Beyond Grammatical Error Correction: Improving L1-influenced research writing in English using pre-trained encoder-decoder models

Gustavo Zomer, Ana Frankenberg-GarciaEMNLP 2021 · DOI

Introduction

Standard Grammatical Error Correction (GEC) tools often focus on surface-level errors like spelling or basic syntax. However, international researchers frequently face "L1 interference"—linguistic habits from their native language that result in grammatically correct but unnatural-sounding English. This research introduces a novel approach using pre-trained encoder-decoder models to improve the fluency, acceptability, and diversity of academic writing by specifically targeting these deep-seated transfer errors.

Methodology: Machine Translation as a Proxy

A significant hurdle in training models for academic writing improvement is the lack of large-scale, human-annotated datasets. The authors overcome this by using Machine Translation (MT) as a proxy for non-native errors.

  1. Data Synthesis: High-quality parallel corpora (EUR-Lex) were used. The Portuguese and Spanish portions were machine-translated back into English. The resulting "clunky" translations served as a simulation of L1-influenced English, preserving the syntax of the source language.
  2. Model Architecture: A T5 (Text-to-Text Transfer Transformer) model was fine-tuned to "translate" this literal MT-style English into polished, human-authored English.
  3. L1-Awareness: The model was conditioned with language-specific prompts (e.g., improve_english <L1>:) to learn patterns unique to different native languages.
  4. Token Preservation: To protect technical terminology and mathematical symbols, a pre-processing step wrapped out-of-vocabulary tokens in [KEEP_ID] tags to prevent accidental corruption.

Evaluation Framework

The models were evaluated on the BrACE (Brazilian Academic Corpus of English) and LACE (Latin-American Academic Corpus of English) datasets. Beyond standard grammaticality (GRM), the researchers employed metrics to capture stylistic quality:

  • Acceptability (ACP): Measured via the Syntactic Log-Odds Ratio (SLOR) using a GPT-2 model to assess how "natural" a sentence sounds compared to expert writing.
  • Syntactical Diversity (SxD) and Lexical Diversity (LxD): Metrics used to quantify the variety and sophistication of the language used.

Key Results

The proposed "L1-aware" models were compared against GECToR, a leading GEC system. While GECToR is highly conservative and often makes minimal changes, the T5-based model demonstrated a greater willingness to restructure sentences for better flow and professional tone.

Table 1: Qualitative comparison of sentence improvements and diversity scores between GECToR and the proposed model.

As shown in the quantitative results, the L1-aware-large model significantly outperformed GECToR in acceptability and diversity. On the BrACE corpus, it achieved an ACP score of 0.517, compared to GECToR's -0.009.

Table 2: Quantitative performance metrics on the BrACE corpus, highlighting the superior acceptability and diversity of L1-aware models.

Discussion and Limitations

The model's "adventurous" nature allows it to catch 34% of overused L1-influenced phrases, whereas GECToR only identifies 14%. However, this approach leads to higher rates of "over-correction," where the model replaces perfectly acceptable words with synonyms (e.g., swapping "furthermore" for "moreover"). Additionally, despite the preservation tags, the model occasionally struggles with highly specialized technical symbols and word order.

Conclusion

By leveraging machine translation as a proxy for L1 interference, this research demonstrates that encoder-decoder models can move beyond simple error correction toward true stylistic improvement in academic writing. Future work will focus on expanding L1 support and refining the model's precision to minimize unnecessary edits.

Summary updated

BibTeX
@inproceedings{zomer2021beyond,
  author = {Gustavo Zomer and Ana Frankenberg-Garcia},
  title = {{Beyond Grammatical Error Correction: Improving L1-influenced research writing in English using pre-trained encoder-decoder models}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.findings-emnlp.216},
}
Figure