YasalmaGEC: Tatar grammatical error correction with error annotations

649 manually corrected Tatar documents with 3,620 annotated corrections in 26 error types. The layout follows UA-GEC: document-level source and target, inline annotations, and error types grouped into coarse categories. On top of that, every rewrite is aligned to the word level so you can train token classification models, and an M2 block is included for ERRANT / M2Scorer evaluation.

Columns (one row per document)

column	type	description
id	int	document id
source / target	str	original text / corrected text
annotated	str	UA-GEC-style inline markup {error=>correction:::error_type=Type}. A moved fragment appears twice: :::move=<id> where it is inserted and :::move_from=<id> where it is removed.
corrections	list[struct]	every original annotation field (id, operation, start, end, original, replacement, error_type_id, error_type, linked_removals) plus error_type_raw, error_category, target_start/target_end (the replacement's span in target), and source_token_start/source_token_end
source_tokens / target_tokens	list[str]	word and punctuation tokens
source_token_offsets	list[[int,int]]	character offsets of each source token
token_labels	list[str]	per-word error type (O = correct)
token_categories	list[str]	per-word coarse category (Spelling, Punctuation, Grammar, Fluency, Other, or O)
token_ops	list[str]	per-word operation: KEEP, REPLACE, DELETE, APPEND (insert after this word), MOVE_FROM
token_correction_ids	list[str]	per word, the id of the correction that applies to it (look up the rewrite in corrections)
start_insertion_id	str	id of a correction that inserts text before the first token, if any
m2	str	M2 block (S … / A i j|||type|||correction|||REQUIRED|||-NONE-|||0)
n_corrections, error_types, error_categories		per-document summaries
reconstructs_target	bool	applying corrections to source gives exactly target (647/649)
m2_reconstructs_target	bool	applying the M2 edits to source_tokens gives target_tokens (648/649)

The repo also contains data/yasalma_gec.m2, the whole corpus in M2 format.

Example

source:    Хаерле көн. Бу группада Казанның Майский теплицасында эшләүчеләр булса аларның киңәшләре кирәк иде.Эш урыны эзли идем.
target:    Хәерле көн. Бу төркемдә Казанның "Майский" теплицасында эшләүчеләр булса, аларның киңәшләре кирәк иде.Эш урыны эзли идем.
annotated: {Хаерле=>Хәерле:::error_type=Spelling} көн. Бу {группада=>төркемдә:::error_type=CodeSwitch} Казанның {Майский=>"Майский":::error_type=Punctuation} теплицасында эшләүчеләр булса{=>,:::error_type=Punctuation} аларның киңәшләре кирәк иде.Эш урыны эзли идем.

S Хаерле көн . Бу группада Казанның Майский теплицасында эшләүчеләр булса аларның киңәшләре кирәк иде . Эш урыны эзли идем .
A 0 1|||Spelling|||Хәерле|||REQUIRED|||-NONE-|||0
A 4 5|||CodeSwitch|||төркемдә|||REQUIRED|||-NONE-|||0
A 6 7|||Punctuation|||" Майский "|||REQUIRED|||-NONE-|||0
A 10 10|||Punctuation|||,|||REQUIRED|||-NONE-|||0

token_labels: [('Хаерле', 'Spelling'), ('көн', 'O'), ('.', 'O'), ('Бу', 'O'), ('группада', 'CodeSwitch'), ('Казанның', 'O'), ('Майский', 'Punctuation'), ('теплицасында', 'O'), ('эшләүчеләр', 'O'), ('булса', 'Punctuation'), ('аларның', 'O')]

Statistics

Operations: replace 3,024, insert 393, move 128, delete 75. Categories: Spelling 1,846, Punctuation 893, Grammar 439, Fluency 355, Other 87. Tokens: 21,352 source tokens, 6,652 of them labelled as erroneous.

error_type	category	count
Spelling	Spelling	1,140
Punctuation	Punctuation	796
Merge	Spelling	414
CodeSwitch	Fluency	263
WordOrder	Grammar	254
Split	Spelling	110
Particle	Grammar	107
SentenceBoundaries	Punctuation	97
CapitalLowerLetter	Spelling	94
Undefined	Other	87
Hyphen	Spelling	70
Case	Grammar	36
Style	Fluency	35
Dialect	Fluency	21
WordChoice	Fluency	20
Script	Spelling	18
MissingWord	Grammar	17
Possessive	Grammar	15
Calque	Fluency	8
VerbTense	Grammar	6
Pleonasm	Fluency	5
SubjectVerbAgreement	Grammar	2
Paronym	Fluency	2
MissingSubjVerb	Grammar	1
Collocation	Fluency	1
VerbVoice	Grammar	1

Notes

• A move correction inserts replacement at [start, end) and removes each linked_removals span. Both places are labelled: the removed words get MOVE_FROM, and the word before the insertion point gets APPEND.
• Normalization: the Undefiined type is renamed to Undefined. Possessive appears under two ids (6 and 24), so the original id is kept in error_type_id and the raw name in error_type_raw.
• In 2 documents (511 and 544), the target contains small edits that were not annotated. See reconstructs_target.
• When one M2 edit covers several overlapping corrections, its type is the joined list of their types, e.g. Merge+Spelling.

from datasets import load_dataset
ds = load_dataset("gaydmi/yasalmaGEC", split="train")