README.md

---
annotations_creators:
- machine-generated
language_creators:
- found
- machine-generated
language:
- uk
license:
- cc-by-4.0
multilinguality:
- monolingual
size_categories:
- 10K<n<100K
source_datasets:
- original
task_categories:
- text-classification
- feature-extraction
task_ids:
- semantic-similarity-classification
- topic-classification
pretty_name: "verba: Ukrainian Proverbs Corpus"
tags:
- linguistics
- cultural-heritage
- paremiology
- historical-spelling
- open-data
configs:
- config_name: default
  data_files:
  - split: train
    path: corpus.csv
dataset_info:
  features:
  - name: id
    dtype: string
  - name: text
    dtype: string
  - name: normalized_text
    dtype: string
  - name: modern_text
    dtype: string
  - name: keyword
    dtype: string
  - name: explanation
    dtype: string
  - name: category
    dtype: string
  - name: sources
    dtype: string
  - name: source_refs
    dtype: string
  - name: variant_group
    dtype: string
  splits:
  - name: train
    num_bytes: 15427267
    num_examples: 48787
---

# Data Card — verba: Ukrainian Proverbs Corpus

**Version:** 1.0.2 · **Released:** 2026-06-24 · **Author:** Dmytro Yemelianov ([ORCID](https://orcid.org/0009-0002-9244-7426))
**Home:** https://verbacorpus.org · **Repo:** https://github.com/dmytro-yemelianov/verbacorpus

---

## Motivation

Ukrainian paremiology lacked a single unified, machine-readable corpus that (a) attributed every proverb to its documented historical source, (b) preserved original orthography alongside a modern-spelling rendering, and (c) carried thematic labels usable for search and NLP. Existing digitizations were isolated (PDFs, scanned books, siloed databases) and none covered the full span from 1841 to the present.

**verba** was created to fill that gap: a canonical, deduplicated, source-attributed corpus drawn from five major published collections spanning 185 years of Ukrainian paremiology. It is built and maintained by Dmytro Yemelianov as an open scholarly resource. The dataset is enriched with modern-spelling renderings and a 27-theme taxonomy generated by batched Claude Code agents, then quality-audited. It powers a searchable PWA and multi-format REST API at https://verbacorpus.org.

---

## Composition

**48,787 instances** — one row per proverb or adage. Language: `uk` (Ukrainian).

### Schema (10 columns, `corpus.csv`)

| Column | Meaning |
|---|---|
| `id` | Stable identifier (`pNNNNNN`) |
| `text` | Verbatim proverb in its source orthography (never modified) |
| `normalized_text` | Lowercased, punctuation-stripped match key |
| `modern_text` | Modern standard Ukrainian spelling (LLM-generated) |
| `keyword` | Lemma/term (Franko), if any |
| `explanation` | Scholarly note (Franko-preferred), cleaned |
| `category` | 1–3 theme keys from the 27-theme taxonomy, `;`-joined, primary first |
| `sources` | `;`-joined source citation keys |
| `source_refs` | `;`-joined per-source references |
| `variant_group` | Id linking probable dialectal variants |

### Per-source counts

| Source | Entries | Notes |
|---|---|---|
| Franko 1901 — Іван Франко, *Галицько-руські народні приповідки* | 30,906 | Primary source; carries scholarly explanations |
| Номис 1864 — Матвій Номис, *Українські приказки, прислів'я і таке інше* | 9,785 | 1864 orthography preserved; best-effort OCR (~75–80%) |
| Бобкова — В.І. Бобкова та ін. (упоряд.), *Українські народні прислів'я та приказки* | 5,613 | Modern Ukrainian; tesseract-OCR'd from PDF |
| Ількевич 1841 — Григорій Ількевич, *Галицкіи приповѣдки и загадки* | 2,702 | Oldest collection (1841) |
| Млодзинський 2009 — *Практичний російсько-український словник приказок* | 2,261 | Modern bilingual collection |

### Additional statistics

- **With explanation:** 30,532 (62.6%) — drawn primarily from Franko 1901 critical apparatus
- **With modern_text:** 48,787 (100%) — all entries carry a modern-spelling rendering
- **Categorized:** 48,787 (100%) — all entries carry 1–3 thematic categories from the 27-theme taxonomy
- **Variant groups:** 5,064 (2,638 contain Nomis variants linked to other sources)
- **No train/test splits** — this is a reference corpus, not a benchmark dataset
- `modern_text`, `category`, cleaned `explanation`, and tuned `variant_group` are **LLM-generated data artifacts** (not deterministically reproducible; see `enrich/REPORT.md`)

---

## Collection Process

The corpus unifies five digitized collections via distinct ingestion pipelines:

**Historical OCR sources (Nomis 1864, Bobkova):**
- `pdftoppm -r 300` → tesseract 5.5.2 (`tessdata_best` ukr model, `--psm 6 --oem 1`) over page images
- For Nomis 1864: per-column crop (`expand/nomis_ocr.py`) because the 1864 edition is two-column; batched sonnet LLM extraction of proverb text + modern_text from the critical apparatus
- For Bobkova: rule-based segmentation, de-hyphenation, hunspell-flagged residuals cleaned by haiku agents; 36 unrecoverable rows dropped

**Other sources** (Franko 1901, Ilkevich 1841, Mlodzynskyi 2009) were ingested from existing digital transcriptions.

**Deduplication and merging:** exact-match on `normalized_text` + rapidfuzz `token_set_ratio ≥ 85` fuzzy variant linking. Nomis 1864 merge: 1,442 proverbs exact-merged into existing entries (adding `Nomis1864` as a second attestation), 2,232 cross-source variant-linked, 5,914 genuinely distinct.

See [`expand/REPORT.md`](expand/REPORT.md) for per-source ingestion details and [`enrich/REPORT.md`](enrich/REPORT.md) for enrichment pipeline details.

---

## Preprocessing / Cleaning / Labeling

All preprocessing is non-destructive — the original `text` field is never modified.

**`normalized_text`** — deterministic: lowercase + punctuation-stripping, used as deduplication/merge key only.

**`modern_text`** (LLM artifact) — generated by batched Claude sonnet agents (Pass B) converting source orthography to modern standard Ukrainian spelling. For Bobkova (modern source) and Mlodzynskyi 2009, `modern_text` equals cleaned `text`. Quality: ~95% acceptable (quality audit n=40; see `enrich/REPORT.md`).

**`category`** (LLM artifact) — 1–3 theme keys from the fixed 27-theme taxonomy, generated by batched Claude haiku agents (Pass A). ~85% acceptable; ~15% are debatable or wrong, clustering where the proverb's theme falls outside the 27-key vocabulary or where secondary tags are over-eager. The primary tag is the most reliable. Out-of-taxonomy keys from agents were dropped deterministically at merge time.

**`explanation`** (LLM-cleaned) — scholarly notes drawn from Franko 1901 critical apparatus, cleaned by haiku agents. Present for 30,532 entries.

**`variant_group`** (computed) — rapidfuzz `token_set_ratio ≥ 85` grouping, capped: groups larger than 8 members dissolved to prevent over-linking. Final: 5,064 groups. Groups are link-only and non-destructive (records are never merged or modified).

The 27-theme taxonomy is fixed and defined in [`enrich/taxonomy.csv`](enrich/taxonomy.csv):
work_labor, poverty_wealth, food_hunger, drink_alcohol, family_kinship, marriage_gender, speech_lying, wisdom_folly, fate_luck, time_seasons, death_illness, religion_god, social_relations, class_power, justice_truth, animals, body_health, home_household, conflict_enmity, friendship_love, travel_distance, trade_money, ethnic_local, emotion_mood, nature_weather, appearance_reputation, idiom_expressive.

---

## Uses

**Suitable uses:**
- NLP research: text normalization, historical spelling variation, paremiology, language modeling
- Linguistics and cultural studies: diachronic analysis of Ukrainian folk wisdom (1841–present)
- Education: searchable reference for Ukrainian proverbs with explanations and thematic access
- Cultural preservation: unified, attributed digital record of five major published collections
- Search and retrieval: the live API at https://verbacorpus.org supports lexical and semantic search

**Cautions and limitations:**
- OCR noise persists in `text` for Nomis 1864 (best-effort ~75–80% character fidelity) and residual artifacts in Bobkova (~2–3%)
- `modern_text`, `category`, and cleaned `explanation` are LLM-generated and orientational, not gold-standard labels — do not use as ground truth for model training without independent verification
- The primary category tag is more reliable than secondary tags; ~15% of any tag may be debatable
- Variant groups are link-only and heuristic; they should not be treated as definitive orthographic equivalence

**Out-of-scope:**
- The corpus is not a benchmark dataset and has no train/test splits
- Bobkova and Mlodzynskyi 2009 texts remain under their publishers' rights; included for research/education per `sources` attribution; removed on request

---

## Distribution

- **GitHub Releases:** CSV, JSON, JSONL, XML assets at https://github.com/dmytro-yemelianov/verbacorpus/releases/tag/v1.0.0 (see `scripts/release.sh` for the build)
- **REST API:** multi-format (`json`, `jsonl`, `xml`, `csv`, `tsv`) at https://verbacorpus.org/api/v1 — search, filter, export, semantic search; OpenAPI 3 spec at `/api/v1/openapi.json`
- **Croissant metadata:** [`croissant.json`](croissant.json) at repo root (MLCommons schema.org JSON-LD)

