哑炮
What Is a Paraphrase?
∗
Rahul Bhagat
USC Information Sciences Institute
∗∗
Eduard Hovy
USC Information Sciences Institute
Paraphrases are sentences or phrases that convey the same meaning using different wording.
Although the logical definition of paraphrases requires strict semantic equivalence, 语言学
accepts a broader, 近似, equivalence—thereby allowing far more examples of “quasi-
paraphrase.” But approximate equivalence is hard to define. 因此, the phenomenon of para-
短语, as understood in linguistics, is difficult to characterize. 在本文中, we list a set
的 25 operations that generate quasi-paraphrases. We then empirically validate the scope and
accuracy of this list by manually analyzing random samples of two publicly available paraphrase
语料库. We provide the distribution of naturally occurring quasi-paraphrases in English text.
1. 介绍
Sentences or phrases that convey the same meaning using different wording are called
paraphrases. 例如, consider sentences (1) 和 (2):
(1)
(2)
The school said that their buses seat 40 students each.
The school said that their buses accommodate 40 students each.
Paraphrases are of interest for many current NLP tasks, including textual entailment,
machine reading, question answering, information extraction, and machine translation.
Whenever the text contains multiple ways of saying “the same thing,” but the applica-
tion requires the same treatment of those various alternatives, an automated paraphrase
recognition mechanism would be useful.
One reason why paraphrase recognition systems have been difficult to build is
because paraphrases are hard to define. Although the strict interpretation of the
term “paraphrase” is quite narrow because it requires exactly identical meaning,
in linguistics literature paraphrases are most often characterized by an approxi-
mate equivalence of meaning across sentences or phrases. De Beaugrande and Dressler
(1981, 页 50) define paraphrases as “approximate conceptual equivalence among
∗ 24515 SE 46th Terrace Issaquah, WA 98029. 电子邮件: me@rahulbhagat.net.
∗∗ 24515 SE 46th Terrace Issaquah, WA 98029. 电子邮件: hovy@isi.edu.
提交材料已收到: 5 七月 2012; revised submission received: 21 一月 2013; 接受出版:
6 行进 2013.
土井:10.1162/大肠杆菌a 00166
© 2013 计算语言学协会
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
计算语言学
体积 39, 数字 3
outwardly different material.” Hirst (2003, slide 9) defines paraphrases as “talk(英)
about the same situation in a different way.” He argues that paraphrases aren’t fully
synonymous: There are pragmatic differences in paraphrases, 即, difference of eval-
uation, connotation, viewpoint, 等等. According to Mel’cuk (2012, 页 7) “An
approximate synonymy of sentences is considered as sufficient for them to be produced
from the same SemS.” He further adds that approximate paraphrases include implica-
系统蒸发散 (not in the logical sense, but in the everyday sense). Taking an extreme view, 克拉克
(1992, 页 172) rejects the idea of absolute synonymy by saying “Every two forms (在
语言) contrast in meaning.” Overall, there is a large body of work in the linguistics
literature that argues that paraphrases are not restricted to strict synonymy.
在本文中, we take a broad view of paraphrases. To avoid the conflict between
the notion of strict paraphrases as understood in logic and the broad notion in linguis-
抽动症, we use the term quasi-paraphrases to refer to the paraphrases that we deal with here.
In the context of this article, the term “paraphrases” (even without the prefix “quasi”)
means “quasi-paraphrases.” We define quasi-paraphrases as ‘sentences or phrases that
convey approximately the same meaning using different words.’ We ignore the fine
grained distinctions of meaning between sentences and phrases, introduced due to the
speaker’s evaluation of the situation, connotation of the terms used, change of modality,
等等. 例如, consider sentences (3) 和 (4).
(3)
(4)
The school said that their buses seat 40 students each.
The school said that their buses cram in 40 students each.
这里, seat and cram in are not synonymous: They carry different evaluations by the
speaker about the same situation. 我们, 然而, consider sentences (3) 和 (4) 成为
(quasi) paraphrases. 相似地, consider sentences (5) 和 (6).
(5)
(6)
The school said that their buses seat 40 students each.
The school is saying that their buses might accommodate 40 students each.
这里, said and is saying have different tenses. 还, might accommodate and seat are not
synonymous, due to the modal verb might. We consider sentences (5) 和 (6) 成为
quasi-paraphrases, 然而.
