gabelev/selex-rt-outputs

SeLex-RT Outputs: Multilingual Round-Trip Synthetic GEC Data

Dataset Summary

This dataset contains synthetic training data for low-resource Grammatical Error Correction (GEC), generated via a round-trip machine translation (RT) pipeline inspired by SeLex-RT from Low-Resource Grammatical Error Correction: Selective Data Augmentation with Round-Trip Machine Translation (Gomez & Rozovskaya, 2025). The pipeline targets lexical errors — the class of errors most systematically underrepresented in standard synthetic GEC corruption methods.

For each target language, clean sentences from a monolingual corpus are translated into a pivot language and back. Divergences between the original and back-translated text yield confusion sets of plausible lexical substitutions, which are applied to generate (corrupted, clean) sentence pairs for GEC pretraining.

This release covers three target languages (Russian, Ukrainian, Slovene) across three pivot distances (close, medium, distant), producing a 3×3 matrix of pivot-language combinations.

---

Languages and Pivot Matrix

Target Language	Close Pivot	Medium Pivot	Distant Pivot
Russian (ru)	Bulgarian (bg)	English (en)	Chinese (zh)
Ukrainian (uk)	Bulgarian (bg)	English (en)	Chinese (zh)
Slovene (sl)	Croatian (hr)	English (en)	Chinese (zh)

Close pivots are linguistically proximate Slavic languages selected to maximize lexical interference signal. Medium and distant pivots provide complementary confusion sets covering different regions of the substitution space.

---

Dataset Structure

selex-rt-outputs/
├── synthetic/           # (corrupted, clean) sentence pairs — primary training data
│   ├── ru_bg.jsonl
│   ├── ru_en.jsonl
│   ├── ru_zh.jsonl
│   ├── uk_bg.jsonl
│   ├── uk_en.jsonl
│   ├── uk_zh.jsonl
│   ├── sl_hr.jsonl
│   ├── sl_en.jsonl
│   └── sl_zh.jsonl
├── confusion/           # Token-level confusion sets per language/pivot
│   ├── ru_bg.json       # token → {substitute: frequency} dictionary
│   ├── ru_en.json
│   └── ...
├── translations/        # Intermediate MT outputs
│   ├── ru_bg_forward.json   # 5 forward hypotheses per sentence
│   ├── ru_bg_backward.json  # 3 back-translations per forward hypothesis
│   └── ...
└── eval/                # Coverage evaluation against gold learner errors
    ├── ru_coverage.json
    ├── uk_coverage.json
    └── sl_coverage.json

Synthetic data format (synthetic/*.jsonl)

Each line is a JSON object:

{
  "source": "Он сделал большой ошибку в своей работе.",
  "target": "Он допустил большую ошибку в своей работе.",
  "pivot": "bg",
  "lang": "ru"
}

Confusion set format (confusion/*.json)

{
  "сделал": {"допустил": 12, "совершил": 7, "произвёл": 3},
  "большой": {"крупный": 9, "значительный": 4}
}

---

Generation Pipeline

Step 1 — Forward translation. Each sentence from the monolingual source corpus is translated into the pivot language using OPUS-MT, generating 5 hypotheses per sentence via beam search.

Step 2 — Back-translation. Each forward hypothesis is translated back into the target language, producing 3 back-translations per hypothesis (15 RT paths per sentence total).

Step 3 — Alignment and confusion set construction. Original sentences are token-aligned with their back-translations. Divergences that are single-word, non-spelling, and non-morphological substitutions are retained as lexical confusion pairs. Frequency counts across all 15 paths determine substitution confidence.

Step 4 — Corpus corruption. Confusion sets are applied to a large monolingual corpus to generate (corrupted, clean) training pairs.

Source corpora

• Russian: Leipzig Corpora Collection (Russian News 2023 100K)
• Ukrainian: CC-100 Ukrainian subset
• Slovene: CC-100 Slovene subset

MT systems

NLLB-200-distilled-1.3B.

---

Coverage Evaluation

Coverage is measured as the percentage of gold lexical errors from the MultiGEC-2025 development sets that appear in the generated confusion sets. Gold errors are extracted from MultiGEC annotated (original, corrected) pairs after discarding multi-word substitutions, non-word tokens, spelling errors, and morphological errors.

Language	Gold Lexical Pairs	Close Pivot	Medium Pivot	Combined
Russian	237	bg: 41.4%	en: 43.9%	46.0%
Ukrainian	59	bg: TBD	en: TBD	TBD
Slovene	220	hr: 47.3%	en: TBD	TBD

Combined = union of close + medium pivot confusion sets. Distant pivot (zh) results pending.

---

Intended Use

Primary use: Synthetic pretraining data for multilingual GEC models, specifically to improve coverage of lexical error types prior to fine-tuning on gold annotated data such as MultiGEC-2025.

Recommended combination: This dataset is designed to complement tagged corruption corpora such as Stahlberg & Kumar (BEA 2024), which cover morphological, orthographic, and syntactic errors. The two sources together provide full ERRANT error type coverage for Russian.

Not recommended for: Direct evaluation or as a substitute for gold annotated learner corpora.

---

Related Work

---

Citation

If you use this dataset, please cite the original SeLex-RT paper:

@inproceedings{gomez-rozovskaya-2025-low-resource,
    title = "Low-Resource Grammatical Error Correction: Selective Data Augmentation with Round-Trip Machine Translation",
    author = "Gomez, Steven and Rozovskaya, Alla",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    year = "2025",
    publisher = "Association for Computational Linguistics",
}

And the MultiGEC dataset if you use the coverage evaluation:

@inproceedings{multigec2025,
    title = "{MultiGEC}: A Multilingual Grammatical Error Correction Dataset",
    author = "Masciolini, Arianna and others",
    booktitle = "Proceedings of the NLP4CALL Workshop",
    year = "2025",
}

---

License

Corruption edits and confusion sets are released under CC BY 4.0. Source sentence content derives from the respective monolingual corpora and is subject to their original licenses.

---

Dataset Card Authors

Gabe Levine and Aleksandra Kiszkiel