Document Information Extraction via Semantic and Pixel Space Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for extracting information from documents are either rigid and require intensive maintenance when using templates or are costly and labor-intensive due to the need for large amounts of labeled training data.
Innovation Solution
A computer-implemented method and system that identifies candidates within a document, generates embedding and location vectors, compares them to previous entity values in semantic and pixel spaces, and sorts candidates based on weighted comparison results to efficiently extract information without extensive manual labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If templates are used for information extraction, then the system structure is simple, but the system becomes rigid and requires intensive maintenance
Solution Approach 1:
The patent replaces the mechanical template-matching system with a machine learning-based semantic analysis system. Instead of using fixed templates that require manual maintenance, the system employs trained models that automatically learn and adapt to extract entity values from documents, thereby eliminating the rigidity and maintenance burden of template-based approaches while preserving structural simplicity through automated processing.
2Measurement precision
If large amounts of labelled training data are used, then the extraction accuracy is improved, but the acquisition cost and time increase significantly
Solution Approach 1:
The patent performs preliminary actions by pre-training machine learning models on comprehensive datasets before deployment. The models are trained in advance to learn semantic relationships and entity patterns, so that during actual extraction operations, they can achieve high accuracy without requiring additional manual labeling or data preparation time for each specific extraction task.
Solution Approach 2:
The system enables self-service by allowing the machine learning models to automatically learn and adapt from the data provided during extraction operations. The models can refine their understanding of entity values and relationships through continuous operation, reducing the need for external manual intervention and extensive pre-labeling of training data.
3Reliability
If manual labelling is performed to acquire training data, then the model can be trained, but the process becomes expensive and labor-intensive
Solution Approach 1:
The patent replaces the manual mechanical process of data labeling with automated machine learning systems. The models are designed to learn from unlabelled or minimally labelled data, using self-supervised learning techniques where the system automatically identifies and learns from patterns in the data without human intervention, thereby maintaining training quality while dramatically reducing labor requirements and costs.
Data Source
AI summary
A method for extracting information from a document, comprising: receiving an identification of an entity to be extracted from the document; identifying candidates from the document, each candidate corresponding to a given element contained in the document and having a given location within the document; embedding the candidates, thereby obtaining an embedding vector for each candidate; for each candidate, comparing in a semantic space the respective embedding vector to previous embedding vectors associated with previous entity values previously chosen for the entity, thereby obtaining a first comparison result; for each candidate, comparing in a pixel space the given location within the document of the candidate to a location associated with the previous entity values previously chosen for the entity, thereby obtaining a second comparison result; sorting the candidates using the first and second comparison results obtained for each candidate, thereby obtaining sorted candidates; and outputting the sorted candidates.