by word reordering. We extracted
the paraphrase pairs in which the morpheme
number of the original sentence is equal to that
of the paraphrase and each morpheme of the
original sentence appears in the paraphrase and
vice versa. One example is shown in 3-1.
[3−1] An extracted paraphrase pair.
Original: 42/VV 38/AD 555448/VV
50/P 39/PN 49/VA 47/SP
(Please call me again, could you?)
Paraphrase: 42/VV 38/AD 50/P 39/PN
555448/VV 49/VA 47/SP
Guided by the extracted paraphrase pair, we
can in fact paraphrase the original sentence by
reordering its words according to the word order
of the paraphrase. The extracted paraphrase
pairs of this kind provided instances for learning
word order paraphrasing patterns.
3.1.2 Negative Expressions
In some paraphrase pairs, we observed that
paraphrasing phenomena were related to
negative expressions. For example, original
sentences include negative words “75(do not
)” or “52(did not)” , but their corresponding
paraphrases appear as aﬃrmative forms with-
out these negative words. This fact implied
that the sentences could be simplified by delet-
ing the negative expressions. For this purpose,
the paraphrase pairs were extracted in which
the original sentences included the words “75”
or “52” and the corresponding paraphrases did
not. One example is shown in 3-2.
[3−2]
Original:�/VV�/AD��/VV 39/PN 53
/DEG?�/NN
(Do you know my telephone number?)
Paraphrase:��/VV 39/PN 53/DEG?�
/NN 47/SP
3.1.3 Expression of “32”
The Chinese language has a few grammatical
markers. The particle “76” is one of such mark-
ers. The sentences with the form “S(subject)
V(verb) O(object) C(complement)” may be
changed into the form “S 76 OVC”by
inserting the particle “76” (Zhang and Sato,
1999). The usage of “76” emphasizes the
object by moving it before the verb. When
the particle “76” is in a sentence, it is easier
to identify the object. So the insertion of “76”
will supply more information about syntactic
structure and reduce syntactic ambiguities.
Moreover, paraphrasing the sentences with
particle “76” may be more exact because the
identification of the object is more accurate.
We extracted the paraphrase pairs in which the
original sentences included the particle “76”
and the corresponding paraphrases did not.
See example 3-3 below.
[3−3]
Original: 34/DT 33/M 5132/NN 42/VV 45
/PN 4049/VV
(Could you fill out this form, please.)
Paraphrase: 42/VV 45/PN 58/BA 34/DT
33/M 5132/NN 4049/VV
(Could you make this form filled out, please.)
3.2 Automatic Generalization of
Instances
Then we attempted to generalize the extracted
instances in order to obtain paraphrasing pat-
terns. For each extracted paraphrase pair, the
original sentence is generalized to make the
matching part of the pattern, and the para-
phrase is generalized to make the generation
part of the pattern. The matching part spec-
ifies the components that will be paraphrased
as well as the context conditions. The genera-
tion part defines how to construct a paraphrase.
When the constituted pattern is applied to one
input sentence, if the input matches with the
matching part, a new sentence will be generated
according to the generation part.
In fact, the purpose of generalization is to
get a regular expression from the original sen-
tence and to get an operation expression con-
taining substitutions from the paraphrase. As
shown in 3-3, both the original sentence and the
paraphrase are series of morphemes, and each
morpheme consists of a part-of-speech and an
orthographic expression. The important thing
in paraphrasing is to maintain meaning. To
what extent the series of morphemes will be
generalized depends on each paraphrasing pair.
First, parts-of-speech keep the syntactic infor-
mation and therefore they should be kept. Sec-
ond, orthographic expressions of verbs, auxil-
iary verbs, adverbs, etc., are important in de-
ciding the main meaning of the sentence and
therefore they should also be kept. The ortho-
graphic expressions of other categories, such as
nouns, pronouns and numerals, can be general-
ized to an abstract level by replacing each or-
thographic expression with a wild card.
The pattern generalized from 3-3 is illus-
trated in 3-4. The left part is the matching part
and the right part is the generation part. The
lexical information may be an orthographic
expression or a variable represented by symbol
X
i
.X
i
in the matching part is in fact a
wild card, which means it can match with
any orthographic expression in the matching
operation. X
i
in the generation part defines a
substitution operation.
[3−4] A generalized pattern.
34/DT 33/M X
1
/NN 42/VV X
2
/PN 4049/VV
→42/VV X
2
/PN 58/BA 34/DT 33/M X
1
/NN
4049/VV
However, we found two problems in this kind
of automatic generalization. The first is that re-
strictions on the patterns generalized from long
sentences are too specific at the lexical level. In
fact, the clauses and noun phrases used as modi-
fiers have no eﬀect on the considered paraphras-
ing phenomena and can be generalized further.
The second is that some orthographic expres-
sions with important meanings are generalized
to wild cards, for instance, the numeral “7247
(how many)” may imply that the sentence is
interrogative. Therefore, a method is needed
to prevent some orthographic expressions from
being automatically replaced with wild cards.
3.3 Semi-Automatic Generalization of
Instances
Specifying which morphemes should be general-
ized and which orthographic expressions should
be kept requires human experience. In order to
integrate human experience into automatic gen-
eralization, we developed a semi-automatic gen-