**Licensing (layered):**

| Layer | Content | License |
|---|---|---|
| Compilation + enrichment | `modern_text`, `category`, cleaned `explanation`, `variant_group`, the unified corpus structure | **CC BY 4.0** |
| Historical texts | Franko 1901, Nomis 1864, Ilkevich 1841 | Public domain |
| Modern collections | Bobkova, Mlodzynskyi 2009 | Texts remain under publishers' rights; included for research/education, attributed per `sources`, removed on request |

Cite as: Yemelianov, Dmytro (2026). *verba — Ukrainian Proverbs Corpus* (v1.0.2). https://verbacorpus.org

---

## Maintenance

**Maintainer:** Dmytro Yemelianov ([ORCID 0009-0002-9244-7426](https://orcid.org/0009-0002-9244-7426))

**Versioning policy** (semantic versioning for datasets):
- `MAJOR` — schema changes or breaking restructuring
- `MINOR` — new source added or significant corpus additions
- `PATCH` — corrections to existing entries, OCR fixes, category corrections

**Reporting issues:** GitHub Issues at https://github.com/dmytro-yemelianov/verbacorpus/issues. Include the entry `id` and the nature of the error (OCR, category, modern_text, variant group).

**Update cadence:** as new historical sources are digitized and ingested or as corrections accumulate. The tesseract + LLM pipeline (`expand/`) is reusable for additional archive.org sources.

**Reproducibility note:** `modern_text`, `category`, and `variant_group` are committed LLM artifacts. Regenerating them requires Claude Code; the outputs are not byte-reproducible but quality is audited (see `enrich/REPORT.md`).

---

## Known Limitations

- **Nomis 1864 OCR (~75–80% character fidelity):** two-column 1864 scan; per-column crop + tesseract errors persist in `text`; rare LLM normalization toward familiar forms (~1–2%). Identifiable via `Nomis1864` in `sources`. See [`expand/REPORT.md`](expand/REPORT.md).
- **Category accuracy (~85% acceptable):** ~15% of tags are debatable or wrong; misses cluster where the theme falls outside the 27-key vocabulary or where secondary tags are over-eager. Primary tag is the most reliable. See [`enrich/REPORT.md`](enrich/REPORT.md).
- **`modern_text` accuracy (~95% acceptable):** two wrong cases in quality audit (n=40): an un-modernized archaic future construction and a dropped prefix; minor cases retain dialectal reflexive `ся` intentionally.
- **Variant groups are link-only and heuristic:** groups larger than 8 were dissolved; threshold (85) and cap are tunable. Not a definitive orthographic equivalence claim.
- **Bobkova residual OCR (~2–3%):** front-matter fragments and cross-page hyphen breaks that per-page segmentation cannot rejoin.
- **Enrichment is not deterministically reproducible:** regenerating requires Claude Code agents; committed artifacts are the authoritative version.

---

## Sources

The corpus unifies text from five published collections. Full bibliographic data is in [`references.bib`](references.bib) and [`references.csl.json`](references.csl.json).

1. Іван Франко (зібрав, упоряд. і пояснив). Галицько-руські народні приповідки. Етнографічний збірник, тт. X, XVI, XXIII, XXVII, XXVIII (у 3 т., 6 вип., 1901–1910). Львів: Накладом Наукового товариства імені Шевченка, 1901. https://www.i-franko.name/uk/Folklore/1901/GalRusProverbs.html
2. Матвій Номис (упоряд.). Українські приказки, прислів'я і таке інше. Санкт-Петербург: Друкарня Тиблена і Ко (К. Куліша), 1864. https://archive.org/details/nomis1864
3. Бобкова В., Лавров Ф., Ліздвой М., Сухобрус Г., Ткаченко Ф. (упоряд.). Українські народні прислів'я та приказки. Вид. 2-ге, випр. і доп. Київ: Видавництво АН УРСР, 1961.
4. Григорій Ількевич (зібрав). Галицькі приповідки і загадки. Відень: Друкарня оо. Мехитаристів, 1841. http://irbis-nbuv.gov.ua/dlib/item/0000436
5. Млодзинський Г. (упоряд.). Практичний російсько-український словник приказок. Репринт вид. 1929 р. Київ: Інститут енциклопедичних досліджень НАН України, 2009. ISBN 978-966-02-5147-2. https://repository.encyclopedia.kyiv.ua/items/show/48

---

## Citation

```
Yemelianov, Dmytro (2026). verba — Ukrainian Proverbs Corpus (v1.0.2). https://verbacorpus.org
```

Machine-readable: [`CITATION.cff`](CITATION.cff) (Citation File Format 1.2.0).