Application-driven Statistical Paraphrase Generation
Zhao, Shiqi and Lan, Xiang and Liu, Ting and Li, Sheng

Article Structure

Abstract

Paraphrase generation (PG) is important in plenty of NLP applications.

Introduction

Paraphrases are alternative ways that convey the same meaning.

Related Work

Conventional methods for paraphrase generation can be classified as follows:

Statistical Paraphrase Generation

3.1 Differences between SPG and SMT

Experimental Setup

Our SPG decoder is developed by remodeling Moses that is widely used in SMT (Hoang and Koehn, 2008).

Results and Analysis

We use our method to generate paraphrases for the three applications.

Conclusions and Future Work

This paper proposes a method for statistical paraphrase generation.

Topics

sentence compression

Appears in 14 sentences as: Sentence compression (3) sentence compression (12)
In Application-driven Statistical Paraphrase Generation
  1. We consider three paraphrase applications in our experiments, including sentence compression , sentence simplification, and sentence similarity computation.
    Page 1, “Introduction”
  2. On the contrary, SPG has distinct purposes in different applications, such as sentence compression , sentence simplification, etc.
    Page 2, “Statistical Paraphrase Generation”
  3. The application in this example is sentence compression .
    Page 3, “Statistical Paraphrase Generation”
  4. Paraphrase application: sentence compression
    Page 4, “Statistical Paraphrase Generation”
  5. We consider three applications, including sentence compression , simplification, and similarity computation.
    Page 4, “Statistical Paraphrase Generation”
  6. o Sentence compression : Sentence compression2 is important for summarization, subtitle generation, and displaying texts in small screens such as cell phones.
    Page 4, “Statistical Paraphrase Generation”
  7. Results show that the percentages of test sentences that can be paraphrased are 97.2%, 95.4%, and 56.8% for the applications of sentence compression , simplification, and similarity computation, respectively.
    Page 6, “Results and Analysis”
  8. Further results show that the average number of unit replacements in each sentence is 5.36, 4.47, and 1.87 for sentence compression , simplification, and similarity computation.
    Page 6, “Results and Analysis”
  9. A source sentence s is paraphrased in each application and we can see that: (l) for sentence compression , the paraphrase t is 8 bytes shorter than s; (2) for sentence simplification, the words wealth and part in t are easier than their sources asset and proportion, especially for the nonnative speakers; (3) for sentence similarity computation, the reference sentence s’ is listed below t, in which the words appearing in t but not in s are highlighted in blue.
    Page 7, “Results and Analysis”
  10. After analyzing the results, we find that 24.95%, 8.79%, and 7.16% of the paraphrases achieve sentence compression , simplification, and similarity computation, respectively, which are much lower than our method.
    Page 7, “Results and Analysis”
  11. Previous research regarded sentence compression , simplification, and similarity computation as totally different problems and proposed distinct method for each one.
    Page 7, “Results and Analysis”

See all papers in Proc. ACL 2009 that mention sentence compression.

See all papers in Proc. ACL that mention sentence compression.

Back to top.

language model

Appears in 7 sentences as: Language Model (1) language model (7)
In Application-driven Statistical Paraphrase Generation
  1. Our SPG model contains three sub-models: a paraphrase model, a language model , and a usability model, which control the adequacy, fluency,
    Page 3, “Statistical Paraphrase Generation”
  2. Language Model: We use a trigram language model in this work.
    Page 4, “Statistical Paraphrase Generation”
  3. The language model based score for the paraphrase t is computed as:
    Page 4, “Statistical Paraphrase Generation”
  4. where J is the length of t, 253- is the j-th word of t, and Alm is the weight for the language model .
    Page 4, “Statistical Paraphrase Generation”
  5. Here, the frequency of a unit is measured using the language model mentioned above3.
    Page 4, “Statistical Paraphrase Generation”
  6. 3To compute the language model based score, the matched patterns are instantiated and the matched collocations are connected with words between them.
    Page 4, “Statistical Paraphrase Generation”
  7. The language model is trained using a 9 GB English corpus.
    Page 6, “Experimental Setup”

See all papers in Proc. ACL 2009 that mention language model.

See all papers in Proc. ACL that mention language model.

Back to top.

development set

Appears in 6 sentences as: development set (7)
In Application-driven Statistical Paraphrase Generation
  1. = cdev(+7“)/cdev(7“), where cdev(7“) is the total number of unit replacements in the generated paraphrases on the development set .
    Page 5, “Statistical Paraphrase Generation”
  2. Replacement rate (rr): rr measures the paraphrase degree on the development set , i.e., the percentage of words that are paraphrased.
    Page 5, “Statistical Paraphrase Generation”
  3. We define rr as: 77 = wdev(7“)/wdev(s), where wdev(7“) is the total number of words in the replaced units on the development set, and wdev (s) is the number of words of all sentences on the development set .
    Page 5, “Statistical Paraphrase Generation”
  4. For each application, we first ask two raters to manually label all possible unit replacements on the development set as correct or incorrect, so that rp, rr, and rf can be automatically computed under each set of parameters.
    Page 5, “Statistical Paraphrase Generation”
  5. The parameters that result in the highest rf on the development set are finally selected.
    Page 5, “Statistical Paraphrase Generation”
  6. In our experiments, the development set contains 200 sentences and the test set contains 500 sentences, both of which are randomly selected from the human translations of 2008 NIST Open Machine Translation Evaluation: Chinese to English Task.
    Page 6, “Experimental Setup”

See all papers in Proc. ACL 2009 that mention development set.

