Elamur Research⌕
☰
home / papers / 2015 / Universal Dependencies for Croatian (that work for Serbian, too)

Universal Dependencies for Croatian (that work for Serbian, too)

Summary

The paper addresses dependency parsing for Croatian and Serbian, South Slavic languages that lack annotated resources. Heterogeneous annotation schemes across treebanks make it hard to compare parsers and to transfer them between languages, whereas UD offers a single cross-lingual format. The authors therefore built Croatian UD: newspaper texts from the SETIMES.HR corpus received new layers of part-of-speech, morphological and syntactic annotation following the UD guidelines. MTE4 tags were manually converted to UPOS and universal morphological features. Four experts annotated the dependency trees from scratch, without automatic conversion. The resource also includes Croatian and Serbian test sets from news and Wikipedia, annotated with the same scheme. Parsers trained on Croatian UD outperformed models using the SETIMES.HR scheme in UAS and LAS. On Serbian texts they are practically as accurate as on Croatian ones. Delexicalized transfer from treebanks of ten European languages is examined separately. On non-unified CoNLL data it yields low scores that do not reflect language relatedness. The work matters because Croatian gained a UD-compatible treebank, and Serbian gained a strong parser without training data of its own.

Key points

  • Resource. Croatian UD is annotated on top of the original SETIMES.HR. It comprises 3,557 training and 200 development sentences of newspaper text (3,757 in total), as well as four test sets of 100 sentences each: news and Wikipedia in Croatian and Serbian. The same sentences carry two parallel annotations, SETIMES.HR and UD. The data and the tag mapping table are distributed under a free license (github.com/ffnlp/sethr).
  • Morphology. 14 of the 17 UPOS tags are used: DET, INTJ and SYM simply did not occur in the texts. Some MTE4 features are absent from the UD specification, so the number of full morphosyntactic tags dropped from 662 to 540. The authors introduced no language-specific extensions, either here or in the syntax.
  • Syntax. 39 of the 40 universal relations are used (only reparandum is missing), whereas SETIMES.HR has 15 labels. The most frequent relations are nmod (≈17%), punct (≈12.9%) and case (≈9.8%). Following the UD principle, content words become heads, so coordination, subordinate clauses and prepositional phrases are represented differently than in SETIMES.HR. Non-projective sentences make up 7.6% of the UD layer, 10.1% of SETIMES.HR, and about 20% of the PDT-compatible HOBS.
  • Lexicalized parsers. The experiments use gold POS and features. According to previous studies, LAS with automatic tags is on average 2–3 points lower. The Croatian UD model with UPOS and features achieves an overall UAS of 85.3 / LAS of 79.6, while the equivalent SETIMES.HR model achieves 82.7 / 76.4. On the Serbian test sets this model reaches LAS 81.5 (news) and 77.9 (Wikipedia), on par with 81.5 and 77.3 for Croatian.
  • Cross-lingual transfer. Delexicalized parsers, whose only feature is CPOS, were trained on treebanks of 10 European languages in their CoNLL 2006–2007 and UD 1.0 versions, i.e. 20 models in total. By UAS, Greek turned out best among the CoNLL sources, while Bulgarian and Czech, which are closer to Croatian and Serbian, ended up considerably lower. Czech, despite the large size of its treebank, came second to last; the authors attribute this to annotation heterogeneity.

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: Linguistics & languages

Topics: Universal Dependencies · Treebanks

Cited by

11 papers of our library cite it (of the 140673 whose bibliographies we read); 1 of them have a page here, 10 not yet.

Cite

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

@inproceedings{agic2015universal,
  title = {{Universal Dependencies for Croatian (that work for Serbian, too)}},
  author = {Agić, Željko and Ljubešić, Nikola},
  booktitle = {The 5th Workshop on Balto-Slavic Natural Language Processing},
  year = {2015},
  url = {https://aclanthology.org/W15-5301/}
}

