Bi-DCSpell: A Bi-directional Detector-Corrector Interactive Framework for Chinese Spelling Check

TL;DR

Bi-DCSpell presents a novel bi-directional detector-corrector framework for Chinese Spelling Check, enabling mutual enhancement of detection and correction through a pair of dedicated encoders and a bi-directional interactive learning module. The approach jointly optimizes detection and correction with a multi-objective loss and leverages bi-directional cross-attention, learnable control gates, and a merged feed-forward network to refine representations. Empirical results on SIGHAN13-15 show state-of-the-art correction F1 and robust detection across backbones, with ablations confirming the value of bi-directional interactions and moderate interaction levels. The work highlights the practical impact of interactive learning for CSC and suggests broader applicability to related linguistic error-correction tasks, while acknowledging limitations in output consistency, language scope, and training efficiency.

Abstract

Chinese Spelling Check (CSC) aims to detect and correct potentially misspelled characters in Chinese sentences. Naturally, it involves the detection and correction subtasks, which interact with each other dynamically. Such interactions are bi-directional, i.e., the detection result would help reduce the risk of over-correction and under-correction while the knowledge learnt from correction would help prevent false detection. Current CSC approaches are of two types: correction-only or single-directional detection-to-correction interactive frameworks. Nonetheless, they overlook the bi-directional interactions between detection and correction. This paper aims to fill the gap by proposing a Bi-directional Detector-Corrector framework for CSC (Bi-DCSpell). Notably, Bi-DCSpell contains separate detection and correction encoders, followed by a novel interactive learning module facilitating bi-directional feature interactions between detection and correction to improve each other's representation learning. Extensive experimental results demonstrate a robust correction performance of Bi-DCSpell on widely used benchmarking datasets while possessing a satisfactory detection ability.

Figures (6)

• Figure 1: Comparison of three types of interactions: 'correction-only', 'detection-to-correction' and 'bi-directional interactions'. The wrongly/ground-truth corrected characters in the candidates are in red/blue.
• Figure 2: Overview of our Bi-DCSpell framework. $\oplus$ is the matrix addition operation, and $\sigma$ denotes learnable control gate.
• Figure 3: The average value of $\alpha$ and $\beta$ with different bi-directional interaction layers.
• Figure 4: The F1 scores for detection and correction vary as the number of bi-directional interaction layers increases.
• Figure 5: The F1 scores with respect to $\lambda$.