Elamur Research⌕
☰
home / papers / 2018 / SUD or Surface-Syntactic Universal Dependencies: An annotation scheme near-isomorphic to UD

SUD or Surface-Syntactic Universal Dependencies: An annotation scheme near-isomorphic to UD

Summary

The authors argue that the UD scheme, in which the head is always a content word, makes syntactic trees too flat. It also hides the hierarchy among function words. Moreover, UD relation names include the part of speech of the dependent, so a single syntactic position receives different labels. To address these problems, the authors propose SUD, a surface-syntactic annotation scheme. In it, the head is determined by the distribution of the word, so prepositions, auxiliaries and copulas become heads. Conjuncts form a chain rather than a "bouquet". Function labels do not depend on the part of speech: for example, a single label subj instead of nsubj and csubj. The relations are organized into a hierarchy that allows underspecified labels. On the Grew platform, the authors wrote graph rewriting rules for conversion in both directions: UD→SUD and SUD→UD. The round-trip conversion was tested on the test sets of all 122 UD 2.2 treebanks: information is preserved almost without loss. Unstable trees often point to inconsistent annotation, so the method is suitable for detecting errors in UD.

Key points

  • After the round-trip conversion UD→SUD→UD, the median LAS is 94.1% and the median UAS is 98.8% (122 UD 2.2 corpora, test sets).
  • The lowest LAS is found for UD_Korean-Kaist (71.8%), UD_Japanese-BCCWJ (74.0%) and UD_Japanese-GSD (74.4%). The highest is UD_Hungarian-Szeged (98.6%). UD_Warlpiri-UFAL and UD_Telugu-MTG reached a UAS of 100%, while UD_Yoruba-YTB has the lowest UAS (85.0%).
  • The main source of discrepancies is a relation label that does not match the part of speech of the dependent. For instance, in UD_Korean-PUD, the dependent of 22.4% of advmod relations is tagged with a part of speech other than ADV.
  • SUD trees are deeper but have shorter dependencies. In an example from the EWT corpus, the depth increases from 3 to 5, while the maximum number of dependents of a node drops from 8 to 5. The authors relate this to the principle of dependency length minimization.
  • All 122 UD 2.2 treebanks converted to SUD, as well as the Grew grammars, are available at gitlab.inria.fr/grew/SUD.

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

Institutions

🔬Centre National de la Recherche ScientifiqueParis · France83 works · 59 287 citations🔬Institut national de recherche en sciences et technologies du numériqueLe Chesnay-Rocquencourt · France55 works · 16 624 citations🔬Laboratoire Lorrain de Recherche en Informatique et ses ApplicationsNancy · France19 works · 3 389 citations🎓Université de LorraineNancy · France13 works · 772 citations🎓Université Sorbonne NouvelleParis · France5 works · 457 citations🔬SEMAGRAMME: Analyse Sémantique de la LangueVillers-lès-Nancy · France6 works · 221 citations🔬Modèles Dynamiques CorpusNanterre · France3 works · 190 citations🎓Université Paris NanterreNanterre · France3 works · 190 citations🔬Laboratoire de Phonétique et PhonologieParis · France2 works · 174 citations

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/w18-6008

@inproceedings{gerdes2018sud,
  title = {{SUD or Surface-Syntactic Universal Dependencies: An annotation scheme near-isomorphic to UD}},
  author = {Gerdes, Kim and Guillaume, Bruno and Kahane, Sylvain and Perrier, Guy},
  booktitle = {Proceedings of the Second Workshop on Universal Dependencies (UDW 2018)},
  year = {2018},
  doi = {10.18653/v1/w18-6008},
  url = {https://doi.org/10.18653/v1/w18-6008}
}

References

7 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Leonard Bloomfield. 1933. Language. find
  2. Igor Mel’čuk. 1988. Dependency Syntax: Theory and Practice, SUNY Press. find
  3. Ines Rehbein, Julius Steen, Bich-Ngoc Do. 2017. Universal Dependencies are hard to parse–or are they? Proceedings of the Fourth International Conference on Dependency Linguistics (Depling 2017). find
  4. Guglielmo Cinque, 1999. Adverbs and functional heads: A cross-linguistic perspective. Oxford University Press. find
  5. Marie-Catherine de Marneffe, Bill MacCartney, Christopher D. Manning. 2006. Generating typed dependency parses from phrase structure parses. Proceedings of LREC. find our card
  6. Ivan A. Sag. 2003. Coordination underspecification. Proceedings of the International Conference on HPSG. find
  7. Richard Futrell, Kyle Mahowald, Edward Gibson. 2015. Large-scale evidence of dependency length and 9th Agata Savary, et al. 2015. PARSEME–PARSing and Multiword Expressions within a European multilingual network. 7th Language & Technology Conference: Human Language Technologies as a Challenge for Computer Science and Linguistics (LTC 2015). find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/gerdes-2018-sud-or-surface-syntactic-universal/