Note that this article focuses on defining quasi-paraphrases. It does not provide
direct implementation/application results of using them. 我们相信, 然而, 那
this work will allow computation-oriented researchers to focus their future work more
effectively on a subset of paraphrase types without concern for missing important
材料, and it will provide linguistics-oriented researchers with a blueprint of the
overall distribution of the types of paraphrase.
2. Paraphrasing Phenomena Classified
Although approximate equivalence is hard to characterize, it is not a completely un-
structured phenomenon. By studying various existing paraphrase theories—Mel’cuk
(2012), 哈里斯 (1981), Honeck (1971)—and through an analysis of paraphrases obtained
from two different corpora, we have discovered that one can identify a set of 25 类
of quasi-paraphrases, with each class having its own specific way of relaxing the re-
quirement of strict semantic equivalence. 在这个部分, we define and describe these
类.
464
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
Bhagat and Hovy
What Is a Paraphrase?
The classes described here categorize quasi-paraphrases from the lexical perspec-
主动的. The lexical perspective defines paraphrases in terms of the kinds of lexical changes
that can take place in a sentence/phrase resulting in the generation of its paraphrases.
1. Synonym substitution: Replacing a word/phrase by a synonymous word/phrase,
in the appropriate context, results in a paraphrase of the original sentence/phrase. 这
category covers the special case of genitives, where the clitic ’s is replaced by other
genitive indicators like of, 的, 等等. This category also covers near-synonymy,
那是, it allows for changes in evaluation, connotation, 等等, of words or phrases
between paraphrases. 例子:
(A)
(乙)
Google bought YouTube. ⇔ Google acquired YouTube.
Chris is slim. ⇔ Chris is slender. ⇔ Chris is skinny.
2. Antonym substitution: Replacing a word/phrase by its antonym accompanied by
a negation or by negating some other word, in the appropriate context, results in a
paraphrase of the original sentence/phrase. This substitution may be accompanied by
the addition/deletion of appropriate function words. 例子:
(A)
Pat ate. ⇔ Pat did not starve.
3. Converse substitution: Replacing a word/phrase with its converse and inverting
the relationship between the constituents of a sentence/phrase, in the appropriate
语境, results in a paraphrase of the original sentence/phrase, presenting the sit-
uation from the converse perspective. This substitution may be accompanied by the
addition/deletion of appropriate function words and sentence restructuring. 例子:
(A)
Google bought YouTube. ⇔ YouTube was sold to Google.
4. Change of voice: Changing a verb from its active to passive form and vice versa re-
sults in a paraphrase of the original sentence/phrase. This change may be accompanied
by the addition/deletion of appropriate function words and sentence restructuring.
This often generates the most strictly meaning-preserving paraphrase. 例子:
(A)
Pat loves Chris. ⇔ Chris is loved by Pat.
5. Change of person: Changing the grammatical person of a referenced object results in
a paraphrase of the original sentence/phrase. This change may be accompanied by the
addition/deletion of appropriate function words. 例子:
(A)
Pat said, “I like football.” ⇔ Pat said that he liked football.
6. Pronoun/Co-referent substitution: Replacing a pronoun by the noun phrase it
co-refers with results in a paraphrase of the original sentence/phrase. This also often
generates the most strictly meaning-preserving paraphrase. 例子:
(A)
Pat likes Chris, because she is smart. ⇔ Pat likes Chris, because Chris
is smart.
465
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
计算语言学
体积 39, 数字 3
7. Repetition/Ellipsis: Ellipsis or elliptical construction results in a paraphrase of the
original sentence/phrase. 相似地, this often generates the most strictly meaning-
preserving paraphrase. 例子:
(A)
Pat can run fast and Chris can run fast, 也. ⇔ Pat can run fast and Chris
能, 也.
8. Function word variations: Changing the function words in a sentence/phrase with-
out affecting its semantics, in the appropriate context, results in a paraphrase of the
original sentence/phrase. This can involve replacing a light verb by another light verb,
replacing a light verb by copula, replacing certain prepositions with other prepositions,
replacing a determiner by another determiner, replacing a determiner by a preposition
and vice versa, and addition/removal of a preposition and/or a determiner. 例子:
(A)
(乙)
Results of the competition have been declared. ⇔ Results for the
competition have been declared.
