Elamur Research
☰
home / papers / 2001 / Extracting Paraphrases from a Parallel Corpus

Extracting Paraphrases from a Parallel Corpus

What it is about

The paper extracts paraphrases from several English translations of the same foreign source text, a comparable corpus that differs in wording but not in meaning. The data are five classic literary books translated more than once. Sentences are aligned across translations by dynamic programming on shared words. Each candidate word or phrase is described by its context (neighbouring words and part-of-speech n-grams, and lexico-syntactic features), and a co-training algorithm builds two classifiers from two views of the same pairs and labels new pairs in rounds, starting from a few seeds. The output is 9,483 lexical paraphrase pairs and 25 morpho-syntactic rules. In an evaluation of 500 sampled pairs against the guideline "approximate conceptual equivalence", a set of 70 paraphrases found by a human judge was matched by the algorithm for 48 of them (69%). The paper argues that these pairs cover paraphrase phenomena that had not been studied empirically.

Ideas

For Mova

The paper addresses the lexicon (open problem 2) and the author delta, with verse pairs that already exist in tools/translate. Its precision figure is from a single sample judged by people and should not be taken as a target. Summary only; the source text is not carried into the build.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/P01-1008.pdf

Summary

Barzilay and McKeown address the lack of reusable paraphrase resources. Hand-built collections are costly and tied to one application, and WordNet-style resources cover only single-word relations. Their data are 11 English translations of 5 foreign literary works, including Madame Bovary, Andersen's Fairy Tales and Twenty Thousand Leagues Under the Sea. Translations of the same source keep its meaning but differ in wording, so they are a natural source of paraphrases. After aligning corresponding sentences, they apply an unsupervised, iterative co-training procedure. Identical words in aligned sentences serve as seeds for learning which surrounding contexts reliably signal equivalence. Those contexts then yield new paraphrase pairs, which in turn train new contexts, until nothing new is found. The method extracts single-word, multi-word and morpho-syntactic paraphrases. Human judges rated its output more precise than a bilingual lexicon induction method from MT (Melamed, 2001), at 71.6% versus 52.7%. A WordNet analysis shows that real paraphrases often go beyond synonymy. The work showed that parallel monolingual text can feed paraphrase acquisition for summarization and generation.

Key points

  • Preprocessing:
    • Sentences were aligned with Gale & Church-style dynamic programming, weighted by shared words. On average, 42% of words in matched sentences are identical.
    • This produced 44,562 sentence pairs (1,798,526 words). A manual check of 127 pairs found 94.5% correct.
    • Mikheev's tagger/chunker supplied noun and verb phrases as the basic units, and CELEX supplied derivational roots.
  • Coverage: in 50 sentences, a human judge identified 70 paraphrases. The algorithm recovered 48 of them (69%).
  • Comparison with Melamed's system:
    • Both systems were run on the public 60% of the data, and the top 6,826 pairs were taken from each.
    • Six native speakers judged 500 random pairs from each system, with no context shown.
    • About 29% of this system's pairs were multi-word, against roughly 10% for Melamed's.
  • WordNet relations: among 112 frequent pairs (each occurring at least 20 times), 35% were synonyms and 32% hyperonyms. Another 18% were hyperonym-tree siblings and 10% were unrelated.
  • Data release: the copyright-free portion of the corpus (9 translations) was released by the authors.

Interesting ideas

  • Identical words as free supervision. Words shared by two aligned sentences act as positive examples for learning which contexts predict equivalence. This gives labeled data without annotation and could carry over to any setting with redundant parallel text.
  • Local context over global alignment. Literary translations diverge too much for clean word alignment. Matching only the immediate surroundings of candidate phrases is more robust, which also makes it suitable for noisy or comparable corpora.
  • Bootstrapping between two views. Contexts and paraphrase pairs are learned in alternation, and the loop stops when no new pairs appear. This co-training pattern is a reusable template for expanding small seed lexicons.
  • Generalizing pairs into rules. Extracted pairs are abstracted into morpho-syntactic patterns using POS tags and morphological roots. The result is productive rules rather than just a list of pairs, which could be valuable for morphologically rich languages.
  • Mapping output onto a thesaurus. Projecting extracted pairs onto WordNet relations shows empirically what kinds of relations people use when paraphrasing. This is an easy diagnostic for any lexical resource, and here it shows that relying only on synonymy would miss many valid paraphrases.

