Self-Improving Document Classification for RPA
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current robotic process automation (RPA) lacks effective solutions for splitting electronic computer files into document types suitable for document processing, which hinders efficient document classification and processing.
Innovation Solution
A method for classifying and splitting electronic files by iteratively determining whether each portion belongs to an existing document or represents a new one, using a trained classifier to assess confidence based on word vectors and document layout, and adjusting confidence scores based on document length, allowing for accurate extraction and output of documents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional document processing methods are used in RPA, then document classification and splitting cannot be performed effectively, but implementing a new classification system increases system complexity
Solution Approach 1:
The system performs self-training by automatically learning from document examples without requiring manual configuration or complex pre-programming. The classifier improves its performance autonomously by processing documents and updating its internal models, eliminating the need for complex manual setup while achieving high processing efficiency
Solution Approach 2:
The system performs preliminary training actions by processing a set of example documents before actual document classification begins. This preliminary learning phase prepares the classifier to efficiently handle subsequent document processing tasks, establishing a foundation for high productivity without requiring complex real-time decision-making logic
2Productivity
If manual document classification is performed, then accuracy can be maintained, but processing time and labor costs increase significantly
Solution Approach 1:
The system incorporates feedback mechanisms where classification results are continuously evaluated and used to refine the classifier's performance. Confidence scores are calculated for each classification, and the system can learn from corrections or adjustments, maintaining high accuracy while operating at automated speeds
Solution Approach 2:
The system dynamically adjusts classification parameters such as confidence thresholds and learning rates based on the complexity and characteristics of the documents being processed. This allows the system to maintain high accuracy across different document types while optimizing processing speed, achieving both goals simultaneously
3Measurement precision
If document splitting is performed without confidence assessment, then processing speed increases, but classification accuracy decreases
Solution Approach 1:
The system applies confidence assessment selectively rather than uniformly to all documents. For high-confidence classifications, the system proceeds quickly without additional verification, while only applying deeper validation to low-confidence cases. This partial application of verification maintains accuracy where needed while preserving processing speed for the majority of straightforward cases
Data Source
AI summary
Systems and methods for classifying and splitting an electronic file into a plurality of extracted documents are provided. The electronic file is received. An initial portion of the electronic file is classified using a trained classifier and extracted from the electronic file as an extracted document associated with the classification. It is iteratively determined whether each respective next portion of the electronic file should be added to the extracted document until it is determined that the respective next portion should not be added to the extracted document. In response to determining that the respective next portion should be added to the extracted document, the respective next portion is extracted from the electronic file and added to the extracted document. In response to determining that the respective next portion should not be added to the extracted document, the classifying and the iteratively determining are repeated using the respective next portion as the initial portion. The extracted documents are output. The trained classifier can be trained to learn sets of word vectors and other relevant information associated with document classifications, in order to improve accuracy.