Pat showed a nice demo. ⇔ Pat’s demo was nice.
9. Actor/Action substitution: Replacing the name of an action by a word/phrase denot-
ing the person doing the action (actor) and vice versa, in the appropriate context, 结果
in a paraphrase of the original sentence/phrase. This substitution may be accompanied
by the addition/deletion of appropriate function words. 例子:
(A)
I dislike rash drivers. ⇔ I dislike rash driving.
10. Verb/“Semantic-role noun” substitution: Replacing a verb by a noun correspond-
ing to the agent of the action or the patient of the action or the instrument used for
the action or the medium used for the action, in the appropriate context, results in
a paraphrase of the original sentence/phrase. This substitution may be accompanied
by the addition/deletion of appropriate function words and sentence restructuring.
例子:
(A)
(乙)
(C)
Pat teaches Chris. ⇔ Pat is Chris’s teacher.
Pat teaches Chris. ⇔ Chris is Pat’s student.
Pat tiled his bathroom floor. ⇔ Pat installed tiles on his bathroom floor.
11. Manipulator/Device substitution: Replacing the name of a device by a word/
phrase denoting the person using the device (manipulator) and vice versa, 在里面
appropriate context, results in a paraphrase of the original sentence/phrase. 这
substitution may be accompanied by the addition/deletion of appropriate function
字. 例子:
(A)
The pilot took off despite the stormy weather. ⇔ The plane took off despite
the stormy weather.
12. General/Specific substitution: Replacing a word/phrase by a more general or more
specific word/phrase, in the appropriate context, results in a paraphrase of the original
466
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
Bhagat and Hovy
What Is a Paraphrase?
sentence/phrase. This substitution may be accompanied by the addition/deletion of ap-
propriate function words. Hypernym/hyponym substitution is a part of this category.
This often generates a quasi-paraphrase. 例子:
(A)
(乙)
I dislike rash drivers. ⇔ I dislike rash motorists.
Pat is flying in this weekend. ⇔ Pat is flying in this Saturday.
13. Metaphor substitution: Replacing a noun by its standard metaphorical use and
vice versa, in the appropriate context, results in a paraphrase of the original sentence/
短语. This substitution may be accompanied by the addition/deletion of appropriate
function words. 例子:
(A)
(乙)
I had to drive through fog today. ⇔ I had to drive through a wall of fog
今天.
Immigrants have used this network to send cash. ⇔ Immigrants have used
this network to send stashes of cash.
14. Part/Whole substitution: Replacing a part by its corresponding whole and vice
versa, in the appropriate context, results in a paraphrase of the original sentence/
短语. This substitution may be accompanied by the addition/deletion of appropriate
function words. 例子:
(A)
American airplanes pounded the Taliban defenses. ⇔ American airforce
pounded the Taliban defenses.
15. Verb/Noun conversion: Replacing a verb by its corresponding nominalized noun
form and vice versa, in the appropriate context, results in a paraphrase of the original
sentence/phrase. This substitution may be accompanied by the addition/deletion of
appropriate function words and sentence restructuring. 例子:
(A)
(乙)
The police interrogated the suspects. ⇔ The police subjected the suspects to
an interrogation.
The virus spread over two weeks. ⇔ Two weeks saw a spreading of the
病毒.
16. Verb/Adjective conversion: Replacing a verb by the corresponding adjective form
and vice versa, in the appropriate context, results in a paraphrase of the original
sentence/phrase. This substitution may be accompanied by the addition/deletion of
appropriate function words and sentence restructuring. 例子:
(A)
Pat loves Chris. ⇔ Chris is lovable to Pat.
17. Verb/Adverb conversion: Replacing a verb by its corresponding adverb form and
vice versa, in the appropriate context, results in a paraphrase of the original sentence/
短语. This substitution may be accompanied by the addition/deletion of appropriate
function words and sentence restructuring. 例子:
(A)
Pat boasted about his work. ⇔ Pat spoke boastfully about his work.
