Lexically conditioned realization ambiguity in Korean predicate morphology

Why a sequence of Korean morphemes and POS labels does not always determine the surface form: homonymous predicates in different conjugation classes

arXiv 2026

Wonjun Oh, KyungTae Lim, Jungyeul Park*

Korea Advanced Institute of Science & Technology (KAIST), South Korea

*Corresponding author

Abstract

This paper examines Korean surface realization as distinct from morphological analysis. It asks whether a sequence of canonical morphemes and grammatical category labels uniquely determines the corresponding surface form. The answer is negative for a restricted but theoretically revealing class of Korean predicates. In these cases, formally identical or near-identical stem-ending configurations yield different outputs depending on lexical identity and realization class membership. We analyze this phenomenon as homonymy with inflectional divergence, focusing on regular versus digeut irregular pairs, regular versus bieup irregular pairs, and reu irregular versus reo irregular pairs. These cases show that stem shape and ending alone do not always determine surface realization. Instead, lexical meaning, subcategorization, and semantic role structure help identify the intended predicate; the predicate determines the realization class; and the realization class determines the surface form. Korean realization thus reveals a limit of bare morphological representation.

Keywords Korean morphology Irregular conjugation Surface realization Morphological generation Homonymy with inflectional divergence ㄷ / ㅂ / 르 / 러 irregular predicates Eojeol Korean NLP KLUE

Analysis

The paper treats realization as the inverse of morphological analysis: given canonical morphemes with category labels, can the grammar recover a unique eojeol (the Korean spacing unit)? It distinguishes three relations between analysis and realization:

  1. Nearly transparent: {먹+었+다} → 먹었다 meogeotda ‘ate’.
  2. Non-transparent but locally predictable: {나서+었+다} → 나섰다 naseotda ‘stepped forward’, once the lexical item and morphophonological conditions are known.
  3. Underdetermined: {걷+었+다} yields 걸었다 georeotda ‘walked’ or 걷었다 geodeotda ‘cleared away’, depending on the intended lexeme.

The third case is called homonymy with inflectional divergence: predicates with the same or nearly the same stem shape belong to different realization (conjugation) classes. Using subcategorization frames from the Sejong dictionary, the paper argues for an indirect chain of explanation:

argument frame / semantic role pattern → lexical identity → realization class → surface form

Semantic roles do not choose the surface form directly; they help identify the lexeme, which carries its conjugation class. Derived predicates such as 치걷다 chigeotda, 되묻다 doemutda, and 파묻다 pamutda inherit the behavior of their simplex base.

Main Findings

  • Three ambiguity classes. Realization ambiguity among simplex predicates falls into regular vs. ㄷ digeut irregular (걷다, 닫다, 묻다), regular vs. ㅂ bieup irregular (곱다, 굽다), and 르 reu vs. 러 reo irregular (누르다, 이르다); the inventory in Table 2 is exhaustive for the Standard Korean Language Dictionary.
  • Stem shape marks only the site of alternation. The final consonant and ending show where an alternation could occur, but only the lexeme's realization class decides whether it does, so bare morphological representation is not a reversible encoding of surface form.
  • Attested in corpus data. In the KLUE dependency treebank (train + dev), the 걷다 family has 6 regular and 124 ㄷ-irregular tokens, 묻다 6 and 19, 닫다 6 and 13; 이르다 has 13 르-irregular and 20 러-irregular tokens; no token of 곱다 was found.
  • Practical consequence for reconstruction. Rebuilding surface eojeols from morpheme and POS input over KLUE, a refined rule-based system with explicit realization-class constraints reaches 92.56 target-eojeol F1 and 85.59 sentence-level accuracy, compared with 64.96 F1 for a previous rule-based baseline and 73.42 F1 (54.57 accuracy) for GPT-5.5.
Table 2. Lexically distinct homonymous predicates whose realization classes differ (homograph numbers from the Standard Korean Language Dictionary in parentheses). The same stem and ending surface differently depending on the lexeme.
PredicateMeaningClassRealization examples
Regular vs. ㄷ digeut irregular
걷다 geotda (03)‘clear away’regular걷어 geodeo, 걷으니 geodeuni
걷다 geotda (02)‘walk’ㄷ digeut irregular걸어 georeo, 걸으니 georeuni
닫다 datda (02)‘close’regular닫아 dada, 닫으니 dadeuni
닫다 datda (01)‘run’ㄷ digeut irregular달아 dara, 달으니 dareuni
묻다 mutda (02)‘bury’regular묻어 mudeo, 묻으니 mudeuni
묻다 mutda (03)‘ask’ㄷ digeut irregular물어 mureo, 물으니 mureuni
Regular vs. ㅂ bieup irregular
곱다 gopda (01)‘suffer a loss’regular곱아 goba, 곱으니 gobeuni
곱다 gopda (02)‘be beautiful’ㅂ bieup irregular고와 gowa, 고우니 gouni
굽다 gupda (02)‘bend’regular굽어 gubeo, 굽으니 gubeuni
굽다 gupda (01)‘roast’ㅂ bieup irregular구워 guwo, 구우니 guuni
르 reu irregular vs. 러 reo irregular
누르다 nureuda (01)‘press’르 reu irregular눌러 nulleo, 누르니 nureuni
누르다 nureuda (02)‘be yellowish’러 reo irregular누르러 nureureo, 누르니 nureuni
이르다 ireuda (02)‘say’르 reu irregular일러 illeo, 이르니 ireuni
이르다 ireuda (01)‘reach’러 reo irregular이르러 ireureo, 이르니 ireuni

BibTeX

@misc{oh2026lexicallyconditionedrealizationambiguity,
      title={Lexically conditioned realization ambiguity in Korean predicate morphology},
      author={Wonjun Oh and KyungTae Lim and Jungyeul Park},
      year={2026},
      eprint={2608.27966},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.27966},
}
More projects from KAIST MLP Lab →