Synthetic Document Training Data With Auto-Generated Text Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The availability of training data for machine learning models, particularly language models, is limited due to the time-consuming and resource-intensive task of manually generating label data, which hampers the development of accurate and reliable models, especially for non-English languages.
Innovation Solution
An automated training dataset generation system (TDGS) that generates both training images and associated label data without human intervention, enabling the efficient production of diverse training datasets for various languages, including English and non-English languages, by varying fonts, formats, and incorporating text, image, and handwritten content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual label data generation is used, then label accuracy is maintained, but productivity is severely limited and time consumption increases
Solution Approach 1:
The system uses synthetic text generation to create copy-like representations of real document images. Instead of manually labeling each real image, the system generates synthetic images with embedded ground truth labels, effectively copying the essential structure and content patterns of real documents while automating the labeling process at scale
Solution Approach 2:
The system performs self-service by automatically generating both the training images and their corresponding label data without human intervention. The synthetic document generation process inherently creates the ground truth labels along with the images, eliminating the need for separate manual annotation efforts
2Productivity
If manual label data generation is used, then data quality is ensured, but resource consumption increases and scalability is limited
Solution Approach 1:
The system replaces resource-intensive manual labeling with synthetic image generation that copies the essential document structures and text patterns. This approach generates large volumes of training data with automatic ground truth labels while requiring fewer human resources and computational overhead compared to manual processing of equivalent data volumes
Solution Approach 2:
The system changes the fundamental parameter of data generation from manual human labor to automated synthetic generation. By transforming the generation process parameters (using AI models to create synthetic documents with embedded labels), the system achieves higher productivity and scalability while reducing the complexity of manual coordination and resource management
3Adaptability or versatility
If diverse training data is generated automatically, then model robustness improves, but ensuring label accuracy becomes more challenging
Solution Approach 1:
The synthetic document generation system serves itself by automatically creating ground truth labels as an inherent part of the image generation process. Since the system generates both the synthetic document images and their corresponding correct labels simultaneously, the labels are inherently accurate by construction, eliminating the accuracy challenges associated with manual annotation of diverse data
Data Source
AI summary
Techniques are described for automatically, and substantially without human intervention, generating training data where the training data includes a set of training images containing text content and associated label data. Both the training images and the associated label data are automatically generated. The label data that is automatically generated for a training image includes one or more labels identifying locations of one or more text portions within the training image, and for each text portion, a label indicative of the text content in the text portion. By automating both the generation of training images and the generation of associated label data, the techniques described herein are very scalable and repeatable and can be used to generate large amounts of training data, which in turn enables building more reliable and accurate language models.