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.

CN122433719APending Publication Date: 2026-07-21泰山科技学院
View PDF 0 Cites 0 Cited by

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

Technical Problem

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.

Method used

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.

Benefits of technology

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.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122433719A_ABST
    Figure CN122433719A_ABST
Patent Text Reader

Abstract

The application discloses a Korean text error correction system based on deep learning and belongs to the technical field of language error correction, and specifically comprises the following steps: a sequence construction module splices a character embedding vector and a boundary state vector into a character boundary joint sequence; a feature extraction module outputs a multi-dimensional joint feature matrix through an attention network; a graph convolution module dynamically generates weights through a hypernetwork and performs graph convolution; a parallel decoding module calculates a position adjustment and a joint probability distribution of literal replacement; an instruction generation module synchronously analyzes and generates a space boundary adjustment instruction and a character correction instruction; and a text reconstruction module completes literal mapping and physical space reconstruction according to the two types of instructions and outputs a corrected text. Through the joint representation of characters and boundary states and the parallel joint decision mechanism, the application eliminates the error transmission of a traditional series structure and effectively improves the error correction robustness of Korean text in a space and literal compound error scenario.
Need to check novelty before this filing date? Find Prior Art