INTERNALS · MORPHOLOGY

Morphology

Analysis direction

A query atom contains a lemma and an optional part of speech. Lexicon lookup produces compatible analyses. Each analysis compiles into a CandidateProgram containing an anchor, core span, particle and ending consumption, boundary conditions, and provenance.

Only source positions containing an anchor are verified against the program. This differs from analyzing every observed token. kfind returns spans reachable from the query and their origins, not a complete morpheme sequence for the sentence.

Lexicon layers

The core lexicon stores essential irregulars, POS ambiguity, function words, and surface exceptions. The enriched layer contains inflection and derivation relations supported by pinned Korean dictionary snapshots. The full-POS lexicon supplies automatic and fine-grained POS candidates, while the user lexicon adds project terms and replacement rules.

Multiple analyses for one lemma are preserved. Programs with identical execution conditions may be merged, but every POS and rule origin remains in provenance. An explicit coarse POS also receives supplemental analyses for missing fine-grained categories.

Particles and allomorphs

Nominals do not enumerate every particle chain as a complete surface. After a core hit, a verifier reads the remaining suffix through a particle transition table. Vowel-final hosts select 는·가·를·로, consonant-final hosts select 은·이·을·으로, with a dedicated condition for 로 after final ㄹ.

학교 + 에서 + 는  → 학교에서는
길 + 으로       → 길로
집 + 로         → rejected

Only registered particle transitions are accepted. smart and token verify the token boundary after consumption.

Endings and prefinal endings

Predicate programs separate stem substitution from ending chains. Prefinal endings express tense, honorific meaning, and modality before connective or terminal endings. 먹었겠지만, for example, is 먹/VV + 었/EP + 겠/EP + 지만/EC. kfind represents 었 and 겠 as typed consumption states and closes the token with the permitted 지만 transition.

The rule catalog includes present declaratives, adnominal and nominalizing endings, connectives, and bounded continuations. A chain that is absent from the versioned rule data is not inferred merely because it appears grammatically plausible.

Irregular conjugation and contraction

A lexicon entry selects a substitution class, and the generator combines it with the phonological environment of an ending. ㄷ irregular 걷다 yields 걷 + 어 → 걸어; ㅂ irregular 돕다 yields 돕 + 아 → 도와. The same composition handles 모르다 → 몰라, 낫다 → 나아, and 파랗다 → 파란.

Contractions such as 보아 → 봐 and 주어 → 줘 decompose and recombine Hangul syllable components under explicit rules. When both regular and irregular analyses are valid, both programs are retained for recall.

Derivation and compound structure

inflection preserves POS. derivation also includes registered suffixes that create a new POS. Paths such as 안정 + 하 + 다 and 잠식 + 당하 + 기 are accepted only when source components prove a complete POS and span sequence.

Auxiliary and compound-internal search uses the same constraint. Internal 가다 in 들어가다 requires a preceding predicate and connective ending, while particle 가 in 친구가 cannot become a verb candidate.

A compound-noun decomposition is used when the source declares components for the complete particle host and no independent analysis competes on the same span. This finds 물 in 물/NNG + 줄기/NNG + 는/JX but rejects 길 from an alternate decomposition that competes with an independent 산길/NNG analysis.

When the resource completes an auxiliary chain left by a generated program, the VX + E* path after the connective must be complete. A source path is used under any of these conditions.

• The connective lies outside the core.
• Every exact analysis of the contracted core is a predicate.
• A query-compatible single-predicate source edge extends beyond the core and is followed by a complete ending path.
• The complete token has an exact predicate + EC + VX + E+ analysis.
• The suffix belongs to the resultative -아/어지다 family.

These conditions retain 빼놓을, 생겨났던, and 극심해지겠지만 without expanding 해가며 from the ambiguous 해 surface alone.

Local structural verification

smart decodes only bounded tokens around an anchor from the compact component resource. A StructuralConstraint compares the POS sequence, component spans, continuations, and neighboring-token conditions required by the query. It uses only observable morphology and does not interpret meaning.

Because no corpus-wide tokenization is built, large-file scanning stays on the byte-search path. Source-proven internal components and particle-ending boundaries still make the result more precise than a plain substring search.

For a nominal followed by particles, the exact whole-nominal path is selected first. 길 found only in an alternative decomposition of 산길을 is rejected, while the 주의 component declared by the whole analysis of 자본주의 is retained.