467
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
计算语言学
体积 39, 数字 3
18. Noun/Adjective conversion: Replacing a verb by its corresponding adjective form
and vice versa, in the appropriate context, results in a paraphrase of the original
sentence/phrase. This substitution may be accompanied by the addition/deletion of
appropriate function words and sentence restructuring. 例子:
(A)
I’ll fly by the end of June. ⇔ I’ll fly late June.
19. Verb-preposition/Noun substitution: Replacing a verb and a preposition denoting
location by a noun denoting the location and vice versa, in the appropriate context,
results in a paraphrase of the original sentence/phrase. This substitution may be
accompanied by the addition/deletion of appropriate function words and sentence
restructuring. 例子:
(A)
The finalists will play in Giants stadium. ⇔ Giants stadium will be the
playground for the finalists.
20. Change of tense: Changing the tense of a verb, in the appropriate context, 结果
in a paraphrase of the original sentence/phrase. This change may be accompanied
by the addition/deletion of appropriate function words. This often generates a quasi-
paraphrase, although it might be semantically less accurate than many other quasi-
paraphrases. 例子:
(A)
Pat loved Chris. ⇔ Pat loves Chris.
21. Change of aspect: Changing the aspect of a verb, in the appropriate context, 结果
in a paraphrase of the original sentence/phrase. This change may be accompanied by
the addition/deletion of appropriate function words. 例子:
(A)
Pat is flying in today. ⇔ Pat flies in today.
22. Change of modality: Addition/deletion of a modal or substitution of one modal
by another, in the appropriate context, results in a paraphrase of the original sen-
tence/phrase. This change may be accompanied by the addition/deletion of appro-
priate function words. This often generates a quasi-paraphrase, although it might be
semantically less accurate than many other quasi-paraphrases. 例子:
(A)
(乙)
Google must buy YouTube. ⇔ Google bought YouTube.
The government wants to boost the economy. ⇔ The government hopes to
boost the economy.
23. Semantic implication: Replacing a word/phrase denoting an action, 事件, 所以
向前, by a word/phrase denoting its possible future effect, in the appropriate context,
results in a paraphrase of the original sentence/phrase. This may be accompanied by
the addition/deletion of appropriate function words and sentence restructuring. 这
often generates a quasi-paraphrase. 例子:
Google is in talks to buy YouTube. ⇔ Google bought YouTube.
The Marines are fighting the terrorists. ⇔ The Marines are eliminating
the terrorists.
(A)
(乙)
468
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
Bhagat and Hovy
What Is a Paraphrase?
24. Approximate numerical equivalences: Replacing a numerical expression (a word/
phrase denoting a number, often with a unit) by an approximately equivalent nu-
merical expression (even perhaps with change of unit), in the appropriate context,
results in a paraphrase of the original sentence/phrase. This often generates a quasi-
paraphrase. 例子:
(A)
(乙)
At least 23 我们. soldiers were killed in Iraq last month. ⇔ About 25 我们.
soldiers were killed in Iraq last month.
Disneyland is 32 miles from here. ⇔ Disneyland is around 30 minutes
from here.
25. External knowledge: Replacing a word/phrase by another word/phrase based on
extra-linguistic (世界) 知识, in the appropriate context, results in a paraphrase
of the original sentence/phrase. This may be accompanied by the addition/deletion of
appropriate function words and sentence restructuring. This often generates a quasi-
paraphrase, although in some cases preserves meaning exactly. 例子:
(A) We must work hard to win this election. ⇔ The Democrats must work hard
to win this election.
The government declared victory in Iraq. ⇔ Bush declared victory in Iraq.
(乙)
3. Analysis of Paraphrases
在部分 2, we presented a list of lexical changes that define quasi-paraphrases. 在这个
部分, we seek to validate the scope and accuracy of this list. Our analysis uses two
criteria:
1. Distribution: What is the distribution of each of these lexical changes in a paraphrase
语料库?
2. Human judgment: If one uses each of the lexical changes, on applicable sentences,
how often do each of these changes generate acceptable quasi-paraphrases?
3.1 Distribution
We used the following procedure to measure the distribution of the lexical changes:
1. We downloaded paraphrases from two publicly available data sets containing
sentence-level paraphrases: the Multiple-Translations Corpus (MTC) (黄, Graff,
and Doddington 2002) and the Microsoft Research (MSR) paraphrase corpus (Dolan,
Quirk, and Brockett 2004). The paraphrase pairs come with their equivalent parts
manually aligned (Cohn, Callison-Burch, 和拉帕塔 2008).
