Unstructured Document Label Extraction via Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems struggle to accurately extract label-value pairs from unstructured documents with variable formats, as labels and values can be dynamically placed, making it challenging to train extraction systems using fixed location-based techniques.
Innovation Solution
The system employs a deep learning model pre-trained with data including complete document layouts, which predicts the likely location of labels and values, and uses geometrical association and OCR techniques to extract label-value pairs, even when values exist without associated labels or are in the form of logos.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fixed location based extraction techniques are used, then extraction accuracy for standardized documents is improved, but adaptability to documents with variable formats deteriorates
Solution Approach 1:
The system transitions from static fixed-location extraction to dynamic location prediction using a trained machine learning model. The model learns from multiple document templates and dynamically predicts the location of label-values based on document content and structure, allowing the extraction point to vary adaptively across different document formats while maintaining accuracy.
Solution Approach 2:
The system changes the parameter of extraction location from fixed to variable by training a model on spatial attributes from multiple templates. The model outputs predicted locations as continuous parameters rather than fixed coordinates, enabling adaptation to different document layouts while maintaining extraction precision.
2Adaptability or versatility
If training data from multiple templates is used, then adaptability to different document formats is improved, but training complexity and data requirements increase
Solution Approach 1:
The system creates a universal extraction model that can handle multiple document templates and formats through a single trained system. Rather than creating separate extraction logic for each template, the model learns common patterns across templates and applies them universally, reducing overall system complexity while maintaining broad adaptability.
Solution Approach 2:
The system uses template copying where spatial attributes and structural patterns from multiple document templates are extracted and used to train a generalizable model. The model learns to replicate the labeling patterns across different templates without requiring template-specific extraction rules, simplifying the training process while maintaining format versatility.
3Measurement precision
If deep learning models are trained on spatial attributes, then extraction accuracy for variable formats is improved, but computational resources and training time increase
Solution Approach 1:
The system performs preliminary training of the deep learning model on a comprehensive dataset of spatial attributes from multiple templates before deployment. This upfront training investment creates a pre-trained model that can quickly and accurately extract label-values from variable formats without requiring extensive computation during actual extraction operations, trading initial training time for operational efficiency.
Data Source
Figure 1~2A
Figure 2B
Figure 3A
AI summary
This disclosure relates generally to method and system for determining label value for labels in unstructured documents. Typical systems have challenge in understanding variations in layout of unstructured documents and extract information therefrom. The disclosed method and system facilitate systematically identifying sections and bounding boxes in the page images, taking image portion of the bounding boxes and extracting labels and label values therefrom. In case the label values are not present in the same bounding box having the label, the neighboring labels are examined for the matching label values. The system also obtains label-label value pairs from the document by utilizing a trained deep learning model, and compares the output with the label-label value pairs extracted earlier. An aggregated confidence score is assigned to the text in the bounding box.