See all papers in Proc. ACL that mention development set.

Back to top.

score function

Appears in 5 sentences as: score function (3) score functions (2)
In Application-driven Statistical Paraphrase Generation
  1. The PTs used in this work are constructed using different corpora and different score functions (Section 3.5).
    Page 3, “Statistical Paraphrase Generation”
  2. Let (51,72) be a pair of paraphrase units, their paraphrase likelihood is computed using a score function ¢pm(§i,fi).
    Page 3, “Statistical Paraphrase Generation”
  3. Suppose we have K PTs, (ski, {1%) is a pair of paraphrase units from the k-th PT with the score function gbk(§ki, £191.
    Page 3, “Statistical Paraphrase Generation”
  4. We use five PTs in this work (except the self-paraphrase table), in which each pair of paraphrase units has a score assigned by the score function of the corresponding method.
    Page 5, “Statistical Paraphrase Generation”
  5. The details of the corpora, methods, and score functions are presented in (Zhao et al., 2008a).
    Page 5, “Statistical Paraphrase Generation”

See all papers in Proc. ACL 2009 that mention score function.

See all papers in Proc. ACL that mention score function.

Back to top.

machine translation

Appears in 4 sentences as: Machine Translation (1) machine translation (3)
In Application-driven Statistical Paraphrase Generation
  1. PG shows its importance in many areas, such as question expansion in question answering (QA) (Duboue and Chu-Carroll, 2006), text polishing in natural language generation (NLG) (Iordanskaja et al., 1991), text simplification in computer-aided reading (Carroll et al., 1999), and sentence similarity computation in the automatic evaluation of machine translation (MT) (Kauchak and Barzilay, 2006) and summarization (Zhou et al., 2006).
    Page 1, “Introduction”
  2. o Sentence similarity computation: Given a reference sentence s’, this application aims to paraphrase s into t, so that t is more similar (closer in wording) with s’ than s. This application is important for the automatic evaluation of machine translation and summarization, since we can paraphrase the human translations/summaries to make them more similar to the system outputs, which can refine the accuracy of the evaluation (Kauchak and Barzilay, 2006).
    Page 4, “Statistical Paraphrase Generation”
  3. In our experiments, the development set contains 200 sentences and the test set contains 500 sentences, both of which are randomly selected from the human translations of 2008 NIST Open Machine Translation Evaluation: Chinese to English Task.
    Page 6, “Experimental Setup”
  4. (1) It is the first statistical model specially designed for paraphrase generation, which is based on the analysis of the differences between paraphrase generation and other researches, especially machine translation .
    Page 8, “Conclusions and Future Work”

See all papers in Proc. ACL 2009 that mention machine translation.

See all papers in Proc. ACL that mention machine translation.

Back to top.

rule-based

Appears in 4 sentences as: Rule-based (2) rule-based (3)
In Application-driven Statistical Paraphrase Generation
  1. Rule-based methods: Rule-based PG methods build on a set of paraphrase rules or patterns, which are either hand crafted or automatically collected.
    Page 1, “Related Work”
  2. In the early rule-based PG research, the paraphrase rules are generally manually written (McKeown, 1979; Zong et al., 2001), which is expensive and arduous.
    Page 1, “Related Work”
  3. Some researchers then tried to automatically extract paraphrase rules (Lin and Pantel, 2001; Barzilay and Lee, 2003; Zhao et al., 2008b), which facilitates the rule-based PG methods.
    Page 1, “Related Work”
  4. But they are either rule-based (Murata and Isa-hara, 2001; Takahashi et al., 2001) or thesaurus-based (Bolshakov and Gelbukh, 2004), thus they have some limitations as stated above.
    Page 2, “Related Work”

See all papers in Proc. ACL 2009 that mention rule-based.

See all papers in Proc. ACL that mention rule-based.

Back to top.

objective function

Appears in 3 sentences as: objective function (3)
In Application-driven Statistical Paraphrase Generation
  1. In SMT, however, the optimization objective function in MERT is the MT evaluation criteria, such as BLEU.
    Page 5, “Statistical Paraphrase Generation”
  2. We therefore introduce a new optimization objective function in this paper.
    Page 5, “Statistical Paraphrase Generation”
  3. Replacement f-measure (rf): We use rf as the optimization objective function in MERT, which is similar to the conventional f-measure and lever-agesrp and rr: 7“f = (2 X 7”]?
    Page 5, “Statistical Paraphrase Generation”

See all papers in Proc. ACL 2009 that mention objective function.

See all papers in Proc. ACL that mention objective function.

Back to top.

WordNet

Appears in 3 sentences as: WordNet (3)
In Application-driven Statistical Paraphrase Generation
  1. In the first phase, it extracts all synonyms from a thesaurus, such as WordNet , for the words to be substituted.
    Page 2, “Related Work”
  2. They paraphrase a sentence s by replacing its words with WordNet synonyms, so that s can be more similar in wording to another sentence s’.
    Page 8, “Results and Analysis”
  3. has also been proposed in (Zhou et al., 2006), which uses paraphrase phrases like our PT—l instead of WordNet synonyms.
    Page 8, “Results and Analysis”

See all papers in Proc. ACL 2009 that mention WordNet.

See all papers in Proc. ACL that mention WordNet.

Back to top.

PDF available from www.aclweb.org.
Send us an email