Neural Key-Value Extraction for Variable Document Layouts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for detecting key-value associations in unstructured electronic documents rely heavily on manual heuristics and require extensive re-training for documents with different layouts and types, leading to inefficiencies and reduced accuracy.
Innovation Solution
A neural network system comprising embeddings models, a document context model, key hypotheses model, value hypotheses model, and a KVA model to generate and evaluate key-value hypotheses, leveraging embeddings to encode object properties and context, and utilizing multiple subnetworks for efficient identification across diverse document types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual heuristics and templates are used for object detection, then the system can be configured with specific detection rules, but the process requires extensive human operations and is time-consuming
Solution Approach 1:
The patent replaces manual configuration of heuristics and templates with an automated neural network system. The neural network model automatically learns detection patterns from training data, eliminating the need for manual rule creation and configuration. This substitution of mechanical human operations with an automated intelligent system resolves the contradiction by maintaining detection accuracy while dramatically reducing processing time and human intervention.
Solution Approach 2:
The patent implements preliminary training of the neural network model on extensive document data before actual detection tasks. This pre-processing phase allows the system to learn and store detection patterns in advance, so that during actual document processing, the model can quickly apply learned knowledge without requiring manual rule configuration for each new document type, thus reducing real-time processing time while maintaining accuracy.
2Adaptability or versatility
If conventional object detection methods are used, then the system can identify objects in documents, but it requires extensive re-training for documents with different layouts and types
Solution Approach 1:
The patent implements a universal neural network model that can detect multiple types of objects across different document layouts and types using a single system. The model is designed to handle various document formats (invoices, receipts, forms, etc.) and object types (text, tables, images, handwritten content) without requiring separate detection systems for each document type, thereby increasing adaptability while managing system complexity through unified architecture.
Solution Approach 2:
The patent utilizes parameter changes in the neural network model to adapt to different document types and layouts. By adjusting model parameters and hyperparameters based on the specific document characteristics, the system can optimize detection performance for different document formats without requiring complete re-training, thus improving versatility while controlling the complexity of adaptation procedures.
3Reliability
If manual configuration of heuristics is used, then the system can detect key-value associations, but it involves a large number of human operations
Solution Approach 1:
The patent replaces manual configuration of detection heuristics with an automated neural network system that learns key-value association patterns from training data. The system automatically identifies relationships between document elements (such as form fields and their corresponding values) without requiring manual rule creation, thereby maintaining detection reliability while dramatically improving ease of operation by eliminating complex manual configuration steps.
Solution Approach 2:
The patent implements a self-training mechanism where the neural network model automatically learns and adapts to different document types and key-value association patterns through exposure to training data. The system serves itself by automatically adjusting its detection parameters and patterns based on the data it processes, eliminating the need for continuous manual configuration and heuristic updates, thus improving operational simplicity while maintaining reliability.
Data Source
AI summary
Aspects of the disclosure provide for mechanisms for identification of text fields in documents using neural networks. A method of the disclosure includes obtaining vectors, representative of objects in a document and processing the vectors to generate key hypotheses associating key(s) with one or more objects and value hypotheses associating value(s) with zero or more objects. The method further includes generating key-value association (KVA) hypotheses associating a selected key hypothesis with a selected value hypothesis and characterized by a KVA likelihood score that is based on at least a key likelihood score associated with the selected key hypothesis and a value likelihood score associated with the selected value hypothesis. The method further includes identifying one or more target KVAs of the document using the KVA likelihood scores of the generated KVA hypotheses.