A Chinese text correction method based on model fusion and scene self-adaption
By using parallel encoding of cross-model attention and Chinese character structure awareness text scene recognition algorithms, and integrating BERT, ERNIE and T5 models, a three-path mutual error correction model is constructed and character-by-character voting is performed. This solves the problems of insufficient language representation and scene adaptation in existing Chinese error correction models, and achieves efficient self-adaptive error correction for specific scenes.
Patent Information
- Application Number
- CN202511090344.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2026-03-03
- Estimated Expiration
- 2045-08-05
AI Technical Summary
Existing Chinese error correction models suffer from limited language representation capabilities and insufficient generalization ability when used individually. Traditional model cascading methods suffer from information redundancy and static fusion issues. Furthermore, general error correction models are not optimized for specific scenarios, resulting in limited error correction performance.
We adopt a model fusion and scene self-adaptation approach, which integrates BERT, ERNIE and T5 models through parallel encoding of cross-model attention paths, and combines them with the Chinese character structure perception text scene recognition algorithm CSA-TextCNN to construct a three-path mutual error correction model. We then perform character-by-character voting, and load the corresponding parameters for error correction after recognizing the text scene.
It improves language representation capabilities, mitigates the error correction bias and data noise effects of a single model, achieves self-adaptive error correction performance optimization in specific scenarios, solves information redundancy and static fusion problems, and improves error correction accuracy.
Smart Images

Figure CN120975078B_ABST
Abstract
Citation Information
Patent Citations
Text error correction model training method, application method, device and equipment
CN115146621A
Text error correction method and device based on GRU-Transform, electronic equipment and storage medium
CN120106053A