Document Text Extraction Models Decoupled From Document Types
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document extraction systems face challenges in efficiently handling new document layouts and require significant setup effort due to client-specific learning and model redundancy, leading to increased costs and time-to-solution.
Innovation Solution
A method that attributes learning models to extraction entities rather than document types, enabling independent improvement and faster configuration of new document types, utilizing a 'swarm effect' for model sharing across users, with automated validation and retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If learning models are bound to document types, then extraction accuracy for specific document types is improved, but setup time and cost increase significantly for each new document type
Solution Approach 1:
The patent segments the learning model into two independent components: a document type classifier and an extraction entity model. The extraction entity model is trained once and reused across multiple document types, while only the classifier needs to be trained for new document types. This segmentation eliminates the need to retrain extraction models for each new document type, significantly reducing setup time while maintaining extraction accuracy.
Solution Approach 2:
The extraction entity model is designed to be universal and document-type independent. Once trained on extraction entities from various document types, the model can be applied to any new document type without retraining. This multi-functionality allows the same extraction model to serve multiple document types, reducing both setup time and cost while preserving extraction accuracy through the specialized entity-focused training.
2Reliability
If client-specific learning is implemented, then extraction performance is optimized for each client, but system complexity and training costs increase
Solution Approach 1:
The patent extracts the extraction entity model training from the client-specific customization process. Instead of training separate models for each client, the system trains a universal extraction entity model that can be applied across all clients. This extraction of the common training component reduces system complexity while maintaining extraction performance through the specialized entity-focused approach.
Solution Approach 2:
The patent merges the training of extraction entities across multiple clients and document types into a single unified model. By combining training data and models from different clients into one universal extraction entity model, the system reduces complexity while improving reliability through larger, more diverse training datasets that enhance generalization across different client needs.
3Measurement precision
If separate learning models are trained for each document type, then extraction accuracy is maintained, but model redundancy and processing overhead increase
Solution Approach 1:
The patent extracts the extraction entity model as a separate, independent component that is document-type independent. This extracted model is trained once and reused across multiple document types, eliminating the redundancy of training separate models for each document type while maintaining extraction accuracy through the specialized entity-focused training approach.
Solution Approach 2:
The extraction entity model is designed with universal applicability across multiple document types. Instead of creating redundant document-type-specific models, a single universal extraction entity model serves multiple document types, reducing model quantity and redundancy while maintaining extraction accuracy through the generalized entity recognition capability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a process for automatically extracting information from document files by attributing learning models and extraction mechanisms to extraction entities, the process comprising the following partial processes individually, complementarily or in combination: aggregation of the learning data and extraction data from individual fields; logical separation of the learning models and extraction mechanisms from document types; automated learning from document extraction data by: a. generating the learning data from the data validation of the document processing, b. centrally validating all learning data for relevance and quality, c. centrally or distributedly retraining the extraction methods with the aid of the evaluated learning data, d. evaluating the newly created extraction methods and models.