Topics: Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.3115/1073012.1073020

@inproceedings{barzilay2001extracting,
  title = {{Extracting Paraphrases from a Parallel Corpus}},
  author = {Barzilay, Regina and McKeown, Kathleen R.},
  booktitle = {Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics},
  year = {2001},
  doi = {10.3115/1073012.1073020},
  url = {https://doi.org/10.3115/1073012.1073020}
}

References

26 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. R. H. Baayen, R. Piepenbrock, and H. van Rijn, editors. 1993. The CELEX Lexical Database(CD-ROM). Linguistic Data Consortium, University of Pennsylvania. find
  2. R. Barzilay and M. Elhadad. 1997. Using lexical chains for text summarization. In Proceedings of the ACL Workshop on Intelligent Scalable Text Summarization, pages 10–17, Madrid, Spain, August. find
  3. P. Brown, S. Della Pietra, V. Della Pietra, and R. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263–311. find
  4. M. Collins and Y. Singer. 1999. Unsupervised models for named entity classification. In proceedings of the Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora. find
  5. R. de Beaugrande and W. V. Dressler. 1981. Introduction to Text Linguistics. Longman, New York, NY. find
  6. M. Dras. 1999. Tree Adjoining Grammar and the Reluctant Paraphrasing of Text. Ph.D. thesis, Macquarie University, Australia. find
  7. W. Gale and K. W. Church. 1991. A program for aligning sentences in bilingual corpora. In Proceedings of the 29th Annual Meeting of the Association for Computational Linguistics, pages 1–8. find our card
  8. M. Halliday. 1985. An introduction to functional grammar. Edward Arnold, UK. find
  9. V. Hatzivassiloglou and K.R. McKeown. 1993. Towards the automatic identification of adjectival scales: Clustering adjectives according to their meaning. In Proceedings of the 31rd Annual Meeting of the Association for Computational Linguistics, pages 172–182. find
  10. V. Hatzivassiloglou, J. Klavans, and E. Eskin. 1999. Detecting text similarity over short passages: Exploring linguistic feature combinations via machine learning. In proceedings of the Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora. find
  11. L. Iordanskaja, R. Kittredge, and A. Polguere, 1991. Natural language Generation in Artificial Intelligence and Computational Linguistics, chapter 11. Kluwer Academic Publishers. find
  12. C. Jacquemin, J. Klavans, and E. Tzoukermann. 1997. Expansion of multi-word terms for indexing and retrieval using morphology and syntax. In proceedings of the 35th Annual Meeting of the ACL, pages 24–31, Madrid, Spain, July. ACL. find
  13. J.R. Landis and G.G. Koch. 1977. The measurement of observer agreement for categorical data. Biometrics, 33:159–174. find
  14. I. Langkilde and K. Knight. 1998. Generation that exploits corpus-based statistical knowledge. In proceedings of the COLING-ACL. find
  15. Maria Lapata. 2001. A corpus-based account of regular polysemy: The case of context-sensitive adjectives. In Proceedings of the 2nd Meeting of the NAACL, Pittsburgh, PA. find
  16. D. Lin. 1998. Automatic retrieval and clustering of similar words. In proceedings of the COLING-ACL, pages 768– 774. find
  17. Melamed. 2001. Empirical Methods for Exploiting Parallel Texts. MIT press. find
  18. A. Mikheev. 1997. the ltg part of speech tagger. University of Edinburgh. find
  19. G.A. Miller, R. Beckwith, C. Fellbaum, D. Gross, and K.J. Miller. 1990. Introduction to WordNet: An on-line lexical database. International Journal of Lexicography (special issue), 3(4):235–245. find
  20. F. Pereira, N. Tishby, and L. Lee. 1993. Distributional clustering of english words. In proceedings of the 30th Annual Meeting of the ACL, pages 183–190. ACL. find
  21. E. Riloff and R. Jones. 1999. Learning Dictionaries for Information Extraction by Multi-level Boot-strapping. In Proceedings of the Sixteenth National Conference on Artificial Intelligence, pages 1044–1049. The AAAI Press/MIT Press. find
  22. J. Robin. 1994. Revision-Based Generation of Natural Language Summaries Providing Historical Background: Corpus-Based Analysis, Design, Implementation, and Evaluation. Ph.D. thesis, Department of Computer Science, Columbia University, NY. find
  23. S. Siegel and N.J. Castellan. 1988. Non Parametric Statistics for Behavioral Sciences. McGraw-Hill. find
  24. J. Veronis, editor. 2000. Parallel Text Processing: Alignment and Use of Translation Corpora. Kluwer Academic Publishers. find
  25. R. Wechsler. 1998. Performing Without a Stage: The Art of Literary Translation. Catbird Press. find
  26. D. Yarowsky. 1995. Unsupervised word sense disambiguation rivaling supervised methods. In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics, pages 189– 196. find our card