Keyword Identification in Document Images Using Subword Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computerized systems face challenges in accurately processing business documents due to variations in formatting, terminology, and visual differences, leading to errors in extracting key-value pairs from digitized documents, especially when received in image form.
Innovation Solution
A computerized system and method that uses subword embeddings and two-dimensional geometric contexts within a gradient boosted trees classifier to identify and segment keywords in document images, generating feature vectors for automated identification of key-value pairs, enabling the recognition of novel keywords and disambiguation based on context.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional computerized processing methods are used to extract data from business document images, then processing can be performed, but accuracy is reduced due to variations in formatting, terminology, and visual differences
Solution Approach 1:
The patent segments the document image into multiple text segments, then further segments each text segment into subword units (character groups). This hierarchical segmentation allows the system to process documents at multiple granularities, improving accuracy by analyzing local patterns while maintaining adaptability to various document formats and terminologies.
Solution Approach 2:
The patent transforms text segments into n-dimensional vectors using subword embeddings, converting textual information into a numerical parameter space. This parameter transformation enables the gradient boosted trees classifier to process and compare text data mathematically, improving identification accuracy while handling variations through vector space operations.
2Reliability
If simple text extraction is used from document images, then processing speed is maintained, but error rate increases due to visual differences from scanning and formatting variations
Solution Approach 1:
The patent introduces subword embeddings as an intermediary layer between the raw text segments and the final keyword identification. This intermediary transformation converts text into n-dimensional vectors that capture semantic and contextual information, serving as a bridge that improves reliability while managing complexity through a structured processing pipeline.
Solution Approach 2:
The patent replaces traditional mechanical text extraction and pattern matching with a machine learning-based gradient boosted trees classifier. This substitution uses trained models to automatically identify keywords, improving reliability by learning from annotated training documents while the modular architecture manages system complexity.
3Productivity
If manual annotation methods are used for keyword identification, then accuracy can be maintained, but productivity is reduced due to time-consuming processing
Solution Approach 1:
The patent performs preliminary actions by creating annotated training documents in advance, where keywords are pre-identified and associated with feature vectors. This preliminary annotation work is done once to train the gradient boosted trees classifier, enabling automated high-speed processing afterward while maintaining accuracy through the learned model.
Solution Approach 2:
The patent implements a self-service system where the gradient boosted trees classifier automatically identifies keywords in new documents without requiring manual annotation. The system uses the trained model to process documents autonomously, dramatically improving productivity while maintaining the accuracy achieved during the training phase.
Data Source
AI summary
A system and method of automatically learning new keywords in a document image based on context such as when a never before seen keyword exists surrounded by other key-value pairs. A machine learning based approach leverages subword embeddings and two-dimensional geometric contexts in a gradient boosted trees classifier. Keys may be composed of multi-word strings or single-word strings.