K-UniMorph: Korean Morphology Integration

• K-UniMorph Project is a comprehensive initiative that formalizes Korean verbal morphology within a universal morphology framework.
• It maps over 43,000 verb lemmas and nearly 245,000 inflected forms from the Sejong corpus to a standardized feature schema.
• The project advances computational linguistics by benchmarking subword models and enabling improved morphological analysis for NLP tasks.

The K-UniMorph Project is a comprehensive endeavor to integrate Korean language morphology into the global Universal Morphology (UniMorph) paradigm, addressing a longstanding gap in the availability of typologically normalized, computational morphological resources for Korean. Drawing upon the principles, standards, and methodologies established by the UniMorph project and related efforts, K-UniMorph formalizes Korean verbal morphology using a universal feature schema, constructs an extensive lexical resource derived from the Sejong corpus, and contributes methodologies and evaluations relevant both to computational linguistics and cross-lingual morphological analysis (Jo et al., 2023, McCarthy et al., 2018, Batsuren et al., 2022, Kirov et al., 2018).

1. Objectives and Theoretical Framework

K-UniMorph’s primary objective is the systematic extraction, annotation, and dissemination of Korean morphological paradigms in conformance with the cross-lingual UniMorph feature schema. Prior to this initiative, Korean was notably absent from both UniMorph’s core database and international shared tasks such as SIGMORPHON (Jo et al., 2023). The project aims:

• To counter the underrepresentation of Korean in universal morphological resources.
• To render Korean’s complex verbal feature system—encompassing honorifics, tense, mood, evidentiality, and voice—fully compatible with the UniMorph schema as defined by Sylak-Glassman et al. (2015, 2016).
• To enable downstream tasks such as morphological inflection, reinflection, homograph disambiguation, and morphological tagging using resources and tools validated in previous UniMorph releases (McCarthy et al., 2018, Kirov et al., 2018).

These goals are situated within a broader theoretical context that emphasizes the value of harmonizing type-level (dictionary/paradigm) and token-level (corpus/treebank) resources for both computational and descriptive morphology (McCarthy et al., 2018).

2. Data Sources and Extraction Methodology

K-UniMorph is constructed from the Sejong morphologically analyzed corpus, which comprises 676,951 sentences, 7,835,239 eojeols, and 9,537,029 morpheme tags (Jo et al., 2023). The extraction pipeline operates as follows:

1. Sejong’s original POS tag sequences (XPOS) are mapped to Universal POS tags (UPOS) via a manually calibrated lookup table, cross-referenced against UD_Korean-GSD (Jo et al., 2023).
2. Manual correction and classifier retraining are performed based on a sample of 300 development sentences (738 XPOS, 705 UPOS mismatches corrected), driving UPOS prediction error for verbs below 0.1%.
3. Eojeols assigned UPOS=VERB are processed; the first morpheme is designated as the lemma, with subsequent morphemes forming the inflectional suffix sequence.
4. Deduplication removes form-lemma-feature triples seen in both train+dev and test splits.

The resulting resource covers 43,959 verb lemmas and 244,871 unique inflected verb forms. Data is separated into train, dev, and test, following conventions established by Park & Tyers (2019):

Split	#lemmas	#inflected forms
train	41,631	197,774
dev	7,505	19,251
test	7,595	27,846

Lemmatization policy (first morpheme as lemma) and splitting methodology yield higher coverage and granularity than lexicographic or hand-compiled resources previously available for Korean (Jo et al., 2023).

3. Feature Schema and Paradigm Construction

K-UniMorph adopts the UniMorph feature schema, which enumerates over 200 atomic feature values across 23 morphological categories in its global instantiation (Batsuren et al., 2022, Kirov et al., 2018). For Korean verbs, the subset of relevant features is:

$$F = \{\mathrm{Evidentiality},\, \mathrm{Interrogativity},\, \mathrm{Mood},\, \mathrm{Tense},\, \mathrm{Voice},\, \mathrm{Politeness}\}$$

Where

• $$\mathrm{Evidentiality} \in \{\text{hrsy},\,\text{infer}\}$$
• $$\mathrm{Interrogativity} \in \{\text{decl},\,\text{int}\}$$
• $$\mathrm{Mood} \in \{\text{real},\,\text{irr},\,\text{purp},\,\text{oblig}\}$$
• $\mathrm{Tense} \in \{\text{prs},\,\text{pst}\}$
• $\mathrm{Voice} \in \{\text{caus},\,\text{pass}\}$
• $\mathrm{Politeness} \in \{\text{pol}\}$

Every inflected form is annotated as a bundle $(f_1=v_1,\,f_2=v_2,\ldots,f_6=v_6)$, aligning Korean’s rich inventory of verbal endings with schema-conformant morphological tags. Mapping from endings to feature values is systematized; for instance, the past tense “pst” is marked by the non-final ending EP -eoss-, as in $$\textrm{naseo} + \textrm{eoss} + \textrm{da} \rightarrow \textrm{naseossda}$$ (v;decl;pst).

