PiotrSty/justicecorp-slovak

Slovak court decisions (justicecorp_sk) - official published rulings

Slovak court decisions ("súdne rozhodnutia") published under the Open Government Partnership initiative of the Ministry of Justice SR, taken from the justicecorp od-justice v3.0 vertical export compiled by the Ľudovít Štúr Institute of Linguistics (JÚĽŠ SAV) from the open justice data of the Ministry of Justice SR.

• Source documents: 6,537,977
• Retained records: 961,401 (14.7%)
• Measured cl100k_base proxy tokens: 5,245,253,455
• Characters: 11,365,728,126
• Time span (document dates): 1995-11-22 - 2024-05-13
• Source snapshot: JÚĽŠ justicecorp od-justice v3.0, od-justice-3.0.ver.xz (28,383,907,552 bytes, sha256 961527e2908292578404812e72899b5ec092e4171ffb59f481b216fcf31c6507; corpus page https://www.juls.savba.sk/justicecorp.html, download https://www.juls.savba.sk/data/od-justice/od-justice-3.0.ver.xz)

License - evidence at the origin

• The texts are official works - court decisions published by the Ministry of Justice SR under the open government initiative (https://obcan.justice.sk/opendata). Under the Slovak Copyright Act No. 185/2015 Coll., section 5, official works of this kind are not protected by copyright [INFERENCE: statutory reading, not counsel-verified]. Per-record license field: LicenseRef-PD-Official.
• Per-record provenance in artifacts/attribution.jsonl.gz: the direct source document URL (doc.url, e.g. https://obcan.justice.sk/content/public/item/...), court (doc.court), file number (doc.zn), decision date (doc.date), upstream token counts (tokcount, tokcountdd), dropped duplicate-paragraph count and text_sha256.
• The JÚĽŠ SAV compilation (tokenization, morphological annotation with MorphoDita, syntactic annotation with UDPipe) carries no explicit license statement on the origin pages (checked 2026-10-05); this dataset ships only reconstructed plain texts plus metadata, not the annotations. release: None until the review ruling.

Anonymization

The source corpus is anonymized at origin: names of natural persons, dates of birth and addresses appear masked in the published decisions (e.g. X./. V. H., X.. XX.XX.XXXX). This dataset preserves that masking and adds residual pattern redaction (see Filtering); it performs no additional anonymization.

Samples

Verbatim records from artifacts/sample.jsonl, real line breaks preserved, truncated with [...].

Sample 1

id: justicecorp_sk_8d27a2507883ae5a
source: justicecorp_sk
created: 2024-03-10
token_count: 6776

Súd: Okresný súd Vranov nad Topľou Spisová značka: 2Nt/5/2024 Identifikačné číslo súdneho spisu: 8824010051 Dátum vydania rozhodnutia: 11. 03. 2024 Meno a priezvisko sudcu, VSÚ: Mgr. Marcela Holubová ECLI: ECLI:SK:OSVT:2024:8824010051.1 Uznesenie Okresný súd Vranov nad Topľou predsedníčkou senátu Mgr. Marcelou Holubovou v trestnej veci odsúdeného Y. O., B.. XX.XX.XXXX P. P. B. O., O. X. I. XXX/XX, P.c, okres Vranov nad Topľou, o žiadosti odsúdeného o zahladenie odsúdení, takto rozhodol: A) I. Podľa § 470 ods. 1 Trestného poriadku v spojení s § 92 ods. 1 písm. c), ods. 4 Trestného zákona [...]

Sample 2

id: justicecorp_sk_b55f35fec3e02a6d
source: justicecorp_sk
created: 2015-07-15
token_count: 585

Obžalovaný: X. O., nar. XX.XX.XXXX v J. Q., trvalým bydliskom O. O.-Q., K. Z. Č..X, nezamestnaný, slobodný, j e v i n n ý , ž e dňa 24.júna 2015 v čase o 22:45 hod., ako vodič po predchádzajúcom požití alkoholických látok viedol poškodené osobné motorové vozidlo továrenskej značky Š. XXX, ev.č. Z. XXX U. po rýchlostnej ceste R1 v meste Banská Bystrica, v smere jazdy Zvolen - Banská Bystrica, v protismere v km 161,04, kde bol zastavený a kontrolovaný hliadkou OKP Okresného riaditeľstva PZ Banská Bystrica [...]

