DCL-GEC: Dynamic context Learner Grammatical Error correction Model for Paragraphs

Abstract Significant progress has been made in automatic Grammatical Error Correction (GEC) over the past decade. However, existing GEC models have limitations due to their focus on individual sentences, disregarding the importance of contextual information for error correction. To overcome this challenge, some models have started incorporating context with target sentences. Nevertheless, current grammatical error correction algorithms often rely on fixed context boundaries, resulting in the exclusion of crucial information necessary for resolving specific errors. To address this issue, we propose the Dynamic Context Learner (DCL) model, which learns optimal splits within paragraphs or documents to preserve maximum contextual information. Our approach outperforms methods that impose fixed sequence length limits or assume the past few sentences as context. We provide evidence for our approach’s superiority by achieving substantial improvements in F$_{0.5}$ scores compared to state-of-the-art models on CoNLL-2014, BEA-Dev, and FCE-Test datasets, with percentage increases of 77\%, 19.61\%, and 10.49\% respectively. Additionally, we extend our research to scientific writings with varying context lengths and validate our technique by introducing the GEC S2ORC dataset and presenting state-of-the-art findings in scholarly papers.

Status - Pending