Decopy: detect and correct with pinyin for Chinese spelling correction

Area of Science:

• Natural Language Processing
• Computational Linguistics

Background:

• Chinese spelling correction (CSC) faces challenges including misleading error signals, overemphasis on frequent characters, and limited training data.
• Existing CSC models struggle with nuanced errors and data scarcity, hindering performance.

Purpose of the Study:

• To introduce Decopy, a novel Chinese spelling correction model designed to overcome limitations of previous approaches.
• To enhance CSC accuracy by integrating semantic, positional, and phonetic (pinyin) features.

Main Methods:

• Decopy utilizes an advanced detection-correction framework with an innovative error masking strategy incorporating pinyin embeddings.
• The model captures semantic (word embeddings) and positional (position embeddings) information alongside phonetic features.
• A new CSC dataset derived from THUCNews was created for pre-training Decopy, addressing data scarcity.

Main Results:

• Decopy demonstrated significant performance improvements on the SIGHAN15 dataset and three domain-specific datasets (LAW, medical, official documents).
• The model outperformed previous state-of-the-art methods in Chinese spelling correction.
• Evaluation of large language models on CSC tasks was also conducted.

Conclusions:

• Decopy's integration of pinyin features effectively reduces reliance on ambiguous elements and misleading information.
• The proposed model offers a robust solution for Chinese spelling correction, particularly in specialized domains.
• Decopy represents a significant advancement in the field of Chinese spelling correction.