DiktantGEC: Tatar dictation errors for grammatical error correction

Tatar dictations: 570 transcriptions written by people from 6 reference texts. The writer's transcription (source, with errors) is paired with the reference text (target). The original character-level alignment is kept, and from it the dataset derives word-level edits, per-word error labels, UA-GEC-style annotations and M2 edits (the same layout as gaydmi/yasalmaGEC).

Configs

config	rows	description
windows (default)	1,416	GEC examples: each run of sentences with errors, plus one clean context sentence on each side. A window always has at least 3 sentences; if a context sentence also contains an error, the window grows. Deduplicated on (source, target): 1,877 windows reduce to 1,416, and doc_ids / n_writers keep the provenance.
sentences	11,317	every reference sentence of every transcription, with or without errors (6,532 have errors). Useful for error detection, including correct negatives.
documents	570	one row per transcription, with all original fields (alignment, substitution_errors, deletion_errors, insertion_errors, total_number_errors, correct_words, correct_words_percentage) plus derived statistics and flags

Window sizes: 1,041 of 1,416 windows have at most 6 sentences. The largest covers a whole text (27 sentences) because some writers made errors in every sentence. Filter on n_sentences if you need short inputs.

Columns (windows; sentences has the same annotation columns)

column	description
id	diktant_<doc_id>_<first sentence>_<last sentence>
doc_id, text_id	transcription id and reference-text id (0–5)
sentence_start, sentence_end, n_sentences, error_sentence_idx	which reference sentences the window covers, and which of them contain errors
source / target	writer's text / reference text
annotated	inline markup {error=>correction:::error_type=Type}
corrections	list of word-level edits: start, end, original (in source), target_start, target_end, replacement, operation, error_type, error_types, error_category, char_edits (the underlying M/S/D/I character ops), source_token_start/end
source_tokens, target_tokens, source_token_offsets	word and punctuation tokens
token_labels, token_categories, token_ops, token_correction_ids	per-word error type / category / operation (KEEP, REPLACE, DELETE, APPEND) / correction id
m2	M2 block for ERRANT / M2Scorer. The whole set is also in data/windows.m2.
n_corrections, n_char_errors, error_types, error_categories	summaries
has_omitted_sentence	the writer skipped more than half of some sentence in the window (13 windows)
doc_ids, n_writers	every transcription that produced exactly this pair
m2_reconstructs_target	applying the M2 edits to source_tokens gives target_tokens (1416/1416)

Error types

Error types are assigned automatically from the edit and its character operations. A single edit can have several types, e.g. Spelling+TatarLetter. Categories: Punctuation 6,129, Spelling 5,617, Omission 162, Insertion 136, Lexical 22.

type	count	meaning
Punctuation	5,900	missing, extra or wrong punctuation
Spelling	4,113	misspelled word
TatarLetter	1,902	Tatar-specific letter confused with its Russian counterpart (ә/а, ө/о, ү/у, җ/ж, ң/н, һ/х, ә/э)
Capitalization	1,134	upper/lower case
Dash	891	hyphen vs dash (- vs –)
Spacing	788	word split or merged (Казан су → Казансу)
MissingWord	162	word(s) not written
ExtraWord	136	word(s) not in the dictation
WordChoice	67	a different word

Example (windows)

source:    Соңгы елларда дөньядагы икътисади һәм сәяси халәт давыл уйнап торган диңгезгә охшап калды. Ләкин диңгездә гадәттә адашканны коткаручы утыраулар очрый торган була. Арып-талып, хәлдән таеп, шул утрауга барып сыенган кеше, нихаять, тын калып уйланырга, хәл җыярга, тормыш мәгънәсе, халыклар һәм илләр язмышы турында иркенләп фикер тупларга мөмкинлек һәм вакыт таба.
target:    Соңгы елларда дөньядагы икътисади һәм сәяси халәт давыл уйнап торган диңгезгә охшап калды. Ләкин диңгездә, гадәттә, адашканны коткаручы утраулар очрый торган була. Арып-талып, хәлдән таеп, шул утрауга барып сыенган кеше, ниһаять, тын калып уйланырга, хәл җыярга, тормыш мәгънәсе, халыклар һәм илләр язмышы турында иркенләп фикер тупларга мөмкинлек һәм вакыт таба.
annotated: Соңгы елларда дөньядагы икътисади һәм сәяси халәт давыл уйнап торган диңгезгә охшап калды. Ләкин диңгездә{=>,:::error_type=Punctuation} гадәттә{=>,:::error_type=Punctuation} адашканны коткаручы {утыраулар=>утраулар:::error_type=Spelling} очрый торган була. Арып-талып, хәлдән таеп, шул утрауга барып сыенган кеше, {нихаять=>ниһаять:::error_type=Spelling+TatarLetter}, тын калып уйланырга, хәл җыярга, тормыш мәгънәсе, халыклар һәм илләр язмышы турында иркенләп фикер тупларга мөмкинлек һәм вакыт таба.

S Соңгы елларда дөньядагы икътисади һәм сәяси халәт давыл уйнап торган диңгезгә охшап калды . Ләкин диңгездә гадәттә адашканны коткаручы утыраулар очрый торган була . Арып - талып , хәлдән таеп , шул утрауга барып сыенган кеше , нихаять , тын калып уйланырга , хәл җыярга , тормыш мәгънәсе , халыклар һәм илләр язмышы турында иркенләп фикер тупларга мөмкинлек һәм вакыт таба .
A 16 16|||Punctuation|||,|||REQUIRED|||-NONE-|||0
A 17 17|||Punctuation|||,|||REQUIRED|||-NONE-|||0
A 19 20|||Spelling|||утраулар|||REQUIRED|||-NONE-|||0
A 37 38|||Spelling+TatarLetter|||ниһаять|||REQUIRED|||-NONE-|||0

Processing notes

• The original alignment ops are: M match, S substitution (the stored value is the reference character), D a reference character missing from the transcription, I an extra character typed by the writer. Error indices are positions in the alignment. They were checked against the alignment for every document.
• Reference texts are split into sentences, and each sentence is mapped through the alignment to the writer's text.
• Junk transcriptions (character error rate above 0.5, empty, or a 1-character reference) are kept in documents with is_junk = True and excluded from sentences and windows (7 documents).
• Privacy: one transcription (doc 319) began with the writer's name, e-mail and phone number. That header was removed, its alignment ops were adjusted, and the error fields were recomputed (redacted = True). correct_words / correct_words_percentage of that document are the original values.
• Some transcriptions start with the dictation's title or author, or continue past the end of the reference text. These appear as ExtraWord edits.
• The reference texts themselves contain a few typos (e.g. тәртипсз), and target reproduces them as given.
• There is no train/test split. All transcriptions come from only 6 texts, so split by text_id for a fair evaluation.

from datasets import load_dataset
windows = load_dataset("gaydmi/diktantGEC", "windows", split="train")
sentences = load_dataset("gaydmi/diktantGEC", "sentences", split="train")
documents = load_dataset("gaydmi/diktantGEC", "documents", split="train")