Document Classification Platform Using Confidence Threshold Routing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The process of manual classification and data extraction from financial documents is time-consuming, inefficient, prone to human error, and costly, with automated image classification systems being unreliable and text classification systems being computationally intensive.
Innovation Solution
A document classifying platform that uses a combination of image and text classification engines, where documents are initially processed through an image classifying engine and selectively through a text classifying engine based on confidence scores, to quickly and accurately classify documents without sacrificing accuracy or computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual classification and data extraction is used, then accuracy is maintained, but processing time increases and costs increase
Solution Approach 1:
The classification system is segmented into multiple specialized classifiers (image-based classifier, text-based classifier, and hybrid classifier) that process documents through different pathways. The image-based classifier handles visually distinctive documents quickly, the text-based classifier processes documents requiring textual analysis, and the hybrid classifier handles complex cases, thereby reducing overall processing time while maintaining accuracy through specialized handling of different document types
Solution Approach 2:
A confidence score mechanism acts as an intermediary between the image-based classifier and the text-based classifier. When the image-based classifier's confidence score exceeds a threshold, the classification is finalized quickly; when it falls below the threshold, the text-based classifier is invoked as an intermediary to provide additional analysis, ensuring accuracy is maintained only when necessary
2Measurement precision
If text classification engine is always used, then classification accuracy is improved, but computational resources are excessively consumed
Solution Approach 1:
The system applies partial action by using only the image-based classifier for documents where visual features provide sufficient classification confidence. The text-based classifier, which is computationally intensive, is applied only partially to documents where the image-based classifier's confidence score falls below the threshold, thereby consuming computational resources only when necessary while maintaining accuracy for the majority of documents
Solution Approach 2:
The system substitutes the mechanical text processing operation with an image-based classification approach for documents where visual features are sufficient. This replacement eliminates the need for computationally intensive text extraction and analysis for many documents, significantly reducing computational resource consumption while maintaining classification accuracy through the use of appropriate classification pathways
3Productivity
If image classification engine is used, then processing speed is improved, but reliability decreases
Solution Approach 1:
The system implements feedback through the confidence score mechanism that monitors the reliability of image-based classification results. When the confidence score indicates low reliability (below threshold), the system automatically triggers the text-based classifier to re-evaluate the document, using the feedback from the initial image-based classification to determine whether additional text analysis is needed to ensure reliable classification
Solution Approach 2:
The system performs preliminary classification using the faster image-based classifier before committing to a final classification decision. This preliminary action allows most documents to be processed quickly, while only those documents where the preliminary classification shows low confidence (and thus potentially low reliability) undergo the more thorough text-based analysis, ensuring reliability is checked in advance for problematic cases
Data Source
AI summary
A device obtains image data associated with a document. Using a first machine learning model, the device determines, for the document, a first classification of one of a plurality of document types and a first confidence score associated with the first classification, and a second classification of one of the plurality of document types and a second confidence score associated with the second classification based on the image data. The device determines a difference between the first confidence score and the second confidence score, compares the difference and a threshold value, and accept the first classification of the document when the difference satisfies the threshold value.