Elamur Research⌕
☰
home / papers / 1993 / Aligning Sentences in Bilingual Corpora Using Lexical Information

Aligning Sentences in Bilingual Corpora Using Lexical Information

Paper · 1993 · Stanley F. Chen

Summary

Chen addresses sentence alignment in large bilingual corpora. Earlier efficient methods (Brown et al., 1991b; Gale and Church, 1991) rely only on sentence length. They can drift on non-literal translations and small deletions, and they need manual anchors or pre-chunking to handle large gaps. This paper adds lexical information through a simple statistical word-to-word translation model that is estimated while alignment runs. The corpus is treated as a sequence of sentence beads (1:0, 0:1, 1:1, 2:1, 1:2), each generated from word beads. The best alignment is the one that maximises the probability of the corpus under this model. A dynamic-programming search with pruning keeps the cost linear, and a separate rare-word procedure finds where large deletions end. Three English/French collections were aligned: Canadian Hansards, newer Hansard proceedings, and European Economic Community proceedings. The estimated error on the Hansards is about 0.4% with no data discarded. This matches or beats length-based methods that drop the hardest sentences, and the method needs only 100 hand-aligned pairs for each language pair.

Key points

  • Aligned corpora: 3,000,000 sentences of Canadian Hansards, 1,000,000 of newer Hansard proceedings, and 2,000,000 of European Economic Community proceedings, all English/French.
  • Bootstrapping used 100 manually aligned sentence pairs, followed by further training on 20,000 unaligned sentences of the target corpus.
  • The estimated error rate is about 0.4% on the whole old Hansard corpus. Brown reports 0.6% after discarding roughly 10% of the data. Gale reports about 2% with nothing discarded, or 0.4% when 20% is discarded.
  • The model has three parts:
    • word beads of type 1:0, 0:1, and 1:1;
    • a uniform distribution over word order;
    • Poisson sentence-length distributions whose λ parameters are tied across bead types.
  • The method is language independent and parallelizable. Its running time, however, exceeds that of the Brown and Gale algorithms by a factor of tens, a cost the author considers acceptable for a one-time task.

Interesting ideas

  • I1. Learning the translation model during alignment: The word-pair model is trained only on sentence beads that every surviving partial alignment agrees on. This self-training trick could suit other bootstrapping tasks where confident early decisions can safely supervise later ones.
  • I2. Pruning against the best prefix of the same length: The search discards alignment prefixes that are much less probable than the best one of equal length. Because most beads are one-to-one, this turns a quadratic search into a linear one. The same idea fits other long monotone alignment problems.
  • I3. Rare words as markers for the end of a deletion: Rare words are stored in a hash table. A rare word in one text whose translation appears in the other becomes a candidate resumption point. The candidate is accepted only if the next forty sentences on each side align with high probability. This removes the need for hand-picked anchors when matching noisy parallel documents.
  • I4. Tying the length parameters: The Poisson λ values were tied across bead types because independently trained values drifted toward extremes. Tying related parameters is a cheap guard against degenerate estimates.
  • I5. Measuring very low error rates through disagreements: Sampling at random would rarely hit an error, so the author manually checked the 500+ places where his output differed from an existing alignment of known accuracy. This is a practical way to evaluate a strong system against a strong baseline.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Harvard UniversityCambridge · United States73 works · 37,309 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 9,085 worksArtificial Intelligence · 8,888 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 works

Cite

Cite as Stanley F. Chen. 1993. Aligning Sentences in Bilingual Corpora Using Lexical Information. 31st Annual Meeting of the Association for Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.3115/981574.981576

@inproceedings{chen1993aligning,
  title = {{Aligning Sentences in Bilingual Corpora Using Lexical Information}},
  author = {Stanley F. Chen},
  booktitle = {31st Annual Meeting of the Association for Computational Linguistics},
  year = {1993},
  doi = {10.3115/981574.981576},
  url = {https://doi.org/10.3115/981574.981576}
}

