Annotating Korean adnominal ending constructions in corpus data: Beyond relative-clause identification

A construction-sensitive annotation layer for the Korean adnominal ending ETM in the KLUE dependency treebank

arXiv 2026

Jungyeul Park1,†, Chulwoo Park2,†

1KAIST, South Korea 2Anyang University, South Korea

†Corresponding authors

Paper (arXiv) PDF
Sejong-style dependency tree from the KLUE treebank for the sentence 'François Hollande, the President of France, announced the highest level of a terror alert in Paris on this day'. Each eojeol shows Hangul, morpheme segmentation, POS tags, romanization, and gloss; arcs are labeled NP, NP_SBJ, NP_AJT, NP_MOD, NP_OBJ, AP, VP_MOD, and VP, with the root on ballyeonghaetda 'announced'. The adnominal form nop-eun 'high' is tagged VA+ETM and attached as VP_MOD.
Figure 1. An example from the KLUE dependency treebank (sentence meaning: ‘François Hollande, the President of France, announced the highest level of a terror alert in Paris on this day.’). Tokens are eojeol units with Sejong POS tags; the adnominal form nop-eun ‘high’ (VA+ETM) only receives the broad label VP_MOD, which does not say what kind of adnominal construction it realizes.

Abstract

The Korean adnominal ending ETM occurs in diverse noun-modifying constructions, including relative-clause-like modifiers, adjectival and copular forms, bound-noun constructions, and lexicalized expressions. This paper argues that ETM is not a direct marker of relative-clause structure, but a morphological exponent shared by several adnominal constructions. We propose a corpus-based typology that distinguishes these constructions using predicate type, auxiliary structure, argument-structural compatibility, head-noun restriction, and lexicalized patterns. We operationalize the typology as a construction-sensitive annotation layer for the KLUE dependency treebank, implemented through an ordered rule-based procedure and evaluated by manual validation. Productive relative-clause-like uses account for 39.4% of the analyzed instances; the remainder consists mainly of adjectival, copular, bound-nominal, modal, temporal, and collocational constructions. The findings show that Korean relative-clause-like modification cannot be identified from adnominal morphology alone.

Keywords Korean adnominal ending (ETM) Korean relative clauses Adnominal / noun-modifying constructions Corpus annotation Treebank annotation KLUE dependency treebank Sejong corpus Bound nouns Construction typology

Typology and Annotation Procedure

In Sejong-style treebanks such as the KLUE dependency treebank, adnominally marked predicates are grouped under broad labels like vp_mod, while Korean UD spreads them over acl:relcl, amod, advcl, flat, conj, and dep. Neither gives a transparent constructional classification of ETM. The paper adds a separate construction-sensitive layer on top of KLUE: every eojeol in the KLUE training split whose morphological analysis contains ETM is extracted with its morpheme sequence, following nominal head (its dependency head), and local dependents, and the Sejong lexical resource is consulted for predicate class and subcategorization in difficult cases.

Labels are assigned by an ordered rule procedure (Algorithm 1) that tests more specific, constructionally restricted patterns before productive relative-clause-like ones, so that expressions such as -e daehan N ‘N concerning’, Verb+ETM ttae ‘when’, and su itda/eopda ‘can/cannot’ are not over-analyzed as relative clauses:

  1. colloc+etm: lexicalized ETM + noun sequences (collocational or semi-fixed modifiers)
  2. modif+adj+etm: VA+ETM + noun (adjectival modifier)
  3. modif+vcp+etm: copular or negative copular adnominal form + noun
  4. modif+ndaneun+etm: VV + -neundaneun/-ndaneun + noun (content-like or quotative modifier)
  5. modif+su-iss+etm: su ‘possibility/way’ + iss/eops+ETM + noun (modal construction)
  6. modif+etm+ttae: Verb+ETM + ttae ‘time’ (temporal construction)
  7. modif+etm+nnb: Verb+ETM + bound noun, e.g. geos ‘thing’
  8. modif+vx+vx+etm: auxiliary complex + ETM + noun
  9. rc+vv+vx+etm, rc+va+vx+etm, rc+vv+etm: productive relative-clause-like modifiers (verb or adjective with an auxiliary complex, then the canonical VV/XSV+ETM + noun)
  10. undefined: annotation errors or structurally insufficient cases

Main Results

  • Relative clauses are a minority. Of 13,046 ETM tokens in the KLUE training data, the three relative-clause-like classes account for 5,135 (39.4%); non-relative and construction-specific classes account for 7,828 (60.0%), and 83 (0.6%) are undefined.
  • Morphology-based identification overgenerates. Treating every ETM as a relative-clause marker would overgenerate 7,911 of 13,046 instances (60.6%); excluding unresolved cases, the overgeneration rate is still 60.4%.
  • Largest classes. rc+vv+etm (4,816) is the single largest class, followed by adjectival modifiers modif+adj+etm (3,394), bound-noun constructions modif+etm+nnb (2,774), copular modifiers (842), and collocations (563).
  • Reliable automatic labels. On a category-stratified sample of 85 instances, two annotators agreed on 78 (91.8%; 95% Wilson CI 84.0–96.0%; Cohen’s κ = 0.677). Among the 77 instances with two definite judgments, the automatic label was accepted in 70 (90.9%). Errors came mainly from predicate-class ambiguity and the boundary between bound-noun constructions and relative-clause-like modifiers.
Table 3. Distribution of ETM types in the KLUE training data. (Table 3 in the paper.)
TypeCount
colloc+etm563
modif+adj+etm3394
modif+etm+ttae119
modif+etm+nnb2774
modif+ndaneun+etm68
modif+su-iss+etm65
modif+vcp+etm842
modif+vx+vx+etm3
rc+vv+etm4816
rc+vv+vx+etm293
rc+va+vx+etm26
undefined83
Total13,046

BibTeX

@misc{park2026annotatingkoreanadnominalending,
      title={Annotating Korean adnominal ending constructions in corpus data: Beyond relative-clause identification}, 
      author={Jungyeul Park and Chulwoo Park},
      year={2026},
      eprint={2607.03681},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.03681}, 
}
More projects from KAIST MLP Lab →