Named Entity Recognition Using Spatial Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional named entity recognition systems face challenges in accurately identifying named entities in documents with non-standard formatting and OCR errors, as they rely solely on sequential multi-word snippets and fail to incorporate spatial information, leading to inaccuracies in recognizing entities like invoice amounts in tabular structures.
Innovation Solution
The system creates a feature map with a grid of cells corresponding to document portions, calculating text and spatial features, and uses a convolutional neural network to estimate tag locations and recognize named entities, incorporating spatial information to improve accuracy and independence from reading order errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional named entity recognition systems use sequential multi-word snippets only, then the processing is simple, but the recognition accuracy deteriorates due to inability to handle non-standard formatting and OCR errors
Solution Approach 1:
The patent transitions from one-dimensional sequential text processing to two-dimensional spatial-aware processing by creating a feature map that maps text features to their spatial locations in the document image. This allows the system to simultaneously consider both the sequential context and spatial relationships of text elements, improving named entity recognition accuracy without excessive complexity increase.
Solution Approach 2:
The patent segments the document into a grid of cells where each cell corresponds to a specific spatial region and contains text features from that region. This segmentation enables independent processing of different document areas while maintaining spatial context, allowing the system to handle non-standard formatting and OCR errors more effectively.
2Measurement precision
If the system processes documents without spatial information, then the processing is faster, but the recognition accuracy deteriorates for tabular structures and nested text
Solution Approach 1:
The patent performs preliminary action by pre-processing the document image to create the feature map and tagged map structures before named entity recognition. This preliminary organization of spatial and textual information enables faster processing during the actual recognition phase, as the system can directly query the pre-organized data structures rather than analyzing raw documents from scratch.
3Measurement precision
If the system relies on reading order alone, then the processing is straightforward, but it fails to accurately recognize entities in tabular structures where spatial position matters
Solution Approach 1:
The patent introduces an intermediary structure (the feature map) that mediates between the raw document image and the named entity recognition process. This intermediary layer organizes spatial and textual information in a way that preserves both reading order and spatial relationships, enabling accurate entity recognition in complex layouts without excessive complexity.
Data Source
AI summary
Methods and systems for recognizing named entities within the text of a document are provided. The methods and systems may include receiving a document image and recognized text of the document image. A feature map of the document image may be created, a tagged map may be created, and locations of tags within the tagged map may be estimated using a machine learning model. Named entities with the recognized text may be recognized based on the one or more locations of the tags. In some embodiments, the machine learning model is a convolutional neural network. In further embodiments, creating the feature map may include determining, for a subset of the cells of the feature map, one or more features of the recognized text contained in a corresponding portion of the document image.