eralization tool. The tool consists of description
symbols and a transformation program. The
description symbols are designed for people to
define generalization information on instances,
and the transformation program automatically
transforms the defined instances into patterns.
Three description symbols are defined as fol-
lows.
[]: This symbol is followed by a numeral
and is used to enclose a sequence of mor-
phemes. The enclosed part is a syntac-
tic component, e.g., a noun phrase or a
clause. Except for the part-of-speech of the
last morpheme, the enclosed part will be
replaced with a variable. In the Chinese
language, the syntactic property of a se-
quence of words is most likely reflected in
the last word, so we keep the part-of-speech
of the last morpheme. The enclosed parts
in the original sentence and the paraphrase
denoted by the same numerals will be re-
placed with the same variables.
{}: This symbol is used to enclose a mor-
pheme. The orthographic expression of the
morpheme will be kept. In this way, the
lexical information of morphemes can be
utilized to define the context. A few or-
thographic expressions can be defined in-
side one symbol so that words that can be
paraphrased in the same way can be stored
as one pattern.
〈〉: This symbol is used to enclose a mor-
pheme. The orthographic expression of the
morpheme will be replaced with a variable.
In this way, the orthographic expressions of
verbs or adverbs can also be generalized.
The usage of the symbols is explained in 3-5 and
3-6. Example 3-5 is a paraphrase pair in which
description symbols are defined. Example 3-6 is
the paraphrasing pattern generalized from 3-5.
[3−5] A defined instance.
Original: 42/VV 50/VV 39/PN 46/CD 〈56
/M〉 [4337/NN 53/DEG]
1
[3544/NN 4157/NN]
2
(Could you give me two copies of the Japanese
pamphlet, please?)
Paraphrase: [4337/NN 53/DEG]
1
[5236/NN
4157/NN]
2
42/VV 50/VV 39/PN 46/CD 〈56
/M〉
[3−6] The generalized pattern.
42/VV 50/VV X
1
/PN X
2
/CD X
3
/M Y
1
/DEG
Y
2
/NN → Y
1
/DEG Y
2
/NN 42/VV 50/VV
X
1
/PN X
2
/CD X
3
/M
X
i
has the same meaning as that of 3-4.
Y
1
/DEG in the matching part implies that it
can match with any sequence of morphemes in
which the part-of-speech of the last morpheme is
equal to DEG. Y
1
/DEG in the generation part
defines a substitution operation. Y
2
/NN im-
plies the same meaning, but the part-of-speech
of the last morpheme is equal to NN. In addition
to the automatic generalization for morphemes
of category PN and CD, the defined “〈56/M〉”
is also generalized to X
3
/M. The defined “[35
44(guide)/NN 4157/NN]
2
” in Original and “[52
36(tourist guide)/NN 4157/NN]
2
” in Paraphrase
are both generalized to Y
2
/NN, although they
are not exactly the same.
3.4 Construction of the Paraphrasing
Patterns
Using the developed tool, we manually de-
fined generalization information on the ex-
tracted paraphrase pairs and then obtained the
following four groups of paraphrasing patterns
through automatic transformation.
(1) 459 patterns of deleting negative expres-
sions.
(2) 160 patterns of inserting “76”.
(3) 160 patterns of deleting “76”.
(4) 2,030 patterns of reordering words.
The patterns of (3) were obtained by reversing
the matching part and the generation part of
each pattern of (2).
4 Design of the Paraphrasing
Process
In order to generate as many diﬀerent expres-
sions as possible, we designed a mechanism for
applying diﬀerent groups of paraphrasing pat-
terns. As described in Section 2, the paraphras-
ing process can be roughly classified into sim-
plification paraphrasing aimed at simplifying
expressions, and diversity paraphrasing aimed
at increasing variations. Bearing in mind that
simplification paraphrasing can reduce syntac-
tic and semantic ambiguities, we apply this type
of paraphrasing first, and then apply diversity
paraphrasing. Using this strategy, we antici-
pate that the accuracy of diversity paraphrasing
will be higher because there will be fewer am-
biguities in syntax and semantics. In the four
groups of patterns obtained above, group (1)
belongs to simplification paraphrasing, and the
other groups belong to diversity paraphrasing.
For one input sentence, the procedure for ap-
plying the diﬀerent groups of patterns is de-
signed as follows.
(1) Make the input sentence the application
data for all groups of patterns. Set group
number i =1.
(2) In the application of group i, get one pat-
tern from the group and repeat step (2.1)
to step (2.3).
(2.1) Match the input with the match-
ing part of the selected pattern. If
the matching succeeds, generate a sen-
tence according to the generation part
of the pattern.
(2.2) Make the generated sentence
the application data for all groups
j (i<j≤ 4). (At present there are
four groups of patterns.)
(2.3) Get another pattern then go to step
(2.1) until there are no patterns left in
group i.
(3) Set i = i + 1 and go to step (2) until i>4.
(4) When passing the generated sentences to
the transfer, do not pass duplicated ones.
Using this procedure, the generated paraphrases
can be passed to the transfer at any time of
the paraphrasing process. If one of the para-
phrases can be translated by the transfer, the
paraphrasing process will be stopped. In addi-
tion, the generated paraphrases can be para-
phrased further by the patterns of following
groups, therefore more expressions are likely to
be produced. Based on this design, a para-
phraser was implemented.
5 Experiment and Evaluation
A paraphrasing experiment was carried out on
the paraphrase corpus using the implemented
paraphraser and the obtained patterns. In or-