References

19 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. (Bellman, 1957) Richard Bellman. find
  2. Dynamic Programming. Princeton University Press, Princeton N.J., 1957. find
  3. (Brown et al., 1990) Peter F. Brown, John Cocke, Stephen A. DellaPietra, Vincent J. DellaPietra, Frederick Jelinek, John D. Lafferty, Robert L. Mercer, and Paul S. Roossin. A statistical approach to machine translation. Computational Linguistics, 16(2):79-85, June 1990. find
  4. (Brown et al., 1991a) Peter F. Brown, Stephen A. DellaPietra, Vincent J. DellaPietra, and Robert L. Mercer. Word sense disambiguation using statistical methods. In Proceedings 29th Annual Meeting of the ACL, pages 265-270, Berkeley, CA, June 1991. — Word-Sense Disambiguation Using Statistical Methods
  5. (Brown et al., 1991b) Peter F. Brown, Jennifer C. Lai, and Robert L. Mercer. Aligning sentences in parallel corpora. In Proceedings 29th Annual Meeting of the ACL, pages 169-176, Berkeley, CA, June 1991.
  6. (Brown et al., 1993) Peter F. Brown, Stephen A. DellaPietra, Vincent J. DellaPietra, and Robert L. Mercer. The mathematics of machine translation: Parameter estimation. find
  7. Computational Linguistics, 1993. To appear. find
  8. (Catizone et al., 1989) Roberta Catizone, Graham Russell, and Susan Warwick. Deriving translation data from bilingual texts. In Proceedings ]5 find
  9. E1 If there is some evidence that it ... and I will see that it does. E2 \SCM{} Translation \ECM{} Figure 3: An Alignment Error El Motion No. 22 that Bill C-84 be amended in ... and substituting the following therefor : second anniversary of. Figure 4: Another Alignment Error of the First International Acquisition Workshop, Detroit, Michigan, August 1989. find
  10. (Chen, 1993) Stanley 17. Chen. Aligning sentences in bilingual corpora using lexical information. Technical Report TR-12-93, Harvard University, 1993.
  11. (Dagan et al., 1991) Ido Dagan, Alon Itai, and U1-rike Schwall. Two languages are more informative than one. In Proceedings of the 29th Annual Meeting of the ACL, pages 130-137, 1991.
  12. (Dempster et al., 1977) A.P. Dempster, N.M. Laird, and D.B. Rubin. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society, 39(B):1-38, 1977. find
  13. (Gale and Church, 1991) William A. Gale and Kenneth W. Church. A program for aligning sentences in bilingual corpora. In Proceedings of the 29th Annual Meeting of the ACL, Berkeley, California, June 1991.
  14. (Gale et al., 1992) William A. Gale, Kenneth W. Church, and David Yarowsky. Using bilingual materials to develop word sense disambiguation methods. In Proceedings of the Fourth International Conference on Theoretical and Methodological lssues in Machine Translation, pages 101-112, Montr4al, Canada, June 1992. find
  15. (Kay, 1991) Martin Kay. Text-translation alignment. In A C H / A L L C 'gl: "Making Connections" Conference Handbook, Tempe, Arizona, March 1991.
  16. (Klavans and Tzoukermann, 1990) Judith Klavans and Evelyne Tzoukermann. The bicord system. In COLING-gO, pages 174-179, Helsinki, Finland, August 1990. find
  17. (Sadler, 1989) V. Sadler. The Bilingual Knowledge Bank - A New Conceptual Basis for MT. BSO/Research, Utrecht, 1989. find
  18. (Warwick and Russell, 1990) Susan Warwick and Graham Russell. Bilingual concordancing and F1 find
  19. Si on peut prouver que elle ... je verrais & ce que elle se y conforme. \SCM{} Language = French \ECM{} \SCM{} Paragraph \ECM{} F2 F 1 Motion No 22 que on modifie le projet de loi C-84 ... et en la rempla§ant par ce qui suit : ' 18. Deux ans apr~s : '. F2 bilingual lexicography. In E U R A L E X 4th later-national Congress, M~laga, Spain, 1990. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/chen-1993-aligning-sentences-bilingual-corpora-lexical/