Attribute Extraction via Drawing Position Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing attribute extraction systems require manual preparation of document formats, are limited to pre-defined pattern templates, and struggle to distinguish attribute names from values, especially in non-HTML documents and repetitive structures, making them labor-intensive and inefficient.
Innovation Solution
An attribute extraction method that identifies attribute names and values by analyzing the drawing positions of strings or images in documents, calculating an attribute name score to determine aggregation, and selecting groups based on shared positions and co-occurrence frequencies to automatically extract attributes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual preparation of document formats is used, then attribute extraction can be performed, but time and labor are consumed
Solution Approach 1:
The system automatically analyzes document structures and learns attribute patterns without manual intervention. The attribute extraction unit autonomously identifies and extracts attributes by analyzing drawing positions and spatial relationships of text elements, eliminating the need for manual format preparation while maintaining extraction accuracy
Solution Approach 2:
The system performs preliminary analysis of document structures to learn attribute patterns before actual extraction. By pre-analyzing the spatial arrangements and relationships of text elements in documents, the system prepares extraction models in advance, enabling rapid and accurate attribute extraction without manual format preparation during the extraction process
2Ease of manufacture
If pre-defined pattern templates are used, then attribute extraction is possible, but application is limited to specific formats
Solution Approach 1:
The system uses universal spatial relationship analysis that works across different document formats. By analyzing the fundamental spatial arrangements of text elements (such as left-aligned text indicating attribute names and right-aligned text indicating values) rather than format-specific patterns, the system achieves both ease of implementation and broad adaptability to various document types including HTML and non-HTML documents
Solution Approach 2:
The system adapts to different document formats by changing analysis parameters dynamically. Instead of using fixed pattern templates, the system adjusts its analysis based on the actual spatial parameters and structural characteristics of each document, enabling it to handle diverse formats without requiring pre-defined templates for each format type
3Extent of automation
If tree structure analysis of HTML tags is used, then repetitive patterns can be recognized, but attribute name and value identification is difficult
Solution Approach 1:
The system transitions from one-dimensional tag structure analysis to two-dimensional spatial position analysis. By incorporating the spatial dimension (drawing positions, alignment, and relative locations of text elements) into the analysis, the system can automatically recognize patterns and precisely identify which elements are attribute names and which are values, overcoming the limitations of traditional tree structure analysis
4Reliability
If tag tree structure analysis is required, then HTML documents can be processed, but application to other document types is difficult
Solution Approach 1:
The system extracts and analyzes only the essential spatial relationship features from documents, separating these from the specific HTML tag structure. By focusing on the fundamental spatial arrangements of text elements rather than HTML-specific tag trees, the system maintains high accuracy for HTML documents while simultaneously achieving broad applicability to other document formats that lack HTML structure
Data Source
AI summary
Sets of strings of which the drawing positions are arranged in one direction are extracted from a document as attribute groups. An attribute name score is calculated for each attribute group to determine an extent to which each attribute group is a set of attribute names. Based on the attribute name scores, an attribute name group is selected out of the attribute groups. From among the attribute groups, an attribute group which includes a string which is the same as at least one string of the attribute name group and of which the drawing position is the same as that of the string of the attribute name group is selected. From the string at the same drawing position, an attribute name is extracted. From the other strings of the selected attribute group than those at the same drawing position, an attribute value corresponding to the attribute name is extracted.