Positional Embeddings for Document Layout Preservation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current OCR and machine learning techniques for document analysis fail to preserve positional information, leading to inaccurate results in document processing due to the loss of two-dimensional layout and relationships between symbols.
Innovation Solution
The system generates document labels using positional embeddings, which are two-dimensional coordinates that preserve the layout information, by combining these embeddings with word vectors in a neural network to enhance accuracy and efficiency in label generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If traditional OCR and machine learning techniques are used to extract characters and words from document images, then character and word extraction can be achieved, but positional information and layout information are lost
Solution Approach 1:
The patent transforms the one-dimensional sequential processing of document tokens into two-dimensional spatial representation by mapping token positions to their original document coordinates. This dimensional transformation preserves spatial relationships and layout information that would otherwise be lost in traditional sequential processing, enabling more accurate document analysis.
Solution Approach 2:
The patent introduces positional embeddings as an intermediary mechanism that bridges the gap between token extraction and layout preservation. These embeddings encode spatial position information and integrate it with token representations, allowing the model to maintain awareness of document structure while processing sequential data.
2Reliability
If two-dimensional layout information is preserved in document processing, then analysis accuracy is improved, but computational complexity and resource requirements increase
Solution Approach 1:
The patent segments the document processing task into distinct components: token extraction, positional embedding generation, and fused representation processing. By dividing the complex task of layout-aware document analysis into manageable segments, the system achieves high accuracy while maintaining computational efficiency through specialized processing for each component.
Solution Approach 2:
The patent transforms spatial coordinates into embedding vectors with specific dimensional parameters, changing the representation format from continuous two-dimensional coordinates to discrete vector parameters. This parameter transformation enables efficient computational handling of positional information while preserving the essential spatial relationships needed for accurate document analysis.
3Loss of information
If positional embeddings are integrated with token sequences, then layout information is maintained, but processing time and computational resources increase
Solution Approach 1:
The patent generates positional embeddings in advance during the tokenization phase, before the main processing occurs. This preliminary action ensures that spatial information is already encoded and ready for integration, eliminating the need for computationally intensive operations during the main processing loop and thereby maintaining high processing speed.
Solution Approach 2:
The patent merges positional embeddings with token representations through efficient vector operations, combining spatial and semantic information in a single integrated representation. This merging approach avoids separate processing streams and reduces computational overhead, maintaining productivity while preserving layout information.
Data Source
AI summary
Disclosed herein are system, method, and computer program product embodiments for generating document labels using positional embeddings. In an embodiment, a label system may identify tokens, such as words, of a document image. The label system may apply a position vector neural network to the document image to analyze the pixels and determine positional embedding vectors corresponding to the words. The label system may then combine the positional embedding vectors to corresponding word vectors for use as an input to a neural network trained to generate document labels. This combination may embed the positional information with the corresponding word information in a serialized manner for processing by the document label neural network. Using this formatting, the label system may generate document labels in a light-weight and fast manner while still preserving spatial relationships between words.