Document Parsing with Visual Feature Encoding for Resume Data Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current OCR technologies fail to accurately extract information from resume documents due to the lack of consideration for visual features such as formatting, bolding, font size, and color, leading to incorrect interpretation and organization of data.
Innovation Solution
A processor-based system that parses document images to extract both textual and visual features, generates text and visual encodings, and uses machine learning models to identify sections and extract relevant data, incorporating OCR readers like Amazon Textract and Tesseract to recognize and correct textual information based on visual cues.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional OCR technology is used to extract text from document images, then text recognition can be achieved, but information extraction accuracy deteriorates due to lack of visual feature consideration
Solution Approach 1:
The patent merges traditional OCR text recognition with visual feature extraction into a unified processing system. The system simultaneously extracts text content and visual features (formatting, bolding, font size, color) from document images, combining these multiple data streams to improve information extraction accuracy while maintaining adaptability to various document formats.
Solution Approach 2:
The patent creates a composite encoding structure that integrates text encodings with visual encodings. This composite representation combines textual information and visual feature information into a unified data structure that feeds into machine learning models, enabling more accurate information extraction by considering both text content and visual context.
2Measurement precision
If visual features are incorporated into document processing, then information extraction accuracy improves, but system complexity increases
Solution Approach 1:
The patent segments the complex processing task into distinct modules: text extraction module, visual feature extraction module, encoding module, and machine learning model execution module. Each module handles a specific aspect of processing, which manages system complexity by breaking down the overall task while maintaining the benefits of integrated visual and text processing.
Solution Approach 2:
The patent introduces encoding structures as intermediary representations between raw document images and machine learning models. Text encodings and visual encodings serve as intermediate data structures that organize and standardize extracted information, making it more manageable for subsequent processing and reducing the complexity burden on downstream components.
3Manufacturing precision
If multiple encoding types are generated from document images, then data organization quality improves, but processing time increases
Solution Approach 1:
The patent performs preliminary encoding of both text and visual features during the initial processing stage. By generating text encodings and visual encodings upfront, the system prepares structured data that can be efficiently reused by machine learning models, improving data organization quality while managing processing time through advance preparation.
Data Source
AI summary
An apparatus including a processor caused to receive document images, each including representations of characters. The processor is caused to parse each document image to extract, based on structure type, subsets of characters, to generate a text encoding for that document image. For each document, the processor is caused to extract visual features to generate a visual encoding for that document image, each visual feature associated with a subset of characters. The processor is caused to generate parsed documents, each parsed document uniquely associated with a document image and based on the text and visual encoding for that document image. For each parsed document, the processor is caused to identify sections uniquely associated with section type. The processor is caused to train machine learning models, each machine learning model associated with one section type and trained using a portion of each parsed document associated with that section type.