General and Domain-adaptive Chinese Spelling Check with Error-consistent Pretraining

The lack of label data is one the significant bottlenecks for Chinese Spelling Check. Existing researches use automatic generation method by exploiting unlabeled to expand supervised corpus. However, there a big gap between real input scenario and automatically generated Thus, we develop competitive general speller ECSpell, which adopts Error-consistent masking strategy create pretraining. This error-consistency used specify error types sentences consistent with scene. experimental result indicates that our model outperforms previous state-of-the-art models on benchmark. Moreover, spellers often work within particular domain in life. Due many uncommon terms, experiments built domain-specific datasets show perform terribly. Inspired common practice methods, propose add an alterable user dictionary handle zero-shot domain-adaption problem. Specifically, attach User Dictionary guided inference module (UD) token classification-based speller. Our demonstrate ECSpell UD , namely, combined UD, surpasses all other baselines broadly, even approaching performance 1