# uChecker: Masked Pretrained Language Models as Unsupervised Chinese Spelling Checkers

**Authors:** Piji Li  
**arXiv:** [2209.07068](https://arxiv.org/abs/2209.07068)  
**Date:** 2022-09-15  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2209.07068v1/

## Key points

- Detection and correction of spelling errors in text are crucial tasks in the field of
- Manually annotating high-quality datasets for training models is expensive and time-consuming
- Limited training datasets like SIGHAN15 with only 2339 samples pose challenges for
- Introduction of an for CSC and a novel framework called
- Utilization of masked pretrained language models like BERT for robust language diagnosis capabilities
- Proposal of Confusionset-guided masking strategy to fine-tune the masked language model for improved performance in unsupervised spelling error detection and correction
- Experimental results on standard datasets show effectiveness of uChecker in character-level and sentence-level metrics including Accuracy, Precision, Recall, and F1-Measure
- uChecker demonstrates promising results in overcoming data sparsity limitations and addressing overfitting issues faced by supervised-learning based models

## Summary

In the field of , the detection and correction of spelling errors in text are crucial tasks. However, manually annotating a high-quality dataset for training models is expensive and time-consuming. This leads to limited training datasets, such as SIGHAN15 which only contains 2339 samples for training. This poses challenges for , which can suffer from data sparsity and overfitting issues in the era of large language models. To address these challenges, this paper explores an  for CSC and introduces a novel framework called . The framework utilizes masked pretrained language models like BERT as its backbone due to their robust language diagnosis capabilities. By using various masking operations, the authors propose a Confusionset-guided masking strategy to fine-tune the masked language model for improved performance in unsupervised spelling error detection and correction. Experimental results on standard datasets demonstrate the effectiveness of uChecker in terms of character-level and sentence-level metrics including Accuracy, Precision, Recall, and F1-Measure for spelling error detection and correction tasks. The proposed approach showcases promising results in overcoming data sparsity limitations and addressing overfitting issues commonly faced by supervised-learning based models. Overall, uChecker presents a valuable contribution to advancing unsupervised methods for Chinese spelling check applications.