Xinda Wang, Zhengxu Hou, Yangshijie Zhang, Yanbingren, Jialin Liu, ChenZhuo Zhao, Zhibo Yang, Bin-Bin Yang, Feng Xiao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xinda Wang, Zhengxu Hou, Yangshijie Zhang, Yanbingren, Jialin Liu, ChenZhuo Zhao, Zhibo Yang, Bin-Bin Yang, Feng Xiao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In the realm of natural language processing, the Chinese Spelling Correction (CSC) task poses a significant challenge. Regrettably, the progress in enhancing its performance has been rather restricted, mainly due to the constrained incorporation of knowledge. In prior studies, confusion sets were introduced as supplementary knowledge sources. Nevertheless, these sets were characterized by their small scale and merely functioned as ancillary features. To address this, we propose a knowledge recall and selection network (ReSC). Initially, we deployed four distinct recall methods, successfully attaining an average recall rate exceeding 93%. Notably, for each character, we can recall approximately 150 related characters or words. Subsequently, we propose a Knowledge Selection Algorithm to choose the appropriate characters or words from numerous recall sets. The knowledge selection network is highly efficient, as the F1 score nearly reaches 100%. Extensive experiments have proven that ReSC can inject a substantial amount of entities with an even lower False Positive Rate. This novel network achieves better results across six datasets from ECSpell and SIGHAN.