Elamur Research⌕
☰
home / papers / 2008 / Cross-Language Parser Adaptation between Related Languages

Cross-Language Parser Adaptation between Related Languages

Summary

The paper asks whether a usable parser can be built for a resource-poor language without first creating a treebank for it. The authors borrow from McClosky et al.'s domain adaptation work and treat two closely related languages as two "domains" of one language. Danish serves as the source and Swedish as the target. They train the Charniak-Johnson reranking parser on the Danish Dependency Treebank, which they convert to phrase structures, and evaluate it on Talbanken05. To make the comparison fair, they first normalize the annotation styles of the two treebanks and map their tag sets onto each other. Two techniques address the vocabulary mismatch: glosses, which use a Danish-Swedish bitext, and delexicalization, which parses over tags instead of words. They also try self-training on unlabeled Swedish text from JRC-Acquis. The best system, a delexicalized reranking parser, reaches 66.40% labeled F-score against a plain baseline of about 43%. That matches a parser trained on roughly 1,500 native Swedish trees. Building a treebank that size would take months, while this approach could be set up in days.

Key points

  • Data:
    • Danish Dependency Treebank (CoNLL 2006 version): 5,190 sentences, split into 4,900 for training and 290 for testing (276 of them ≤40 words).
    • Talbanken05: 11,042 sentences, split into 10,700 for training and 342 for testing (317 of them ≤40 words).
    • Unlabeled data from JRC-Acquis: 430,808 Swedish sentences (about 6.15M tokens), plus 283,509 aligned Danish-Swedish segments, word-aligned with GIZA++.
  • Pipeline:
    • Dependencies are converted to constituents with the "flattest-possible" algorithm.
    • Head morphological tags are mapped to Penn Treebank-style non-terminals, so the Charniak-Johnson parser can run unmodified.
    • Swedish text was retagged with a Swedish version of Hajič's tagger.
  • Main result:
    • The delexicalized Brown (reranking) parser scored 66.40% F.
    • This is about 23 points absolute above the baseline, or roughly 41% error reduction.
    • On the Swedish learning curve, it equals about 1,546 training trees.
  • Self-training:
    • Swedish Acquis was parsed and added to the training data, without retraining the reranker.
    • This did not beat the best Part One result, and the Danish-trained reranker gave no gain in this setting.
  • Upper bound: A parser trained on delexicalized, automatically tagged Swedish (hybrid tags) reaches 77.01% F. Better treebank and tag mapping and better tagging are named as ways to close the gap.

Interesting ideas

  • I1. Related languages as domains: Domain-adaptation machinery such as reranking and self-training can be reused across closely related languages. This could transfer to other pairs with shared syntax but different spelling.
  • I2. Normalizing annotation conventions first: Before measuring cross-language transfer, the authors harmonize systematic style differences. Examples are how coordination, determiners, numerals and genitives attach. This keeps the result from mixing language effects with guideline effects. It is a useful step whenever treebanks from different teams are combined.
  • I3. Interlingua-style tag conversion: Each tag set gets a driver that decodes tags into a shared feature space and encodes them back.
    • Feature priorities resolve conflicts; for example, part of speech outranks gender.
    • Drivers are written once and reused for any pair of tag sets, a design that UD-style feature schemes echo.
  • I4. Delexicalization with restuffing: The parser runs on tags only, and the original words are put back into the output afterwards. The restored, word-level parses can then train a lexicalized model in the target language.
  • I5. Treebank-size-equivalent evaluation: The authors place the adapted parser's score on the target-language learning curve to see how many annotated trees it is worth. This turns an abstract F-score into a concrete annotation cost, which helps when deciding whether to fund a new treebank.

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

Fields: Natural language processing

Topics: Treebanks · Part-of-speech tagging

Cited by

126 papers of our library cite it (of the 140884 whose bibliographies we read); 20 of them have a page here, 106 not yet.

Cite

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

@inproceedings{zeman2008cross,
  title = {{Cross-Language Parser Adaptation between Related Languages}},
  author = {Zeman, Daniel and Resnik, Philip},
  booktitle = {Proceedings of the IJCNLP-08 Workshop on NLP for Less Privileged Languages},
  year = {2008},
  url = {https://aclanthology.org/I08-3008/}
}

