alea-institute/alea-ocr-correct

ALEA ocr_correct: OCR and keying error detection and correction on clean legal and government text

Proofreading tasks built from clean, born-digital public-record text in the KL3M data collection: the U.S. Code, the Code of Federal Regulations, Federal Register documents, congressional bills, federal agency web pages, and SEC 10-K MD&A sections and contract exhibits. Realistic OCR / keying errors are inserted synthetically; the clean source is the gold. About 28% of rows are clean twins of the same passage whose correct answer is that there are no errors. No language model wrote or labelled any of the data.

Tasks

kind	rows	input	answer
fix_list	37,809	a 1-4 sentence passage with 0-3 errors	each error and its correction ("wrong -> right"), in order of appearance; or "No errors"
fix_full	21,380	one short sentence (10-30 words) with 0-2 errors	the corrected sentence; or "No errors"
detect	10,566	four consecutive sentences labelled A-D, at most one with an error	the letter, or "none"

Total: 69,755 documents, 7.2M tokens (alea-v2b-sw32k tokenizer), from 34,458 source documents. Errors per row: 0 (clean twin) 27.6%, 1 error 49.8%, 2 errors 19.3%, 3 errors 3.2% (fix_list only). Some rows ask for a JSON answer.

Corruptions

Each error is one of: rn<->m, cl<->d, li<->h, l->1 and o->0 inside a word, e<->c, a dropped letter, a doubled letter, a missing space (two words merged), a spurious space (a word split), or stray hyphenation ("regu-lation", "regu- lation"). Mix over all errors: doubled letter 14%, merge 13%, e->c 13%, split 8%, hyphen 8%, o->0 8%, d->cl 7%, dropped letter 6%, c->e 6%, h->li 6%, m->rn 5%, l->1 4%, li->h 1%, cl->d and rn->m under 0.2% each (few words contain them in a way that gives a unique fix).

Every error is checked to be unambiguous:

• only an all-lowercase word of 4+ letters that is common in the clean sample is corrupted; capitalised words, names, citations, numbers, section symbols, defined terms in capitals and sentence-initial words are never touched;
• an in-word corruption must produce a string never seen in the clean sample, and the only common word reachable from it by any single edit (insert, delete, substitute or swap a letter, digit-to-letter, rn/m, cl/d, li/h) must be the original;
• a merged pair must be an unseen string whose only split into two words is the original pair;
• a split or hyphenated word must leave at least one piece that is not a word.

Clean sources

The source passages must themselves be clean. A word list was built from the full clean sample (all eCFR, U.S. Code, Federal Register, bills and agency web page documents of the collection plus the EDGAR passages; 285k word types). A passage is dropped when any word in it occurs in fewer than three documents of the sample, or when it shows a spacing or punctuation quirk a proofreader could flag (space before punctuation, glued sentences, source line-break hyphens, doubled words such as "that that"). Court opinions and other OCR-derived sources are not used. Text is otherwise verbatim (line-wrap whitespace inside a passage is shown as one space).

Surface-cue controls

Clean twins share the passage with a corrupted version, so topic and source carry no signal: for every kind an instruction-only, a length-only and a source-only classifier score exactly the majority prior (detect: "none" 28%, letters A-D 18% each; fix_list / fix_full: 27-28% clean). The answer never repeats a long passage: only fix_full returns the whole (short) sentence.

Composition

U.S. Code 29% of source documents, Federal Register 23%, agency web pages 21%, bills 18%, EX-10 contracts 6%, eCFR 2%, 10-K MD&A 1%. No title, agency, host or filer exceeds 4.6% of source documents. Documents used as evaluation targets elsewhere were excluded, and rows sharing long n-grams with evaluation and held-out sets were removed. Presentation: about 40% plain instruction-then-text, 30% question/answer framing, 30% chat turns; 5% of rows carry a worked example from another document with a different answer.

Review

The builder read 35 released rows against the clean source (33 correct, 2 with harmless source oddities such as footnote numbers or a sentence fragment cut at a page break), after two earlier review rounds that added the source-quirk filters. An independent blind review of 60 rows found no wrong rows; 9 had minor issues (a worked example in a different answer format from the task, leftover page or footnote markers, a sentence split at a page break).

Limitations

• Passages are short (median about 90 tokens per document).
• The corruption mix is synthetic; real OCR noise is burstier and includes errors in capitalised words and numbers, which are excluded here to keep every correction unique.
• A clean passage can still contain a rare source oddity (footnote numbers, a fragment at a page break); such text is not counted as an error.