Yuyongkim/inconvenience-msds

inconvenience-msds

Inconvenience #01 — Material Safety Data Sheets in Korean braille

Accessibility infrastructure for Korean chemical safety information. 48,966 chemicals × up to 16 MSDS sections, encoded as Korean braille following the 2017 한국 점자 규정 (Korean Braille Standards).

What this is

This dataset is the first entry in the inconvenience series — a planned sequence of accessibility-infrastructure datasets targeting domains where the visually impaired user population is small but the consequences of inaccessibility are large.

The most common objection to producing domain-specific braille infrastructure is that the expected user population is small. We argue this observation is not evidence against infrastructure but evidence for it: low observable participation in a profession is a predictable consequence of inaccessible professional information, not an independent fact about preference or capability. The legal basis for producing this infrastructure is already in place under UN CRPD Article 9 and Korea's 장애인차별금지법 제21조; what was missing was the artifact itself.

The naming is deliberate: the inconvenience belongs to the user, not the work. Each entry tries to remove one.

Quick start

from datasets import load_dataset

ds = load_dataset("Yuyongkim/inconvenience-msds", split="train")
print(ds[0]["name_ko"])
print(ds[0]["sections"][1]["braille"][:200])

Source

Korea Occupational Safety and Health Agency (KOSHA) public MSDS database, retrieved via the Korea Public Data Portal API (data.go.kr) under an authorized operational account.

Versions

The braille side of this corpus changed after a rule-by-rule audit of the encoder against the source standard. Pin the version you need.

Cite v1.0 with the paper. The brief report describes the v1.0 artifact, and its figures (232.3M cells, ratio 1.95) refer to that file.

Use v1.2 for new work. The audit corrected the parenthesis cells to the two-cell 2017 form (제54~56항), added the roman terminator (제30항), and implemented the abbreviations of 제12~18항 together with the initial-ㅇ omission of 제2항. The corpus is shorter because those provisions replace multi-cell spellings with single cells. Note that this moves the corpus off Grade 1: v1.0 is uncontracted throughout, while v1.2 uses the standard abbreviations.

See notes/2026-08-13-regulation-audit.md in the code repository for the per-article comparison, and paper/brief_report/errata_2026-08-14.md for what changed relative to the published report.

Statistics

Figures below are for v1.0, the version the paper describes.

MSDS section coverage

Up to 16 standardized sections per chemical (KOSHA convention):

Schema

Each record is a JSON object:

{
  "chem_id": "001008",
  "name_ko": "벤젠 (Benzene)",
  "cas_no": "71-43-2",
  "name_en": "Benzene",
  "sections": [
    {
      "section_no": 1,
      "title": "화학제품과 회사에 관한 정보",
      "text_ko": "제품명: 벤젠 ...",
      "braille": "⠨⠝⠙⠍⠢⠑⠱⠶⠐⠂..."
    }
  ],
  "total_text_chars": 6123,
  "total_braille_chars": 11920
}

Braille encoding

Korean braille is encoded as Unicode braille characters (U+2800–U+28FF) following the 2017 Korean Braille Standards:

• —Hangul syllables: decomposed into initial consonant (초성) + vowel (중성) + optional final consonant (종성), each mapped to a fixed dot pattern.
• —Number indicator ⠼ (dots 3-4-5-6) before digit sequences.
• —Roman letter indicator ⠴ (dots 3-5-6) before Latin characters.
• —Capital indicator ⠠ (dot 6) before uppercase letters.
• —GHS pictograms: filename references (e.g., GHS06.gif) are resolved to descriptive Korean text (e.g., 급성독성) before encoding, so braille output conveys hazard semantics rather than an unreadable filename.

All 147 unique GHS H/P statement codes occurring in the database (63 H-statements, 84 P-statements) are resolved and encoded.

Validation

The encoder is deterministic and lossless on its input; identical input always produces identical output. Validation independent of human transcribers is appropriate for the deterministic encoding regime; a reading-comprehension user study with visually impaired readers is identified as essential follow-up work.

Use cases

• —Accessibility infrastructure: drop-in source for refreshable braille displays and braille embossers reading Korean MSDS
• —Research: largest Korean-language braille corpus to date — usable for braille NLP, accessibility tooling, mixed-script handling research
• —Cross-lingual chemical safety: CAS numbers in records bridge to PubChem, ECHA, EPA, OSHA registries
• —Policy artifact: a concrete demonstration of feasibility, against which procurement and regulatory adoption proposals can be made

Output formats

In addition to this JSONL, the upstream project provides:

• —Unicode braille .txt (U+2800–U+28FF) — for refreshable displays
• —BRF — for braille embossers (40 cells × 25 lines)
• —PDF — sighted-side preview / verification

See github.com/Yuyongkim/inconvenience-msds.

Citation

@article{kim2026inconveniencemsds,
  title   = {KOSHA-Braille: Infrastructure-Grade Accessibility for
             Korean Chemical Safety Information},
  author  = {Kim, Yu Yong},
  journal = {IEEE Access},
  year    = {2026},
  note    = {Submitted}
}

@dataset{inconvenience_msds_2026,
  title     = {inconvenience-msds: Korean Chemical Safety MSDS in Braille},
  author    = {Kim, Yu Yong},
  year      = {2026},
  url       = {https://huggingface.co/datasets/Yuyongkim/inconvenience-msds},
  publisher = {Hugging Face}
}

License

• —This dataset: CC BY 4.0
• —Source MSDS data: Korea Public Data Portal under 공공누리 제1유형 (Korea Open Government License Type 1: Attribution)
• —Encoder source code: MIT

Acknowledgements

Source data: Korea Occupational Safety and Health Agency (KOSHA). Independent encoder reference: hangul-braille-converter by hyonzin and hanbraille by delvier.