/**
 * Subword features for fastText-style embeddings.
 *
 * Korean is the weak spot of plain word2vec here: inflected/rare forms
 * (후킹, 후킹을, 후킹하면) are distinct tokens, most below the frequency floor, so
 * they get no vector. The fix is to represent every word as the sum of its
 * character n-grams — and for Hangul, to first DECOMPOSE each syllable into its
 * jamo so morphologically related words actually share substrings
 * (후 → ㅎㅜ, 킹 → ㅋㅣㅇ). A rare or even unseen Korean word then still gets a
 * composed vector from subwords it shares with common words.
 *
 * Tokenisation is untouched — subwords are an internal representation only — so
 * the embedding vocabulary stays aligned with the search vocabulary.
 */
/** Decompose Hangul syllables to jamo; pass everything else through unchanged. */
export declare function decomposeHangul(text: string): string;
/**
 * The subword set for a token: character n-grams (minN..maxN) over the
 * jamo-decomposed, boundary-marked word, plus the whole word as a distinguished
 * feature. Returns deduped feature strings.
 */
export declare function wordSubwords(word: string, minN?: number, maxN?: number): string[];
/**
 * Derive a subword vector layer from already-trained word vectors: each subword
 * vector is the centroid of the (semantic) vectors of in-vocab words that
 * contain it. This adds coverage for rare / unseen words — their composed vector
 * is the average of the centroids of the substrings they share with known
 * words — WITHOUT perturbing the trained word vectors. Subwords occurring in
 * fewer than `subwordMinCount` words are dropped.
 */
export declare function buildSubwordCentroids(vocab: string[], vectors: Float32Array, dim: number, minN: number, maxN: number, subwordMinCount: number): {
    subwordVocab: string[];
    subwordVectors: Float32Array;
    subwordCounts: number[];
};