Form-Like Document Extraction Using Spatial Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current techniques for extracting information from form-like documents, such as invoices, are inefficient and error-prone, particularly when dealing with diverse templates and formats like PDF or scanned images, lacking spatial presentation hints and requiring manual effort or brittle heuristics.
Innovation Solution
A machine-learned model that generates embeddings for candidate text portions based on their position and neighbor relationships, using self-attention layers to assign these portions to specific fields in a target schema, enhancing accuracy and efficiency in extracting relevant information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual effort or brittle heuristics are used for information extraction, then the system is simple to implement, but the extraction accuracy and reliability are poor
Solution Approach 1:
The patent replaces manual extraction processes and brittle heuristic rules with a machine learning-based system that uses neural networks to automatically extract information from form-like documents. The system converts document images into feature vectors and uses a scoring model to identify and extract relevant information, eliminating the need for manual intervention while achieving high extraction accuracy across diverse document templates
Solution Approach 2:
The patent transforms the extraction process by changing from rule-based parameter matching to a data-driven approach where the system learns patterns from training data. The system uses feature vectors that capture spatial relationships and contextual information, allowing the model to adapt to variations in document layouts without requiring explicit rules for each scenario
2Adaptability or versatility
If traditional text extraction approaches are used, then the system can process natural text, but it fails on form-like documents with tables and grid formatting
Solution Approach 1:
The patent segments the document processing into distinct stages: converting the document image to a structured representation, extracting candidate text portions, generating feature vectors that capture spatial and contextual relationships, and using a scoring model to identify relevant information. This segmentation allows the system to handle the unique characteristics of form-like documents with tables and grids separately from general text processing
Solution Approach 2:
The patent introduces an intermediary representation layer in the form of feature vectors that capture spatial relationships and contextual information. These feature vectors serve as a bridge between the raw document image and the extraction task, enabling the system to understand the structure of form-like documents and accurately extract information regardless of the specific template used
3Adaptability or versatility
If the system processes documents in PDF or scanned image formats without spatial hints, then the processing is more universal, but the extraction becomes more difficult
Solution Approach 1:
The patent performs preliminary processing by converting the document image into a structured representation with extracted text portions and their spatial positions before the actual extraction task. This preliminary action creates a ready-to-process data structure that captures spatial relationships, making the subsequent extraction more efficient and accurate while maintaining support for various input formats like PDF and scanned images
4Ease of manufacture
If the system is trained on a small number of manually labeled examples, then the training data requirement is low, but the model cannot generalize well to unseen templates
Solution Approach 1:
The patent creates a universal extraction system that can handle multiple document templates within a domain using a single trained model. The system learns general patterns from a small number of training examples and applies them to unseen templates through the feature vector representation that captures transferable spatial and contextual relationships, eliminating the need for template-specific training for each document type
Data Source
AI summary
The present disclosure is directed to extracting text from form-like documents. In particular, a computing system can obtain an image of a document that contains a plurality of portions of text. The computing system can extract one or more candidate text portions for each field type included in a target schema. The computing system can generate a respective input feature vector for each candidate for the field type. The computing system can generate a respective candidate embedding for the candidate text portion. The computing system can determine a respective score for each candidate text portion for the field type based at least in part on the respective candidate embedding for the candidate text portion. The computing system can assign one or more of the candidate text portions to the field type based on the respective scores.