ko-finance-asr-corrections

Frequency-annotated Korean ASR confusion pairs from finance/stock YouTube.

• 210 pairs
• mined from 2,391 videos of auto-captions
• across 47 channels
• totalling 1,080.1 hours

Each pair carries how often the term was mangled and how often it was said correctly, plus verification provenance.

한국어 금융·주식 유튜브 자동자막에서 실측한 ASR 오인식→교정 쌍입니다. 모든 쌍에 오표기·정답 표기 빈도(→ 용어별 오인식률)와 검증 메타데이터(2-LLM 합의 감사, 승격 티어)가 붙어 있습니다.

What makes it different

No public "confusion pair dictionary with observed frequencies" exists in any language (as of a 2026-08 survey), and Korean ASR error-correction data is nearly nonexistent — the only prior public pair set is KEBAP (2,478 pairs, error-type labels, frozen since 2023) and no finance-domain resource exists. Frequency is what makes a pair list actionable: it says which errors dominate real traffic:

• 펀더멘탈 → 펀더멘털 ×1,151
• 하이니스 → 하이닉스 ×904
• 변합기 → 변압기 ×682
• FMC → FOMC ×1,106

Those four were 449, 342, 333 and 270 one snapshot ago; the counts grow with the corpus, which is why every number in this card carries a marker the release checker verifies.

Fields

wrong, right, corpus_count, observed_count, tier, evidence, category, auditor_models, approved_at, word_boundary, apply_scope. Field-by-field definitions: docs/SCHEMA.md.

How it was verified

No gold labels exist for this task, so verification is a loop, not an annotation pass: two independent LLM auditors judge each candidate, only consensus is adopted, an external registry check confirms stock names, and pairs are promoted through tiers with instant rollback. Details and per-round numbers: docs/METHODOLOGY.md.

Error rates, not just frequencies

Each pair carries corpus_count (the misrecognized form) and right_count (the verified one) from the same scan, so a row states how often the term came out wrong: FMC → FOMC 74.8% , 엔트로픽 → 앤트로픽 87.3% . 23 pairs have right_count: 0 — across all 2,391 videos the captions never once produced the correct spelling. Stock rows also carry ticker and market, so they join to price data without matching on a name.

data/withdrawn.json lists pairs that shipped and were later removed, with how many of their matches were genuine and which narrower key replaced them. A withdrawal says the key was unsafe as written, not that every correction it made was wrong.

Limitations (read these)

One corpus, one ASR system, one domain; 210 pairs is a seed, not a lexicon; frequency-ranked mining skews toward what these channels talk about; corpus_count counts surface forms, not confirmed errors; some pairs encode a spelling standard rather than a mishearing; tier B pairs were verified by auditor consensus, not by a human; no person-name pairs and no source sentences ship, by policy. The full list is in the repository README — read it before using the data as a general Korean ASR error list, because it is not one.

Benchmark

A standard-library scorer, a 1,100-item synthetic evaluation set, and three dictionary baselines live in the GitHub repository. Fixes and over-corrections are scored together, because a correction dictionary that fixes 100 errors while damaging 20 correct sentences is not a good dictionary.

License and citation

Data: CC BY 4.0. Code (benchmark, scripts): MIT. Cite the concept DOI 10.5281/zenodo.22700919, which always resolves to the latest snapshot; machine-readable metadata is in CITATION.cff in the GitHub repository.

Produced by the 껄무새 pipeline.