MultiGEC

A research project on multilingual grammatical error correction for three low-resource Slavic languages: Russian, Ukrainian, Slovene. Built on the MultiGEC-2025 shared task gold corpora. Done with a research partner.

Two contributions.

SeLex-RT. A round-trip-translation synthetic-data pipeline with a pivot-language ablation across close, medium, and distant pivots. NLLB-200 for translation, SimAlign over multilingual BERT for alignment, 15% corruption rate. Output: 2.8M synthetic (corrupted, clean) pairs across three languages and three pivot distances, published as gabelev/selex-rt-outputs on the HuggingFace Hub.

Auxiliary detection head. An 11-class ERRANT-style token-level classifier bolted onto the mBART-50 encoder during gold fine-tuning. Jointly trained with the seq2seq correction objective. Discarded at inference. Costs nothing at serving time and improves the correction model for having had to also localize errors during training.

Hardware: single A40 (48 GB), mBART-50 at 610M parameters, beam search at inference, seed 42 throughout. Scored with the official MultiGEC-2025 GLEU + ERRANT pipeline. Code, configs, result tables, and the long-form research log at gabelev/MultiGEC-public.

Companion post: Round-trip errors, a throwaway head, three Slavic languages.