der to get the same eﬀect as that of using open
test data, each pattern was not applied to the
sentence from which the pattern was general-
ized. For 45,110 test sentences, 4,908 test sen-
tences (about 10.9%) were paraphrased. From
the 4,908 test sentences, 8,183 paraphrases were
generated and the average number of para-
phrases for one test sentence was 1.66. The gen-
erated paraphrases were evaluated by Chinese
natives from two viewpoints, i.e., naturalness
and meaning-retaining, with their correspond-
ing test sentences. As a result, 7,226 generated
paraphrases were correct and an 88% accuracy
was achieved. The experimental result is shown
in Table 2.
Table 2. Result of Paraphrasing Experiment
# Test Sentences 45,110
# Paraphrased Test Sentences 4,908 (10.9%)
# Generated Paraphrases 8,183
# Correct Paraphrases 7,226 (88%)
Three examples of the paraphrasing results
are given below.
[5−1]
Input: 54755478457168353969776153?
(Could you reserve the earliest plane for me?)
Paraphrase 1: 5478457168353969776157?
Paraphrase 2: 584278457168353969776157?
[5−2]
Input: 383673584271637057
(May I reserve a restaurant here?)
Paraphrase 1: 584238367371637057
Paraphrase 2: 383673716370584257
Paraphrase 3: 637038367358427157
Paraphrase 4: 637058423836737157
Paraphrase 5: 637038367371584257
[5−3]
Input: 65605156696462496745.
(A room with a nice view, please give me.)
Paraphrase 1: 4976656051566964626745.
(Please arrange a room with a nice view for
me.)
Paraphrase 2: 45444065605156696462.
(I would like a room with a nice view.)
Paraphrase 3: 49674565605156696462.
(Please give me a room with a nice view.)
Paraphrase 4: 64624967456560515669.
(As for room, please give me one with a nice
view.)
In the Input of 5-1 there is an expression of
repeated interrogation “547554” that consists of
an aﬃrmation “54(can)” and a negation “7554
(can not)”. After the application of the patterns
of deleting negative expressions, Paraphrase 1
and Paraphrase 2 were generated. Both para-
phrases are in aﬃrmative form and both are cor-
rect. From the Input of 5-2, five paraphrases
were generated only by reordering words. Para-
phrases 1, 2, 3, 4 and 5 are all correct. In the
Input of 5-3, the order of the predicate “496745”
and the object “65605156696462” is inverted. Af-
ter the patterns of inserting “76” were applied,
Paraphrase 1 was obtained. Then, the patterns
of deleting “76” were applied to this generated
paraphrase and Paraphrase 2 and Paraphrase
3 were obtained. In Paraphrase 3 the common
word order was recovered. Finally, the patterns
of reordering words were applied to Paraphrases
1, 2, 3 and the input. Paraphrase 4 was ob-
tained from Paraphrase 3. Paraphrases 1, 2, 3
and 4 are all correct.
From the experimental results we see that the
proposed approach can realize the goal of sim-
plifying the expressions of the inputs and in-
creasing variations with a high level of accu-
racy. If one of the paraphrased results can be
translated, we can say that the paraphraser is
eﬀective in the translation system.
Through the analysis of wrong results, we
found two reasons for paraphrasing errors. One
reason is that some constituents or modification
relations are incorrectly recognized based on the
obtained paraphrasing patterns. For example,
when pattern 3-6 was applied to the sentence
“49674555665033696462(A room for two people,
please.)”, the quantity phrase “5566(two)” was
wrongly recognized as modifying “5033696462(a
room where people can live)”, whereas it in fact
modifies the noun “50(people)”. Because of
this wrong recognition, the generated sentence
was “50336964624967455566 (Please give me two
rooms where people can live.)”. Its meaning is
diﬀerent from the input and therefore the result
is not correct. The other reason for paraphras-
ing errors is that there were errors in part-of-
speech tagging. For example, the word “67”in
“49674579344143365946 (Please gift-wrap this)”
was tagged as a verb while it really acts as a
preposition. The wrong tagging resulted in the
wrong application of patterns.
6 Conclusion
In this paper, a pattern-based approach to the
paraphrasing of Chinese utterances is proposed
and a method of constructing paraphrasing pat-
terns from a corpus is described. Based on the
proposed approach and method, a paraphraser
is implemented and four types of paraphrasing
patterns are constructed. Also, a paraphrasing
experiment is conducted and experimental re-
sults are reported. The proposed approach has
the following advantages.
(1) Because only morphological analysis is re-
quired, it is easy to implement the para-
phraser and the processing time is short.
(2) By using the developed semi-automatic
generalization tool, paraphrasing patterns
can be eﬃciently learned from a paraphras-
ing corpus and human experience. The pat-
terns enhanced by human experience have
a higher accuracy.
(3) The classification of paraphrasing phenom-
ena in pattern learning makes it possible to
restrict the paraphrasing process to some
language phenomena. The mechanism of
applying diﬀerent types of patterns empha-
sizes how to raise the accuracy of para-
phrasing and how to increase variations.
In this research, only four types of paraphras-
ing phenomena are involved. The coverage
achieved using the current patterns is still
low. In the next phase, we are going to use
the proposed approach on other paraphrasing
phenomena in order to be able to paraphrase
more Chinese utterances.
Acknowledgement
This research was supported in part by the
Telecommunications Advancement Organiza-
tion of Japan. We would also like to thank Ms.
Lan Yao for her help in our experiments.

