📄 Hi-GEC: Hindi Grammar Error Correction in Low Resource Scenario

Summary

Research on grammatical error correction (GEC) has concentrated on English and other well-resourced languages. Hindi, with its rich morphology and complex syntax, has had neither a manually annotated benchmark nor a systematic comparison of training-data strategies. The authors address both gaps.

First, they mine the Hindi Wikipedia revision history with a modified Wikiedits tool, filter the edits heavily, label them with the ERRANT toolkit, and rebalance them into HiWikiEdits, a human-edited corpus covering 17 error categories. Second, they compare three sources of synthetic training pairs:

• rule-based corruption of clean text (Direct-Noise);
• Round Trip Translation (RTT) through several pivot languages using IndicTrans2;
• a neural error generation model (EGM) trained to imitate the seed corpus.

Pivoting through English, a distant but high-resource language, produced better training data than closely related Indic languages. Model-generated errors performed on par with the other methods. The best system was pretrained on combined synthetic data and then fine-tuned on HiWikiEdits, reaching 73.86 GLEU. The work gives Hindi its first GEC benchmark and reusable resources for other low-resource languages.

Key points

• HiWikiEdits contains 8,137 sentence pairs (5,696 train / 976 validation / 1,465 test) from a July 2024 Hindi Wikipedia dump. Edits were kept only if sentences had 10–30 words, the pair differed by at most 3 words, and the Levenshtein ratio was below 0.3. A NOOP class holds error-free sentences.
• Direct-Noise provides about 300,000 pairs. Errors are injected at the word and character level following Grundkiewicz et al. (2019), using ASpell substitutions plus Hindi-specific matra (vowel-sign) replacements.
• RTT pivots were English, Assamese, Marathi and Tamil. English gave a 6.25% GLEU improvement over Assamese and Marathi. BLEU-filtered RTT data (scores between 30 and 95) outperformed unfiltered data, suggesting data quality matters more than quantity.
• EGM-generated data (about 100,000 pairs) improved GLEU by 17.54. This was not significantly different from the Direct-Noise-based model (M1).
• Release: the best model (M3) scores 73.86 GLEU and handles adjective and pronoun inflection errors well. The dataset and scripts are on GitHub (ujjwalsharmaIITB/Hi-GEC), and the synthetic data is promised under CC-BY-SA 4.0.