Datasetpublic

martinsr/gec-targeted-corrections-esl

GEC Targeted Corrections — ESL An LLM-generated grammatical error correction (GEC) dataset targeting the specific error patterns that ESL learners most commonly produce. Each example is a (src, tgt) pair where src contains a realistic grammatical error and tgt is the minimally corrected version: only what is necessary is changed. Dataset Summary Split Examples train 2,037 Schema { "src": "She gave me some advices about the… See the full description on the dataset page: https://huggingface.co/datasets/martinsr/gec-targeted-corrections-esl.

sourceHugging Facemitupdated 3mo agoView on Hugging Face

0likes36downloads

Dataset Card

GEC Targeted Corrections — ESL

An LLM-generated grammatical error correction (GEC) dataset targeting the specific error patterns that ESL learners most commonly produce. Each example is a (src, tgt) pair where src contains a realistic grammatical error and tgt is the minimally corrected version: only what is necessary is changed.

Dataset Summary

Split	Examples
train	2,037

Schema

json

{
  "src": "She gave me some advices about the interview.",
  "tgt": "She gave me some advice about the interview.",
  "pattern_id": "mass_noun_homeworks",
  "category": "Noun Number"
}

Field	Description
`src`	Sentence containing a grammatical error
`tgt`	Minimally corrected version of `src`
`pattern_id`	Identifier for the specific error pattern
`category`	Broader error category

Error Patterns

The dataset covers 20 targeted error patterns across 15 categories:

Category	Pattern ID	Description	Examples
Noun Number	`mass_noun_homeworks`	Uncountable nouns used as countable (homeworks, advices, informations, furnitures, equipments)	101
Article	`article_before_uncountable`	Indefinite article before uncountable nouns (a useful advice, a good information)	102
Double Negative	`double_negative_both`	Both negative pronouns need fixing (nothing→anything AND nobody→anybody)	102
Word Form	`adverb_fastly`	Non-existent adverb fastly instead of fast or quickly	102
Tense	`impossible_tense`	Past tense with future time expressions (she went tomorrow)	101
Tense	`complex_tense_reconstruction`	Past perfect continuous required (since many months → for many months + had been doing)	102
Collocation	`collocation_do_make`	do/make confusion (make a research → do research, do a mistake → make a mistake)	102
Collocation	`collocation_redundancy`	Redundant expressions (return back, repeat again, end result, past history)	102
Gerund/Infinitive	`gerund_vs_infinitive`	Verbs requiring gerund used with infinitive (enjoy to swim → enjoy swimming)	101
Gerund/Infinitive	`infinitive_vs_gerund`	Verbs requiring infinitive used with gerund (want doing → want to do)	102
Conditional	`conditional_type2`	Type 2 conditional: If I would have → If I had	102
Conditional	`conditional_type3`	Type 3 conditional: If I would have known → If I had known	102
Passive Voice	`passive_irregular`	Irregular past participles in passive (was wrote → was written, was build → was built)	102
Pronoun	`reflexive_pronoun`	Missing or wrong reflexive pronoun (he hurt him → himself)	102
Pronoun	`pronoun_case`	Wrong pronoun case (between you and I → me, her and I went → she and I)	102
Subject	`missing_it_subject`	Missing expletive it subject (Is very important → It is very important)	102
Comparative	`double_comparative`	Double comparative/superlative (more better → better, most fastest → fastest)	102
Preposition	`since_vs_for`	Since used with duration instead of for (since three years → for three years)	102
Vocabulary	`bring_vs_take`	Bring/take confusion based on direction of movement	102
Overcorrection Prevention	`no_change_needed`	Grammatically correct sentences that should not be changed	102

Design Principles

—Minimal correction: tgt changes only what is grammatically necessary. No paraphrasing or stylistic edits.
—Overcorrection prevention: The no_change_needed pattern includes correct sentences to discourage the model from changing things that are already right.
—Diversity: Sentences span a wide range of topics (work, travel, health, family, sport, technology) and sentence structures to prevent overfitting to surface patterns.
—ESL-realistic: All errors reflect patterns documented in ESL learner corpora rather than synthetic noise.

Usage

python

from datasets import load_dataset

ds = load_dataset("martinsr/gec-targeted-corrections-esl")

# Filter to a specific pattern
conditional = ds["train"].filter(lambda x: x["category"] == "Conditional")

# Use src/tgt for seq2seq training
for example in ds["train"]:
    print(example["src"], "→", example["tgt"])

Intended Use

—Fine-tuning sequence-to-sequence models (T5, BART, mT5) for grammatical error correction
—Targeted evaluation of GEC models on specific error types
—Data augmentation alongside larger corpora such as FCE, W&I+LOCNESS, or NUCLE

Limitations

—English only
—Covers 20 specific patterns; not a general-purpose GEC corpus
—Sentences are authored rather than collected from real learner writing

License

MIT