GEC-DCL: Grammatical Error Correction Model with Dynamic Context Learning for Paragraphs and Scholarly Papers

Abstract

Over the past decade, there has been noteworthy progress in the field of Automatic Grammatical Error Correction (GEC). Despite this growth, current GEC models possess limitations as they primarily concentrate on single sentences, neglecting the significance of contextual understanding in error correction. While a few models have begun to factor in context alongside target sentences, they frequently depend on inflexible boundaries, which leads to the omission of vital information necessary for rectifying certain errors. To address this issue, we introduce the Dynamic Context Learner (DCL) model, which identifies optimal breakpoints within paragraphs or documents to retain maximum context. Our method surpasses those employing fixed sequence lengths or assuming a limited number of preceding sentences as context. Through extensive evaluation on the CoNLL-2014, BEA-Dev, and FCE-Test datasets, we substantiate the efficacy of our approach, achieving substantial F$_{0.5}$ score enhancements: 77% increase, 19.61% boost, and 10.49% rise respectively, compared to state-of-the-art models. Furthermore, we contrast our model’s performance with LLaMA’s GEC capabilities. We extend our investigation to scientific writing encompassing various context lengths and validate our technique on the GEC S2ORC dataset, yielding cutting-edge results in scholarly publications.

A Appendix

Table 9 deduces that corrections dependent on the more extended context, like tense and verb choice, have seen significant improvement with our DCL model and show a mean of 24$\%$ of lift across correction made in missing (M), replacement (R) and unnecessary (U) error in FCE-Test. The formula for calculating the $Diff F_{0.5}$ can be taken from Eq. 1. In the Eq. 1, A stands for the $F_{0.5}$ score of the baseline model, and B stands for $F_{0.5}$ GEC-DCL, the proposed model.

$$\begin{aligned} Diff F_{0.5} = 100 * \frac{2 * |A - B|}{(A + B)} \end{aligned}$$

(1)

Table 8. Types of error and their distribution used in the errorify script.

Table 9. Error type-specific performance comparison of GECToR and DCL-GEC on FCE-Test in F$_{0.5}$, difference between our proposed model and GECToR in last column. 24$\%$ increase shown by our DCL-GEC model.