References

19 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Rens Bod. 2006a. Unsupervised Parsing with U-DOP. In: Proceedings of the Conference on Natural Language Learning (CoNLL-2006). New York, New York, USA. find
  2. Rens Bod. 2006b. An All-Subtrees Approach to Unsupervised Parsing. In: Proceedings of the 21 st International Conference on Computational Linguistics and the 44 th Annual Meeting of the ACL (COLING-ACL-2006). Sydney, Australia. find
  3. Alena Böhmová, Jan Hajič, Eva Hajičová, Barbora Hladká. 2003. The Prague Dependency Treebank: A Three-Level Annotation Scenario. In: Anne Abeillé (ed.): Treebanks: Building and Using Syntactically Annotated Corpora. Kluwer Academic Publishers, Dordrecht, The Netherlands. find
  4. Eugene Charniak, Mark Johnson. 2005. Coarse-to-Fine N-Best Parsing and MaxEnt Discriminative Reranking. In: Proceedings of the 43 rd Annual Meeting of the ACL (ACL-2005), pp. 173–180. Ann Arbor, Michigan, USA. find our card
  5. Michael Collins. 1997. Three Generative, Lexicalized Models for Statistical Parsing. In: Proceedings of the 35 th Annual Meeting of the ACL, pp. 16–23. Madrid, Spain. find
  6. Michael Collins, Jan Hajič, Lance Ramshaw, Christoph Tillmann. 1999. A Statistical Parser for Czech. In: Proceedings of the 37 th Annual Meeting of the ACL (ACL-1999), pp. 505–512. College Park, Maryland, USA. find our card
  7. Jan Hajič. 2004. Disambiguation of Rich Inflection (Computational Morphology of Czech). Karolinum, Charles University Press, Praha, Czechia. find
  8. Rebecca Hwa, Philip Resnik, Amy Weinberg, Clara Cabezas, Okan Kolak. 2004. Bootstrapping Parsers via Syntactic Projection across Parallel Texts. In: Natural Language Engineering 1 (1): 1–15. Cambridge University Press, Cambridge, England. find
  9. Dan Klein, Christopher D. Manning. 2004. Corpus-Based Induction of Syntactic Structure: Models of Dependency and Constituency. In: Proceedings of the 42 nd Annual Meeting of the ACL (ACL-2004). Barcelona, Spain. find
  10. Philipp Koehn, Franz Josef Och, Daniel Marcu. 2003. Statistical Phrase-Based Translation. In: Proceedings of HLT-NAACL 2003, pp. 127–133. Edmonton, Canada. find
  11. Matthias T. Kromann, Line Mikkelsen, Stine Kern Lynge. 2004. Danish Dependency Treebank. At: http://www.id.cbs.dk/~mtk/treebank/. København, Denmark. ↗
  12. Mitchell P. Marcus, Beatrice Santorini, Mary Ann Marcinkiewicz. 1993. Building a Large Annotated Corpus of English: the Penn Treebank. In: Computational Linguistics, vol. 19, pp. 313–330. find our card
  13. David McClosky, Eugene Charniak, Mark Johnson. 2006. Reranking and Self-Training for Parser Adaptation. In: Proceedings of the 21 st International Conference on Computational Linguistics and the 44 th Annual Meeting of the ACL (COLING-ACL-2006). Sydney, Australia. find our card
  14. Joakim Nivre, Jens Nilsson, Johan Hall. 2006. Talbanken05: A Swedish Treebank with Phrase Structure and Dependency Annotation. In: Proceedings of the 5 th International Conference on Language Resources and Evaluation (LREC-2006). May 24-26. Genova, Italy. find
  15. Franz Josef Och, Hermann Ney. 2000. Improved Statistical Alignment Models. In: Proceedings of the 38 th Annual Meeting of the ACL (ACL-2000), pp. 440–447. Hong Kong, China. find our card
  16. Philip Resnik, Noah A. Smith. 2003. The Web as a Parallel Corpus. In: Computational Linguistics, 29(3), pp. 349–380. find our card
  17. Mark Steedman, Miles Osborne, Anoop Sarkar, Stephen Clark, Rebecca Hwa, Julia Hockenmaier, Paul Ruhlen, Steven Baker, Jeremiah Crim. 2003. Bootstrapping Statistical Parsers from Small Datasets. In: Proceedings of the 11 th Conference of the European Chapter of the ACL (EACL-2003). Budapest, Hungary. find
  18. Ralf Steinberger, Bruno Pouliquen, Anna Widiger, Camelia Ignat, Tomaž Erjavec, Dan Tufiş, Dániel Varga. 2006. The JRC-Acquis: A Multilingual Aligned Parallel Corpus with 20+ Languages. In: Proceedings of the 5 th International Conference on Language Resources and Evaluation (LREC-2006). May 24-26. G… find our card
  19. Fei Xia, Martha Palmer. 2001. Converting Dependency Structures to Phrase Structures. In: Proceedings of the 1 st Human Language Technology Conference (HLT-2001). San Diego, California, USA. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zeman-2008-cross-language-parser-adaptation-between/