Table Structure Reconstruction for Accurate Header-Cell Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in accurately extracting information from tables embedded in documents due to varied and low-quality representations, cell spanning issues, and representation errors, particularly in semi-structured formats like SGML, HTML, and XML, which affect the association between cells and their headers.
Innovation Solution
A method to recognize headers and merged cells, create a richer table structure representation, and link cells to their respective row and column headers, enabling accurate extraction of information from heterogeneous tables.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If tables are represented in semi-structured formats like SGML, HTML, or XML to enable structured data storage and retrieval, then data organization and accessibility are improved, but table structure quality deteriorates due to cell spanning, representation errors, and loss of association between cells and headers
Solution Approach 1:
The patent introduces an intermediary processing system that acts as a mediator between the semi-structured table representation and the information extraction process. This system includes components for detecting table structures, identifying headers, resolving cell spanning issues, and correcting representation errors before information extraction, thereby maintaining data accessibility while improving structure quality
Solution Approach 2:
The patent applies preliminary actions by performing table structure analysis, header identification, and error correction before the actual information extraction process. The system pre-processes tables to establish correct cell-header associations and resolve structural issues, ensuring that subsequent extraction operations work with cleaned and validated table data
2Adaptability or versatility
If cells are allowed to span multiple rows or columns to represent complex data relationships, then data representation flexibility is improved, but information extraction accuracy deteriorates due to loss of association between cells and their respective headers
Solution Approach 1:
The patent applies segmentation by dividing spanned cells into multiple individual cell instances, each with explicit associations to their respective headers. The system segments complex spanned structures into manageable units that can be independently processed, thereby maintaining the flexibility of spanned representations while enabling accurate information extraction through clear cell-header mappings
Solution Approach 2:
The patent introduces an additional dimensional layer by creating explicit association structures that link spanned cells to their headers through index terms and positional relationships. This additional dimension of association metadata allows the system to track and maintain cell-header connections even when cells span multiple rows or columns, preserving both flexibility and extraction accuracy
3Loss of information
If OCR and cell breaking are used to improve readability and extract text from tables, then text accessibility is improved, but representation errors increase affecting extraction accuracy
Solution Approach 1:
The patent implements feedback mechanisms that detect representation errors introduced by OCR and cell breaking processes. The system analyzes extracted text for consistency, validates table structure integrity, and uses this feedback to identify and correct errors, thereby maintaining text accessibility while improving representation accuracy through iterative validation and correction
Data Source
AI summary
Much valuable information in documents is presented within tables. However, the information within tables is hard to extract automatically with high accuracy due to the wide variety and low quality of typical tables found in electronic documents. Information extraction technology can provide a method of extracting information from heterogeneous tables by recognizing tables, the header cells, and cells that are merged or should be merged, creating a richer representation of table structure and providing a convenient way of linking cells to their row and column headers. Use of this richer representation allows a few extraction patterns to successfully pull out information from a wide variety of differently formatted tables.