A deep learning-based korean text correction system
By concatenating character embedding vectors and boundary state vectors into a joint sequence in the Korean text error correction system, and using graph convolution and parallel decoding modules to calculate the joint probability distribution, spatial boundary adjustment and character correction instructions are generated. This solves the problem of word structure destruction caused by misjudgment of spaces, and improves the error correction robustness and readability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 泰山科技学院
- Filing Date
- 2026-04-28
- Publication Date
- 2026-07-21
AI Technical Summary
Existing deep learning-based Korean text correction systems are prone to disrupting morpheme structure when dealing with space segmentation errors, making it difficult to accurately extract contextual semantic features. Furthermore, they struggle to balance the joint probability distribution of spatial position adjustments and literal substitutions, resulting in correction results that deviate from the original semantics and lack robustness and readability.
The sequence construction module concatenates the character embedding vector and the boundary state vector into a joint character boundary sequence. The graph convolution module dynamically generates graph convolution weights. The parallel decoding module calculates the joint probability distribution of position adjustment and literal replacement, and generates spatial boundary adjustment instructions and character correction instructions in parallel. Finally, the text reconstruction module performs physical space reconstruction.
It effectively avoids the collapse of morpheme structure caused by misjudgment of spaces, ensuring that the error correction results are highly faithful to the semantic logic of the Korean context while maintaining the rationality of the original sentence's physical layout. This greatly improves the system's error correction robustness and output readability in complex text environments.
Smart Images

Figure CN122433719A_ABST