References

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

  1. Anne Abeillé. 2003. Treebanks: Building and Using Parsed Corpora. Springer. find
  2. Željko Agić and Nikola Ljubešić. 2014. The SE-Times.HR linguistically annotated corpus of Croatian. In LREC, pages 1724–1727. find
  3. Željko Agić and Danijela Merkler. 2013. Three syntactic formalisms for data-driven dependency parsing of Croatian. LNCS, 8082:560–567. find
  4. Željko Agić, Nikola Ljubešić, and Danijela Merkler. 2013a. Lemmatization and morphosyntactic tagging of Croatian and Serbian. In BSNLP, pages 48–57. find
  5. Željko Agić, Danijela Merkler, and Daša Berović. 2013b. Parsing Croatian and Serbian by using Croatian dependency treebanks. In SPMRL, pages 22– 33. find
  6. Željko Agić, Daša Berović, Danijela Merkler, and Marko Tadić. 2014a. Croatian dependency treebank 2.0: New annotation guidelines for improved parsing. In LREC, pages 2313–2319. find our card
  7. Željko Agić, Jörg Tiedemann, Kaja Dobrovoljc, Simon Krek, Danijela Merkler, and Sara Može. 2014b. Cross-lingual dependency parsing of related languages with rich morphosyntactic tagsets. In LT4CloseLang, pages 13–24. find
  8. Željko Agić, Maria Jesus Aranzabe, Aitziber Atutxa, Cristina Bosco, Jinho Choi, Marie-Catherine de Marneffe, Timothy Dozat, Richárd Farkas, Jennifer Foster, Filip Ginter, Iakes Goenaga, Koldo Gojenola, Yoav Goldberg, Jan Hajič, Anders Trærup Johannsen, Jenna Kanerva, Juha Kuokkala, Veronika Laip… find
  9. Željko Agić. 2012. K-best spanning tree dependency parsing with verb valency lexicon reranking. In COLING, pages 1–12. find
  10. Daša Berović, Željko Agić, and Marko Tadić. 2012. Croatian dependency treebank: Recent developments and initial experiments. In LREC, pages 1902–1906. find
  11. Alena Böhmová, Jan Hajič, Eva Hajičová, and Barbora Hladká. 2003. The Prague dependency treebank. In Treebanks: Building and Using Parsed Corpora, pages 103–127. Springer. find
  12. Bernd Bohnet. 2010. Top accuracy and fast dependency parsing is not a contradiction. In COLING, pages 89–97. find our card
  13. Sabine Buchholz and Erwin Marsi. 2006. CoNLL-X shared task on multilingual dependency parsing. In CoNLL, pages 149–164. find our card
  14. Dipanjan Das and Slav Petrov. 2011. Unsupervised part-of-speech tagging with bilingual graph-based projections. In ACL, pages 600–609. find
  15. Marie-Catherine de Marneffe, Timothy Dozat, Natalia Silveira, Katri Haverinen, Filip Ginter, Joakim Nivre, and Christopher D. Manning. 2014. Universal stanford dependencies: A cross-linguistic typology. In LREC, pages 4585–4592. find our card
  16. Bojana Djordjević. 2014. Initial steps in building Serbian treebank: Morphological annotation. In Natural Language Processing for Serbian: Resources and Applications, pages 41–53. find
  17. Jakob Elming, Anders Johannsen, Sigrid Klerke, Emanuele Lapponi, Hector Martinez Alonso, and Anders Søgaard. 2013. Downstream effects of tree-to-dependency conversions. In NAACL, pages 617– 626. find
  18. Tomaž Erjavec. 2012. Multext-East: Morphosyntactic resources for central and eastern European languages. Language Resources and Evaluation, 46(1):131–142. find
  19. Richard Johansson. 2013. Training parsers on incompatible treebanks. In NAACL, pages 127–137. find
  20. Ryan McDonald, Slav Petrov, and Keith Hall. 2011. Multi-source transfer of delexicalized dependency parsers. In EMNLP, pages 62–72. find our card
  21. Ryan McDonald, Joakim Nivre, Yvonne Quirmbach-Brundage, Yoav Goldberg, Dipanjan Das, Kuzman Ganchev, Keith Hall, Slav Petrov, Hao Zhang, Oscar Täckström, Claudia Bedini, Núria Bertomeu Castelló, and Jungmee Lee. 2013. Universal dependency annotation for multilingual parsing. In ACL, pages 92–97. find
  22. Joakim Nivre, Johan Hall, Sandra Kübler, Ryan Mc-Donald, Jens Nilsson, Sebastian Riedel, and Deniz Yuret. 2007. find
  23. The CoNLL 2007 shared task on dependency parsing. In EMNLP-CoNLL, pages 915– 932. find
  24. Joakim Nivre, Cristina Bosco, Jinho Choi, Marie-Catherine de Marneffe, Timothy Dozat, Richárd Farkas, Jennifer Foster, Filip Ginter, Yoav Goldberg, Jan Hajič, Jenna Kanerva, Veronika Laippala, Alessandro Lenci, Teresa Lynn, Christopher Manning, Ryan McDonald, Anna Missilä, Simonetta Montemagni, S… find
  25. Slav Petrov, Dipanjan Das, and Ryan McDonald. 2012. A universal part-of-speech tagset. In LREC, pages 2089–2096. find
  26. Djamé Seddah, Reut Tsarfaty, Sandra Kübler, Marie Candito, Jinho D. Choi, Richárd Farkas, Jennifer Foster, Iakes Goenaga, Koldo Gojenola Gal-letebeitia, Yoav Goldberg, Spence Green, Nizar Habash, Marco Kuhlmann, Wolfgang Maier, Joakim Nivre, Adam Przepiórkowski, Ryan Roth, Wolfgang Seeker, Yanni… find our card
  27. Marko Tadić. 2007. Building the Croatian dependency treebank: The initial stages. Suvremena lingvistika, 63:85–92. find
  28. Jörg Tiedemann. 2014. Rediscovering annotation projection for cross-lingual parser induction. In COL-ING, pages 1854–1864. find
  29. Hans Uszkoreit and Georg Rehm. 2012. Language White Paper Series. Springer. find
  30. David Yarowsky, Grace Ngai, and Richard Wicentowski. 2001. Inducing multilingual text analysis tools via robust projection across aligned corpora. In HLT, pages 1–8. find
  31. Daniel Zeman, Ondřej Dušek, David Mareček, Martin Popel, Loganathan Ramasamy, Jan Štěpánek, Zdeněk Žabokrtskỳ, and Jan Hajič. 2014. HamleDT: Harmonized multi-language dependency treebank. Language Resources and Evaluation, 48(4):601–637. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/agic-2015-universal-dependencies-croatian-that-work/