References

Akira Kataoka, Shigeru Masuyama, and
Kazuhide Yamamoto. 1999. Summarization
by Shortening a Japanese Noun Modifier into
Expression “A no B”. In Proc. of NLPRS’99,
pages 409–414.

Satoshi Sato. 1999. Automatic Paraphrase of
Technical Papers. Transactions of Informa-
tion Processing Society of Japan, 40(7):2937–
2945. (in Japanese).

Satoshi Shirai, Satoru Ikehara, and Tsukasa
Kawaoka. 1993. Eﬀects of Automatic Rewrit-
ing of Source Language within a Japanese to
English MT System. In Proc. of TMI ’93,
pages 226–239.

Fei Xia. 2000. The Part-of-speech Tag-
ging Guideline for the Penn Chi-
nese Treebank (3.0). Available at
http://www.ldc.upenn.edu/ctb.

Kazuhide Yamamoto, Satoshi Shirai, Masashi
Sakamoto, and Yujie Zhang. 2001. Sand-
glass: Twin Paraphrasing Spoken Language
Translation. In Proc. of ICCPOL’01, pages
154–159.

Kazuhide Yamamoto. 2002. Machine Transla-
tion by Interaction between Paraphraser and
Transfer. In Proc. of Coling 2002.

Li Zhang and Haruhiko Sato. 1999. Chi-
nese Expression Grammar-28 Points. Toho-
Shoten. (in Japanese).

Yujie Zhang and Kazuhide Yamamoto. 2001.
Analysis of Chinese Spoken Language for
Automatic Paraphrasing. In Proc. of IC-
CPOL’01, pages 290–293.

Yujie Zhang, Chengqing Zong, Kazuhide
Yamamoto, and Masashi Sakamoto. 2001.
Paraphrasing Utterances by Reordering
Words Using Semi-Automatically Acquired
Patterns. In Proc. of NLPRS’01, pages
195–202.

Chengqing Zong, Yujie Zhang, Kazuhide Ya-
mamoto, Masashi Sakamoto, and Satoshi Shi-
rai. 2001. Approach to Spoken Chinese Para-
phrasing Based on Feature Extraction. In
Proc. of NLPRS’01, pages 551–556.
