Transformer-Based OCR Text Correction for Noisy Documents
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optical character recognition (OCR) technologies often produce erroneous text due to noise in old documents and variation in typesetting, leading to issues in downstream applications.
Innovation Solution
A bidirectional autoregressive transformer model is fine-tuned using datasets like token masking, sentence permutation, and document rotation to detect and correct errors such as oversegmentation, undersegmentation, misrecognized characters, missing characters, and hallucinations in OCR text.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional OCR technology is used to convert scanned images to text, then the process is simple and fast, but the resulting text contains errors due to noise in old documents and variation in typesetting
Solution Approach 1:
The transformer model is pretrained on large corpora before being fine-tuned on OCR-specific tasks. This preliminary pretraining equips the model with general language understanding capabilities, which are then specialized for OCR error correction through fine-tuning on datasets containing OCR errors and their corrections
Solution Approach 2:
A transformer model serves as an intermediary between the noisy OCR output and the desired clean text. The model takes erroneous OCR text as input, processes it through its pretrained language understanding mechanisms, and outputs corrected text, effectively mediating the transformation from inaccurate to accurate text
2Productivity
If OCR text is corrected using traditional methods, then the process requires multiple steps including error detection and separate correction, but this increases processing time and complexity
Solution Approach 1:
The patent combines error detection and correction into a single unified transformer model that performs both functions simultaneously. The model is trained to identify errors and generate corrections in one pass, eliminating the need for separate detection and correction steps while maintaining high accuracy through its pretrained language understanding
Data Source
AI summary
OCR-text correction system and method embodiments are described. The OCR-text correction embodiments comprise or cooperate with a transformer-based sequence-to-sequence language model. The model is pretrained to denoise corrupted text and is fine-tuned using OCR-correction-specific examples. Text obtained at least in part through OCR is applied to the fine-tuned pretrained transformer model to detect at least one error in a subset of the text. Responsive to detecting the at least one error, the fine-tuned pretrained transformer model outputs an updated subset of the text to correct the at least one error.