2. We selected 30 sentence-level paraphrase pairs from each of these corpora at random
and extracted the corresponding aligned and unaligned phrases.1 This resulted in 210
phrase pairs for the MTC corpus and 145 phrase pairs for the MSR corpus.
1 We assume that any unaligned phrase is paired with a null phrase and we discard it prior to the analysis.
469
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
计算语言学
体积 39, 数字 3
3. We labeled each of the phrase pairs with the appropriate lexical changes defined in
部分 2. If any phrase pair could not be labeled by a lexical change from Section 2, 我们
labeled it as unknown.
4. We finally calculated the distribution of each label (lexical change), over all the labels,
for each corpus. 桌子 1 shows the percentage distribution of the lexical changes in the
MTC (柱子 3) and MSR corpora (柱子 4).
3.2 Human Judgment
在这个部分, we explain the procedure we used to obtain the human judgments of the
changes that define paraphrases from the lexical perspective:
1. We randomly selected two words or phrases from publicly available resources (的-
pending on the lexical change) for each of the lexical operations from Section 2 (除了
external knowledge). 例如, to obtain words for synonym substitution, we used
WordNet (Fellbaum 1998) (and selected a word, say buy); to obtain implication rules
for semantic implication, we used the DIRT resource (Lin and Pantel 2001); 等等.
桌子 1
Distribution and Precision of paraphrases. Distribution may not sum to 100% due to rounding.
#
类别
% Distribution MTC % Distribution MSR % Precision
1. Synonym substitution
2. Antonym substitution
3. Converse substitution
4. Change of voice
5. Change of person
6. Pronoun/Co-referent
substitution
7. Repetition/Ellipsis
8. Function word variations
9. Actor/Action substitution
10. Verb/“Semantic-role noun”
substitution
11. Manipulator/Device substitution
12. General/Specific substitution
13. Metaphor substitution
14. Part/Whole substitution
15. Verb/Noun conversion
16. Verb/Adjective conversion
17. Verb/Adverb conversion
18. Noun/Adjective conversion
19. Verb-preposition/
Noun substitution
20. Change of tense
21. Change of aspect
22. Change of modality
23. Semantic implication
24. Approximate numerical
equivalences
25. External knowledge
26. Unknown
470
37
0
1
1
0
1
4
37
0
1
0
4
0
0
2
1
0
0
0
4
1
1
1
0
6
0
19
0
0
1
1
1
4
30
0
0
0
3
1
0
3
0
0
0
0
1
0
0
4
2
32
0
95
65
75
85
80
70
100
85
75
60
30
80
60
65
100
55
65
80
65
70
95
80
70
95
95
NA
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
Bhagat and Hovy
What Is a Paraphrase?
2. For each selected word or phrase, we obtained five random sentences from the Giga-
word corpus. These sentences were manually checked to make sure that they contained
the intended sense of the word or phrase. This gave us a total of 10 sentences for each
现象. 例如, for the word buy, one of the selected sentences might be:
(A)
They want to buy a house.
3. For each sentence selected in step 2, we applied the corresponding lexical changes to
the word or phrase selected in step 1 to generate a potential paraphrase.2 For example,
we might apply synonym substitution to sentence (A) and replace the word buy with its
WordNet synonym purchase. This will result in the following sentence:
(乙)
They want to purchase a house.
4. For the phenomenon of external knowledge, we randomly sampled a total of 10 sen-
tence pairs from the MTC and MSR corpora, such that the pairs were paraphrases based
on external knowledge.
5. We gave the sentence pairs to two annotators and asked them to annotate them as
either paraphrases or non-paraphrases. 例如, the annotator might be given the
sentence pair (A) 和 (乙) and she/he might annotate this pair as paraphrases.
6. We used the annotations from each of the annotators to calculate the precision per-
centage for each lexical change. The final precision score was calculated as the average
of the precision scores obtained from the two annotations. 桌子 1 shows the percentage
precision (柱子 5) of lexical changes in this test corpus.
7. We finally calculated the kappa statistic (Siegal and Castellan Jr. 1988) to measure the
注释者间协议. A kappa score of κ = 0.66 was obtained on the annotation
任务.
