Hybrid Key Value Document Extraction via Vector Embeddings and Cascading Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic data extraction from documents is complex due to the inability of computers to 'read' documents like humans, especially when the document type is unknown, leading to difficulties in identifying relevant information.
Innovation Solution
A method using a vector embedding model to generate term encodings and a cascading classifier model to classify documents through a directed acyclic graph, allowing for the identification of target fields and extraction of corresponding values from the documents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single classification model is used to classify documents into multiple categories, then the device complexity is reduced, but the classification accuracy deteriorates due to the inability to capture hierarchical relationships between different document types
Solution Approach 1:
The patent divides the classification task into multiple independent binary classifiers, each responsible for distinguishing between two specific classes. This segmentation allows each classifier to focus on a specific discrimination task, improving accuracy while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The patent transforms the multi-class classification problem into a multi-dimensional space by creating a hierarchical structure of binary decisions. Each binary classifier represents a dimension in this space, and the combination of these dimensions enables precise classification of complex document types with hierarchical relationships.
2Measurement precision
If multiple binary classifiers are used to accurately classify documents with hierarchical types, then the classification accuracy is improved, but the device complexity increases due to the need for multiple classification models and coordination
Solution Approach 1:
The complex multi-class classification problem is segmented into multiple simple binary classification tasks. Each binary classifier handles a specific discrimination task, reducing the complexity of individual models while collectively achieving high accuracy through their coordinated decisions.
Solution Approach 2:
The patent implements a dynamic classification process where the system adaptively navigates through the hierarchy of binary classifiers based on the document being classified. This dynamic approach allows the system to activate only the necessary classifiers for each document type, reducing effective complexity while maintaining accuracy.
3Ease of operation
If traditional text processing methods are used to extract data from documents, then the ease of operation is maintained, but the extraction accuracy deteriorates because computers cannot understand document context like humans do
Solution Approach 1:
The patent introduces vector embeddings as an intermediary representation that bridges the gap between raw text and semantic understanding. These embeddings capture the contextual meaning of text, enabling accurate extraction while maintaining ease of operation through automated processing without requiring complex human-like interpretation.
4Measurement precision
If vector embeddings are used to capture document semantics for accurate extraction, then the data extraction accuracy is improved, but the computational resources and time required increase
Solution Approach 1:
The patent performs preliminary classification of documents into hierarchical types before executing the full data extraction process. This preliminary action enables the system to optimize subsequent processing by applying extraction methods tailored to specific document types, reducing overall processing time while maintaining high accuracy through context-aware extraction.
Data Source
AI summary
The method may include generating, by a vector embedding model, a vector embedding of multiple terms in an input document to obtain multiple term encodings. The method may also include generating, by a cascading classifier model, a classification of the input document. Generating the classification includes iteratively: traversing a directed acyclic graph ordering multiple class groups, and while traversing the directed acyclic graph, classifying the input document into a first class of a current class group of the class groups using the term encodings, where classifying the input document into the first class uses at least one second class of at least one parent class group of the class groups, and where the classification includes the first class and the at least one second class. The method may furthermore include obtaining a set of target fields corresponding to the classification. The method may in addition include extracting a set of values from the input document matching the set of target fields.