Text Mining Patent Tables via Vector Encoding and Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data extraction tools are inadequate for processing large and complex tables in patent specifications, particularly in chemistry, as they are designed for smaller web tables and lack semantic analysis capabilities, leading to inefficiencies in identifying key information.
Innovation Solution
A method involving the transformation of table cells into vector representations using a sequential 2-D model, such as Q-LSTM, followed by image classification to summarize table semantics and map to probability distributions for labels, leveraging pre-trained models like BERT and ResNet18 for improved table and cell-level classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated data extraction tools are used, then processing speed increases, but accuracy and reliability of information extraction deteriorate due to tools being designed for simpler web tables rather than complex patent tables
Solution Approach 1:
The patent transforms the fundamental parameters of table representation by converting tables into image formats and applying deep learning models (CNN, ResNet) originally designed for image recognition. This parameter change allows the system to process complex patent tables with high accuracy while maintaining automated processing speed, resolving the contradiction between extraction speed and accuracy.
Solution Approach 2:
The patent replaces traditional mechanical text parsing methods with neural network-based semantic analysis. By using pre-trained language models and image classification models, the system substitutes rule-based extraction mechanisms with intelligent models that can understand table semantics, thereby improving accuracy without sacrificing automation efficiency.
2Measurement precision
If manual processing of patent specifications is performed, then information extraction accuracy improves, but time consumption and processing cost increase significantly
Solution Approach 1:
The patent implements a self-service automated system that uses pre-trained language models and image classification models to independently process patent tables without human intervention. The system automatically performs table detection, image conversion, model inference, and result extraction, eliminating the need for manual processing while maintaining high accuracy through intelligent algorithms.
Solution Approach 2:
The patent applies pre-trained language models and pre-trained image classification models that have been trained in advance on large datasets. These pre-trained models provide ready-to-use semantic understanding and table recognition capabilities, allowing the system to achieve high accuracy immediately without requiring time-consuming manual processing or extensive on-the-fly training.
3Ease of operation
If traditional text parsing tools are used, then processing simplicity is maintained, but capability to handle large complex tables deteriorates
Solution Approach 1:
The patent creates a universal processing framework that handles diverse table formats (CSV, Excel, PDF, images) through a unified pipeline. The system converts all table types into image representations and processes them through the same deep learning models, providing both operational simplicity and high adaptability to different table structures and formats encountered in patent specifications.
Data Source
AI summary
A method for text mining from one or more tables is provided. The method includes the steps of: receiving one or more tables, the tables having one or more table labels, and one or more cells to be processed, transforming each of the cells into cell vector representations; encoding the one or more cell vector representations with a sequential 2D model; obtaining one or more table-level vector representations by summarising the semantics of the cell vector representations by an image classification model; and mapping the output of to an output vector which represents the probability of each of the table labels.