4. 结论
A definition of what phenomena constitute paraphrases and what do not has been a
problem in the past. Whereas some people have used a very narrow interpretation
of paraphrases—paraphrases must be exactly logically equivalent—others have taken
broader perspectives that consider even semantic implications to be acceptable para-
短语. 据我们所知, outside of specific language interpretation frame-
作品 (like Meaning Text Theory [Mel’cuk 1996]), no one has tried to create a general,
exhaustive list of the transformations that define paraphrases. In this article we provide
such a list. We have also tried to empirically quantify the distribution and accuracy of
列表. It is notable that certain types of quasi-paraphrases dominate whereas others
are very rare. We also observed, 然而, that the dominating transformations vary
based on the type of paraphrase corpus used, thus indicating the variety of behavior
exhibited by the paraphrases. Based on the large variety of possible transformations that
can generate paraphrases, its seems likely that the kinds of paraphrases that are deemed
useful would depend on the application at hand. This might motivate the creation of
2 The words in the new sentence were allowed to be reordered (permuted) if needed and only function
字 (and no content words) were allowed to be added to the new sentence.
471
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
计算语言学
体积 39, 数字 3
application-specific lists of the kinds of allowable paraphrases and the development of
automatic methods to distinguish the different kinds of paraphrases.
致谢
The authors wish to thank Jerry Hobbs and
anonymous reviewers for valuable
comments and feedback.
参考
克拉克, 乙. V. 1992. Conventionality and
contrasts: Pragmatic principles with
lexical consequences. In Andrienne Lehrer
and Eva Feder Kittay, 编辑, Frame,
Fields, and Contrasts: New Essays in
Semantic Lexical Organization. 劳伦斯
埃尔鲍姆联合公司, Hillsdale, 新泽西州,
pages 171–188.
Cohn, T。, C. Callison-Burch, 和M. 警告.
2008. Constructing corpora for the
development and evaluation of
paraphrase systems. 计算型
语言学, 34(4):597–614.
De Beaugrande, 右. 和W. V. Dressler.
1981. Introduction to Text Linguistics.
朗文, 纽约, 纽约.
Dolan, B., C. Quirk, 和C. Brockett.
2004. Unsupervised construction of
large paraphrase corpora: Exploiting
massively parallel news sources.
In Proceedings of the Conference on
计算语言学 (科林),
pages 350–357, 日内瓦.
Fellbaum, C. 1998. An Electronic Lexical
Database. 与新闻界, 剑桥, 嘛.
哈里斯, Z. 1981. Co-occurence and
transformation in linguistic structure.
In Henry Hiz, 编辑, Papers on Syntax.
D. Reidel Publishing Co., 多德雷赫特,
pages 143–210. First published in 1957.
Hirst, G. 2003. Paraphrasing paraphrased.
Invited talk at the ACL International
Workshop on Paraphrasing, Sapporo.
Honeck, Richard P. 1971. A study of
paraphrases. Journal of Verbal Learning
and Verbal Behavior, 10(4):367–381.
黄, S。, D. Graff, and G. Doddington.
2002. Multiple-translation Chinese
语料库. Linguistic Data Consortium,
费城, PA.
林, D. 和P. 潘特尔. 2001. Dirt: Discovery of
inference rules from text. In ACM SIGKDD
International Conference on Knowledge
Discovery and Data Mining, pages 323–328,
旧金山, CA.
Mel’cuk, 我. 1996. Lexical functions: A tool for
description of lexical relations in a lexicon.
In Leo Wanner, 编辑, Lexical Functions
in Lexicography and Natural Language
加工. John Benjamins Publishing Co.,
费城, PA, pages 37–102.
Mel’cuk, 我。, 2012. 语义学: From Meaning
to Text. John Benjamins Publishing Co.,
费城, PA.
西格尔, S. and N. J. Castellan, 少年. 1988.
Nonparametric Statistics for the Behavioral
科学. 麦格劳-希尔, Columbus, 哦.
472
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
C
哦
我
我
/
我
A
r
t
我
C
e
–
p
d
F
/
/
/
/
3
9
3
4
6
3
1
8
0
1
9
1
2
/
C
哦
我
我
_
A
_
0
0
1
6
6
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3
下载pdf