Form-Like Document Extraction with Spatial Embeddings for Diverse Templates
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 neighborhood, using self-attention layers to score and assign text portions to specific fields in a target schema, enabling automated extraction from both digital and scanned documents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual effort or brittle heuristics are used for information extraction, then the system can handle diverse document templates, but the extraction process is time-consuming and error-prone
Solution Approach 1:
The patent replaces manual information extraction (mechanical human effort) with an automated machine learning system that uses trained models to extract data from documents. The system substitutes human operators with computational algorithms that can process documents automatically, reducing both time and human error while maintaining high accuracy through learned patterns from training data.
Solution Approach 2:
The patent transforms the extraction approach by changing from rule-based heuristics to learned parameters through machine training. The system learns optimal extraction parameters from labeled training examples, allowing it to adapt to diverse document templates. This parameter learning enables the system to maintain high accuracy across different document types without requiring manual rule configuration for each template.
2Productivity
If rule-based heuristics are used for extraction, then the system can process documents quickly, but the system becomes brittle and error-prone with diverse templates
Solution Approach 1:
The patent makes the extraction system dynamic by training machine learning models on diverse document examples. Instead of static rules that break with template variations, the system learns adaptive patterns from training data and can generalize to new document templates. This dynamic learning capability allows the system to maintain both speed and accuracy when processing diverse document types.
Solution Approach 2:
The patent performs preliminary training action by pre-training the machine learning model on labeled document examples before deployment. This preliminary learning phase enables the system to internalize extraction patterns for various document templates, so that during actual processing, the model can quickly and accurately extract information without relying on brittle rule-based heuristics.
3Adaptability or versatility
If traditional NLP approaches are used, then the system can handle natural text, but the system cannot effectively process form-like documents with tables and grid formatting
Solution Approach 1:
The patent segments the document processing task into distinct components: first extracting text and spatial information from the document image, then using this structured information as input to the machine learning model for information extraction. This segmentation allows the system to handle form-like documents with tables and grids by processing their unique spatial structure separately before applying NLP techniques, thereby increasing adaptability without excessive complexity.
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.