📄 The construction of English grammar error correction models based on Seq2Seq model and feedback filtering mechanisms

Summary

The paper addresses English grammatical error correction (GEC). Its starting point is that current systems, including the CAMB model, struggle with long sentences and long-distance dependencies. The authors build a Seq2Seq corrector from a recurrent encoder–decoder with attention. They add a feedback filtering mechanism based on N-grams, which scores incoming user corrections and keeps only plausible ones, so the model can adapt over time without absorbing noisy feedback. They motivate the design with second-language-acquisition theory: learner errors are treated as systematic traces of interlanguage rather than random noise, which makes targeted correction educationally meaningful. Training and evaluation used a merged FCE and Lang-8 Corpus dataset. Against CAMB, the model reports 5.69% higher accuracy and 1.34% higher F0.5. Gains are clearest on context-dependent errors such as articles, prepositions and attributives. Complex corrections take about 32–55 ms each. The work suggests that a fairly light architecture with quality-gated feedback can stay accurate and fast enough for teaching and writing tools.

Key points

• Architecture: an RNN-based Seq2Seq encoder–decoder with attention, plus an N-gram-based filter that judges user feedback before the model uses it to adapt.
• Data: the merged FCE + Lang-8 Corpus. Reported accuracy is 35.44%, recall 40.7% and F0.5 32.6%.
• Compared with CAMB, accuracy is +5.69% and F0.5 is +1.34%. The largest gains are on articles, prepositions and attributive errors.
• Average processing time on complex correction tasks is 32–55 ms. A stress test is reported to confirm robustness.
• The datasets are available only on request. The authors name multilingual GEC and integration with Transformer-based writing tools as future directions.

Interesting ideas

• I1. Gate feedback before learning from it: incoming corrections are scored with a cheap N-gram plausibility check, and only those that pass are allowed to update the model, which protects adaptive systems from poisoned or careless feedback.
• I2. Treat errors as systematic signal, not noise: learner mistakes are viewed as patterned outputs of a developing interlanguage, so they can be modelled and explained rather than just suppressed.
• I3. Precision-weighted metric for corrections: choosing F0.5 reflects the view that a wrong "fix" costs more than a missed one.
• I4. Error-type breakdown in evaluation: reporting results separately for articles, prepositions and attributives shows where the model actually improves, rather than hiding it in one aggregate number.
• I5. Latency as a first-class result: per-task processing time and a stress test are reported alongside quality, since interactive tools must respond quickly.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): put a similar cheap plausibility gate in front of LLM-teacher outputs and self-induced repair rules, so only candidates that pass enter silver data or the rule set.
• I2 (tree): collect recurring non-native patterns (missing articles, wrong prepositions) as named, explainable repair rules applied before or during parsing, instead of letting the parser fail silently.
• I3 (learn): score candidate repair rules and tree-rewrites with a precision-weighted measure, so rules that occasionally damage correct input are rejected even if they have high coverage.
• I4 (answer): break test and QA failures down by the linguistic phenomenon involved, which points directly to the pipeline stage that needs work.
• I5 (tokens): track per-sentence time for the tokenizer, tagger and lemmatizer under load, so speed regressions become visible as soon as quality changes land.