Sample 3

id: justicecorp_sk_e9f7267d2595965a
source: justicecorp_sk
created: 2012-01-31
token_count: 3856

Krajský súd v Bratislave v senáte v zložení: predsedníčka senátu JUDr. Darina Kriváňová a členovia senátu: JUDr. Ayše Pružinec Erenová a JUDr. Jana Vlčková v právnej veci navrhovateľa: U. F., W.. XX.XX.XXXX, F. U. Bratislave, M. G.. Č.. XX, toho času: H. L. L. pre H., G.. L. Č.. XX, F., zast. advokátkou: Mgr. Tatiana Dlhošová, Nobelova ul. č. 34, Bratislava, proti odporkyni: E. F., nar. XX.XX.XXXX, bytom v F., M. G.. Č.. XX, o rozvod manželstva, na odvolanie odporkyne proti rozsudku Okresného súdu Bratislava V zo dňa 7. februára 2011, č. k. 11C 218/2008-73, takto rozhodol:
Súd prvého stupňa napadnutým rozsudkom rozviedol manželstvo účastníkov [...]

Filtering

From the od-justice v3.0 vertical export (per-token word/lemma/tag columns, <g/> glue markers, <p dup="0|1"> paragraph duplicate flags): text reconstructed per document with glue-aware spacing; Unicode/whitespace normalization; layout-line removal; PII pattern redaction; Slovak language gate (langid == sk on the first 1000 characters); exact + near deduplication.

Rejections (5,576,576 records, one final decision per record in artifacts/decisions.jsonl.gz):

{
  "duplicate_id": 1390844,
  "low_letter_ratio": 1705,
  "non_slovak_text": 30,
  "too_little_text": 4183164,
  "too_little_text_after_redaction": 833
}

Source-level paragraph dedup: paragraphs flagged dup="1" in the vertical (the upstream duplicate markers, consistent with tokcountdd) are dropped at extract; the per-record count is in artifacts/attribution.jsonl.gz. Documents whose paragraphs are all source-duplicated end up empty and are rejected as too_little_text - empty reconstruction correlates 1:1 with the source's own tokcountdd == 0 (verified on a 200k-record sample), which is why too_little_text and duplicate_id dominate the rejections (form rulings and re-published decision copies in the source).

PII redaction (redact_pii() in the builder): labelled patterns only - e-mail, labelled phone, labelled IP address ("IP adresa", "adresa IP"), labelled national identifier ("rodné číslo"), 26-digit account candidates. Pattern counts: 8,596 e-mail, 1,762 labelled phone, 165 labelled IPv4, 12 labelled IPv6, 279 national identifier, 10 account candidates. Residual pattern documents after redaction: 0 across all patterns. IP addresses are matched only when labelled - bare dotted quads are usually decimal document numbering ("bodu 2.1.1.1") in this genre and must not be redacted. Pattern redaction is residual scrubbing on top of the origin anonymization, not anonymization.

Layout filter (drop_layout_lines()): bullet-only lines, page fractions (x/y) and Strana n footers; decimal numbering is content and survives (889 line instances removed).

Boilerplate: lines binned to 2^26 64-bit hash bins with document frequency above 12,760 (1% of gated documents) are dropped (377,797 line instances removed).

Deduplication: exact dedup on 128-bit truncated sha256 of the final text; near dedup with Jaccard >= 0.90 on capped 25 x 5-word shingle sets, where candidate pairs come from a first-owner-per-shingle majority vote (>= 20 shared shingles) instead of the full inverted index used at legal_sk scale (6.5M documents); the representative of a duplicate group is the first occurrence in shard order and chained duplicates drop transitively. Records sharing a source id are a separate rejection class (duplicate_id): the first occurrence wins, later copies are rejected.

Known limitations

• Register: administrative/judicial legalese (rulings, orders, payment orders, criminal orders), not conversational Slovak.
• Published decision wording only; consolidated or corrected texts are not included.
• Letter-spaced print emphasis (e.g. r u š í) survives reconstruction as spaced letters.
• created is the decision date at the precision present in the source.
• Pattern redaction is not anonymization; unlabelled identifiers and personal disclosures may remain beyond the origin masking.
• Language gate operates on a leading-text slice of each document.
• Boilerplate detection is hash-binned (rare bin collisions may over-drop by a negligible amount).

See artifacts/ for attribution, per-record decisions, QA counts, samples and checksums. verify() passes: 961,401 rows, unique IDs, token counts recomputed and text_sha256 re-checked against the emitted text. Tests: 10 (pytest).