📄 Grammar Error Correction in Morphologically Rich Languages: The Case of Russian

Summary

Research on grammar error correction (GEC) has focused almost entirely on English, where systems rely on tens of millions of words of annotated learner text. This paper tests how well established English methods transfer to Russian, a highly inflected Slavic language with no comparable annotation. The authors corrected and error-tagged more than 200K words written by non-native authors, both foreign and heritage learners. They also designed an error taxonomy and released the data as a benchmark. On this corpus they compared two families of systems. The first is machine translation (MT), which depends on large parallel learner data. The second is error-specific classifiers trained with minimal supervision, which pair abundant native Russian text with a small annotated learner sample. All scores fall well below English levels, but the classifiers clearly outperform MT trained on the same data. The work matters for two reasons. It supplies a rare non-English GEC resource. It also shows that when annotation is scarce and words have many forms, combining cheap native data with modest annotation is the more practical approach.

Key points

• The new corpus covers more than 200K words of learner Russian, with corrections and error tags. It is distributed via the RULEC-GEC repository on GitHub and includes analysis of error rates for foreign and heritage learners.
• Under the standard F0.5 metric, the minimal-supervision classifier system reaches 21.0, while the MT system trained on identical data reaches only 10.6.
• Russian nouns and related words can take up to 24 inflected forms. This creates lexical sparsity, which large native corpora help offset.
• Classifier precision is fairly good, but recall is low. Noun case and preposition errors reach roughly 25–36% recall, agreement errors about 16%, and verb aspect errors 9.1%.
• Manual re-checking of false positives raised precision by 4–16 points for most error types. For prepositions, about 48% of flagged "false" corrections turned out to be acceptable alternatives.