Feature	Value	Ending Class	Example Suffix
Tense	prs	EF	–nda
	pst	EP	–eoss
Mood	real	ETM	–nun
	irr	ETM	–eul
Evidentiality	hrsy	EF	–lae
	infer	EP	–gess
Voice	caus	EC	–ge
	pass	stem infix	–hi, –li
Interrogativity	int	EF	–kka, –ga
	decl	EF	–da
Politeness	pol	EP	–si

Paradigm formation leverages the cross-product of stem set $S$ and feature-value sets $$\mathrm{Evidentiality} \in \{\text{hrsy},\,\text{infer}\}$$0:

$$\mathrm{Evidentiality} \in \{\text{hrsy},\,\text{infer}\}$$1

With $$\mathrm{Evidentiality} \in \{\text{hrsy},\,\text{infer}\}$$2 and $$\mathrm{Evidentiality} \in \{\text{hrsy},\,\text{infer}\}$$3, the theoretical paradigm space is $$\mathrm{Evidentiality} \in \{\text{hrsy},\,\text{infer}\}$$4 million, but the observed set is filtered to $$\mathrm{Evidentiality} \in \{\text{hrsy},\,\text{infer}\}$$5245,000 existing forms due to language-specific combinatorial constraints (Jo et al., 2023).

4. Evaluation and Benchmarking

K-UniMorph includes systematic evaluation using morphological inflection tasks at three granularities: letter, syllable, and morpheme levels (Jo et al., 2023).

• SIGMORPHON 2018 baseline and a Transformer neural model (fairseq) are benchmarked.
• Morpheme-level input/output outperforms both letter and syllable modeling, reflecting Korean's agglutinative structure.

Model	Letter (%)	Syllable (%)	Morpheme (%)
Baseline	26.88	27.75	31.29
Neural	51.97	49.72	54.26

This substantiates the conclusion that aligned subword modeling is essential for adequate handling of Korean morphology in computational systems.

Furthermore, the K-UniMorph resource is integrated with methods for deterministic UD–UniMorph mapping, as established in (McCarthy et al., 2018). The core converter achieves a macro-average recall of 64.13% (post-edit) in matching UD token-level and UniMorph type-level annotations across 31 languages, demonstrating practical interoperability.

5. Resource Format and Availability

Data are released in CSV format as tuples (lemma, feature_bundle, inflected_form), facilitating direct input to morphological analyzers and reinflection systems.

Example: $$\mathrm{Evidentiality} \in \{\text{hrsy},\,\text{infer}\}$$6

The full resource, code, and splits are distributed via GitHub at https://github.com/jungyeul/K-UniMorph (Jo et al., 2023).

Quantitative characteristics:

• 43,959 verb stems
• ≈245,000 inflected forms
• Average of 8.285 forms per lemma, 5.634 per (lemma, feature-bundle)

These figures confirm K-UniMorph’s scale and comprehensiveness compared to legacy Korean resources.

6. Broader Implications and Extensions

The introduction of K-UniMorph enables the Korean language to be fully represented in UniMorph and SIGMORPHON shared tasks, supporting:

• Paradigm-constrained parsing and tagging pipelines
• Homograph-aware tagging by enumeration of all possible morphological readings for surface forms
• Morphological analysis and generation for downstream tasks in machine translation and low-resource NLP (Jo et al., 2023, McCarthy et al., 2018)

Future work includes:

• Extending coverage to nominal paradigms (case, information structure, topic, comparative marking)
• Annotating valency alternations and additional morphological categories such as quotative evidentiality and honorific lexemes
• Cataloguing case stacking and polypersonal agreement phenomena, enabled by UniMorph’s hierarchical schema (Batsuren et al., 2022)

The K-UniMorph Project thus not only enriches the comparative study of world morphology, but also exemplifies the methodological rigor achievable by close alignment with universal, cross-lingual annotation conventions.