Financial Document Labeling With Confidence-Based Attribute Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic data extraction techniques for financial documents, such as invoices and receipts, are unreliable in efficiently and accurately extracting relevant data like entity names and bank details.
Innovation Solution
A method involving token extraction, preliminary attribute labeling, and confidence-based refinement using an expectation maximization model to determine accurate labels for entity attributes like ABN, BSB, and account numbers, leveraging inherent characteristics and checksums for verification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing automatic data extraction techniques are used, then data extraction speed is maintained, but extraction accuracy and reliability deteriorate
Solution Approach 1:
The patent segments the data extraction process into multiple independent modules: token extraction module, preliminary attribute labeling module, confidence value determination module, and refined label determination module. Each module handles a specific aspect of the extraction process, allowing for specialized optimization and reducing errors that would propagate through a monolithic system.
Solution Approach 2:
The patent introduces confidence values as an intermediary mechanism between preliminary attribute labels and final refined labels. This intermediary layer allows the system to evaluate and filter extraction results, using checksums and inherent characteristics to verify accuracy before finalizing data, thereby improving both precision and reliability.
2Measurement precision
If multiple verification steps are added to improve accuracy, then data extraction accuracy improves, but processing time increases
Solution Approach 1:
The patent performs preliminary attribute labeling on tokens before full document processing, using inherent characteristics and checksums to pre-validate data. This preliminary action filters out obviously incorrect extractions early, reducing the need for extensive verification later and maintaining processing efficiency while improving accuracy.
Solution Approach 2:
The patent replaces manual verification processes with automated confidence value calculation and threshold-based filtering. The system automatically determines confidence values for each attribute label and applies predefined thresholds to accept or reject extractions, eliminating time-consuming manual review while maintaining high accuracy standards.
3Measurement precision
If confidence-based filtering is applied, then data quality improves, but the number of processed documents decreases
Solution Approach 1:
The patent applies confidence-based filtering selectively to different attribute types and document sections rather than uniformly across all data. Critical fields like bank details and entity names undergo strict confidence threshold checking, while less critical fields use more lenient criteria, maintaining data quality for important information while preserving processing volume overall.
Solution Approach 2:
The patent dynamically adjusts confidence thresholds based on document type, attribute importance, and extraction context. For high-stakes extractions like financial account numbers, higher confidence thresholds are applied, while for less critical fields, lower thresholds maintain processing throughput. This parameter adaptation balances data quality requirements with productivity constraints.
Data Source
AI summary
A method comprises determining token(s) from financial documents and determining a set of preliminary attribute labels for the token(s), wherein the set is associated with attribute type(s). The method further comprises providing the set for each token to an attribute prediction model to determine, for the token, a confidence value for each attribute type(s), determining subsets of token, each subset being associated with a respective document of the plurality of documents and determining a set of refined labels for each document based on the confidence values, wherein the set of refined labels comprises a value for attribute type(s).