Automated SDS Information Extraction Using Visual ML and NLP
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current technological solutions fail to effectively index and extract targeted, complex information from digital documents, particularly those with variable and unpredictable text structures, such as Safety Data Sheets (SDS) in the chemical industry, due to inconsistencies in formatting and the need for intelligent machine learning systems capable of handling diverse data structures.
Innovation Solution
A system combining machine learning techniques, including Computer Vision and Natural Language Processing, with rule-based systems to identify and extract chemical ingredient information from SDS documents, using pre-trained models like BERT NER and Cascade Mask R-CNN, to recognize text and extract relevant sections and data items like CAS numbers and weight percentages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If rule-based systems are used to extract information from digital documents, then the system is simple and easy to implement, but it fails to handle variable and unpredictable text structures effectively
Solution Approach 1:
The system transitions from fixed rule-based parameters to dynamic machine learning parameters that can adapt to varying document structures. The ML models are trained on diverse datasets to learn parameter variations across different SDS formats, enabling the system to handle unpredictable text structures while maintaining automated extraction capabilities.
Solution Approach 2:
The patent replaces traditional mechanical rule-based extraction systems with intelligent machine learning systems. Instead of using predefined patterns and rigid parsing rules, the system employs trained ML models (such as NLP and computer vision models) that can automatically adapt to different document structures, thereby solving the limitation of rule-based systems while maintaining automation.
2Adaptability or versatility
If machine learning networks are used to handle diverse data structures, then adaptability improves, but system complexity increases
Solution Approach 1:
The system divides the complex information extraction task into multiple specialized machine learning components. Each ML model is trained to handle specific aspects (e.g., text recognition, table extraction, chemical ingredient identification), allowing the overall system to manage diversity in document structures while keeping individual components relatively simple and modular.
Solution Approach 2:
The patent introduces intermediate processing layers between the input documents and final extraction results. These intermediaries include pre-trained ML models that serve as mediators, transforming diverse input formats into standardized representations that can be processed by downstream extraction modules, thereby reducing overall system complexity.
3Measurement precision
If manual indexing methods are used, then accuracy can be maintained, but productivity decreases due to time consumption
Solution Approach 1:
The system enables automated self-service information extraction using machine learning networks that can independently identify and extract targeted information from digital documents without human intervention. The ML models are trained to achieve accuracy comparable to manual indexing while processing documents at automated speeds, eliminating the trade-off between accuracy and productivity.
Solution Approach 2:
The patent incorporates feedback mechanisms where the ML models learn from extraction results and continuously improve accuracy. Through training on annotated datasets and iterative optimization, the system achieves high precision in information extraction while maintaining automated processing speeds, resolving the contradiction between manual accuracy and automated productivity.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and methods for automated indexing and extraction of information in digital documents are disclosed. A method may comprise selecting a page number of a digital document to identify a page containing targeted information; inputting an image of the page into a visual machine learning network (visual ML), wherein the visual ML is trained to recognize text associated with the targeted information in an image; identifying by the visual ML, a section of the image that contains the targeted information; inputting the page number, the digital document, and coordinates of the section into an extraction module; and extracting the targeted information by the extraction module from the section.