Revisiting the Knowledge Recall and Selection in Chinese Spelling Correction

Anonymous

Abstract: Chinese Spell Check (CSC) task is a challenging natural language processing task. We are currently facing a significant challenge as the improvement in performance is quite limited, it is primarily because the infusion of knowledge is limited, and the injection of knowledge occurs without explicit selection. Previous work involved confusion sets, but the size was small and was only used as additional feature input. To more effectively address the issue of knowledge infusion, we propose a knowledge recall and selection network (ReSC). First through four kinds of recall to achieve an average recall rate above 93%, with individual character recall of around 150 related characters/words. Subsequently, we proposed a Knowledge Selection Algorithm, choosing the appropriate characters or words from numerous recall sets. The knowledge selection network is highly efficient, as the classification accuracy has nearly reached 100%. Extensive experiments have proven that our method achieves SOTA results in six datasets.

Paper Type: long

Research Area: NLP Applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: Chinese