Template Invariant Information Extraction via Dual Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems face challenges in efficiently extracting and classifying dynamic fields from documents with variations, requiring extensive rule creation and frequent updates, especially when dealing with small datasets and diverse document structures.
Innovation Solution
A system utilizing two neural network models, including a graphical neural network, to classify documents by recognizing entities and their locations, generating confidence scores, and determining template membership through a threshold value, allowing for efficient extraction and classification without the need for extensive rule creation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional template matching using images and patterns is used, then template matching can be performed when documents comprise images and patterns, but template extraction fails when documents fail to disclose any image and pattern
Solution Approach 1:
The patent replaces the mechanical image-based template matching system with a text-based entity extraction system using neural networks. Instead of relying on visual patterns and images, the system uses NLP models to extract entities and their relationships from text, enabling template extraction to work on documents without images or patterns while maintaining reliability
Solution Approach 2:
The patent creates a universal entity extraction system that can handle multiple document types and formats through a single neural network model. The system extracts entities and relationships from text regardless of document structure, making it adaptable to various document formats including those without images or patterns
2Measurement precision
If a machine is provided with artificially framed rules for each field, then information extraction can be performed for specific document types, but the number of rules increases exponentially as variations grow
Solution Approach 1:
The patent implements a self-learning neural network system that automatically learns extraction patterns from training data without requiring manual rule creation. The model extracts entities and relationships by learning from examples, eliminating the need for humans to frame and manage complex rules for each document variation while maintaining high extraction precision
Solution Approach 2:
The patent transforms the extraction system from a rule-based approach with discrete parameters to a neural network approach with continuous learned parameters. The model learns optimal extraction parameters automatically from training data, reducing the complexity of managing numerous discrete rules while maintaining precision across document variations
3Adaptability or versatility
If a machine requires a large number of documents for training to handle variations, then the machine can understand document structure and relationships, but the system cannot effectively process small datasets with large variations
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network model on a diverse dataset of document variations before deployment. This pre-training enables the model to learn general document structures and relationships in advance, allowing it to effectively process small datasets with large variations without requiring extensive training data for each specific case
Data Source
AI summary
A system for template invariant information extraction. The system comprises of processor, a first neural network model and a second neural network model. The processor is configured to recognize and extract entities and location of the entities in the input document using the first neural network model. The processor is further configured to classify whether the input document belongs to at least a template of the documents of the first training dataset using the second neural network model. The second neural network model comprises a linear classifier configured to generate a plurality of confidence scores for the input document corresponding to a unique template of the documents of the first training dataset. A threshold value to classify the input document belonging to template of the documents of the first training dataset is determined, Classification is done by comparing the confidence score with the threshold value.