K-UD: Revising Korean Universal Dependencies Guidelines

Revised syntactic dependency guidelines for annotating Korean in Universal Dependencies

arXiv 2024

Kyuwon Kim1, Yige Chen2, Eunkyul Leah Jo3,4, KyungTae Lim5, Jungyeul Park3, Chulwoo Park6

1Seoul National University, South Korea 2The Chinese University of Hong Kong, Hong Kong 3The University of British Columbia, Canada 4Sorbonne Université, France 5SeoulTech, South Korea 6Anyang University, South Korea

Affiliations as printed in the arXiv version (December 2024).

Paper (arXiv) PDF
Dependency analysis of 나는 장동건 사진을 보고 싶다 (na-neun jangdonggeon sajin-eul bogo sipda, 'I want to see a picture of Jang Dong-gun'). Top, current Korean GSD annotation: root on bogo 'see', nsubj to na-neun, obj to jangdonggeon, flat from jangdonggeon to sajin-eul, flat from bogo to sipda. Bottom, revised annotation: root on sipda 'want', xcomp to bogo, obj from bogo to sajin-eul, compound from sajin-eul to jangdonggeon, dislocated:nsubj to na-neun.
Figure 1. Example of the current annotation in Korean_GSD (top) and the revised annotation (bottom) for na-neun jangdonggeon sajin-eul bogo sipda ‘I want to see a picture of Jang Dong-gun’. Currently the first noun (jangdonggeon) and the first verb (bogo) are heads; in the revision, the last noun sajin-eul heads the noun phrase (compound), the catenative verb sipda ‘want’ is the root with an xcomp, and the topic-marked na-neun is dislocated:nsubj.

Abstract

Critique has surfaced concerning the existing linguistic annotation framework for Korean Universal Dependencies (UDs), particularly in relation to syntactic relationships. In this paper, our primary objective is to refine the definition of syntactic dependency of UDs within the context of analyzing the Korean language. Our aim is not only to achieve a consensus within UDs but also to garner agreement beyond the UD framework for analyzing Korean sentences using dependency structure, by establishing a linguistic consensus model.

Keywords Korean Universal Dependencies (UD) Annotation guidelines Korean dependency parsing Dependency treebank Head-final syntax Catenative constructions Sejong treebank KLUE

Proposed Revisions

Current Korean UD treebanks (GSD and Kaist) annotate the first word as the head of noun compounds, multiword proper names, and verb sequences, regardless of grammatical category, and attach many arguments with generic relations. Korean is head-final, so the revised guidelines align the dependency analysis with Korean syntax. In the paper’s figures, ‘top’ arcs show the current annotation and ‘bottom’ arcs the revised one; examples up to Section 3 come from ko_gsd-ud (UD v2.12).

  1. Nominal heads. In a noun compound, the head is the stem that determines the semantic category, usually the last noun; for multiword proper names, the last proper noun is the head because it receives case from the predicate. flat attaches to the last word in Korean (e.g., dates such as 1941nyeon 1wol 11il).
  2. Auxiliaries vs. catenative verbs. Verbs in catenative constructions (e.g., sipda ‘want’ in bogo sipda) are the lexical head and take an xcomp; auxiliaries expressing tense, mood, aspect, voice, or evidentiality are attached with aux.
  3. Clausal arguments. ccomp marks clausal complements, xcomp is reserved for catenative constructions, and csubj:pass is introduced for clausal passive subjects.
  4. Oblique arguments. Adverbial-postposition (jkb) phrases that are complements of the predicate according to the Sejong dictionary’s subcategorization frames are annotated obl:arg, distinguishing them from adjunct obl.
  5. Topic-marked subjects. Most nominals with the topic marker eun/neun, now labeled dislocated, are annotated dislocated:nsubj; expl is not used for Korean.
  6. Function words and MWEs. cop is annotated only when the copula is tokenized apart from its lexeme (e.g., after punctuation); mark marks subordinate clauses; fixed covers grammaticized expressions such as ppun anila ‘in addition to’.

Main Findings

  • Validation on new data. 200 sentences randomly selected from the Sejong project were manually annotated with the revised guidelines, using argument-structure frames from the Sejong verb dictionary (e.g., geolaesikida ‘trade’ with nsubj, iobj, and obj; mudeonada ‘stand out’ with nsubj and obl:arg).
  • Known annotation problems. The current Korean GSD data also contains morphological annotation errors, on average 2.46 words per sentence for language-specific XPOS and 2.35 words per sentence for Universal POS (as reported by Jo et al., 2023).
  • Consistency with Sejong-style dependencies. The paper compares the revision with the Sejong-style dependency structure used by KLUE, which consistently follows a right-to-left (head-final) pattern, as part of building a linguistic consensus model for Korean.
  • Next steps. The authors plan to add full Korean guidelines with examples to the UD relation pages and to align UD-style and Sejong-style treebanks with the KLUE benchmark consortium and the National Institute of Korean Language.
Table 1. Statistics of the available Korean dependency parsing datasets, in number of sentences. (Table 1 in the paper.)
SourceTrainDevTestTotal
GSD44009509896339
Kaist23,0102066228727,363
Penn–––5010
KLUE10,0002000250014,500

BibTeX

@misc{kim2024kudrevisingkoreanuniversal,
      title={K-UD: Revising Korean Universal Dependencies Guidelines}, 
      author={Kyuwon Kim and Yige Chen and Eunkyul Leah Jo and KyungTae Lim and Jungyeul Park and Chulwoo Park},
      year={2024},
      eprint={2412.00856},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.00856}, 
}
More projects from KAIST MLP Lab →