Line Item Extraction Training Using Confidence-Scored Auto Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for training machine learning models to extract information from structured documents are inefficient, costly, and prone to inaccuracies due to the need for extensive human-labeled training data and model hallucinations, especially with varying document formats and content.
Innovation Solution
A method involving optical character recognition (OCR) to extract text and positional data, followed by table detection, and using a natural language processing model to generate structured text with table tags, which is then used to train a compact item extraction model with confidence scores and a noise-aware supervised learning process to prevent hallucinations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If traditional rule-based or structure-based approaches are used for information extraction, then the process is simple to implement, but the accuracy deteriorates due to inability to handle variable document formats and content
Solution Approach 1:
The patent replaces traditional rule-based mechanical extraction systems with a machine learning-based approach. The system uses a trained model that processes document images through optical character recognition and table structure analysis to extract information, substituting rigid mechanical rules with adaptive neural network processing that can handle variable formats.
Solution Approach 2:
The patent changes the approach from fixed extraction parameters to adaptive parameters. The machine learning model learns optimal extraction parameters during training on labeled data, allowing the system to automatically adjust extraction criteria based on document content and structure variations rather than using fixed rules.
2Reliability
If machine learning models are trained with large amounts of human-labeled data, then extraction accuracy improves, but the cost and time required for training deteriorates
Solution Approach 1:
The patent uses synthetic data generation to create copies of labeled training data. The system generates synthetic document images and their corresponding labels through controlled rendering, allowing the model to be trained on artificially created data that replicates real document structures without requiring manual annotation of each sample.
Solution Approach 2:
The system employs self-supervised learning mechanisms where the model can generate its own training data through synthetic document rendering. The process automatically creates labeled datasets by rendering documents with known structures and extracting information algorithmically, eliminating the need for human annotators to manually label each training sample.
3Loss of time
If machine learning models are trained with limited labeled data, then training cost decreases, but extraction accuracy deteriorates due to insufficient format variation coverage
Solution Approach 1:
The patent segments the training process into multiple stages: pre-training on synthetic data with controlled variations, then fine-tuning on real labeled data. This segmentation allows the model to learn general extraction patterns from synthetic examples before adapting to specific real-world document formats, improving accuracy without requiring all training data to be manually labeled.
Solution Approach 2:
The system performs preliminary training on synthetic data that pre-establishes the model's ability to handle various document formats and structures. This preliminary action prepares the model to accurately extract information from real documents without requiring extensive labeled real document data, as the synthetic pre-training already covers format variations.
4Speed
If the extraction model uses greedy autoregressive decoding, then generation speed improves, but model hallucinations increase due to erroneous patterns
Solution Approach 1:
The patent implements feedback mechanisms that monitor and correct model predictions during generation. The system uses confidence scores and validation checks to detect and correct hallucinations, allowing the model to self-correct erroneous patterns while maintaining efficient generation speeds through optimized decoding strategies.
Data Source
AI summary
Aspects of the present disclosure provide techniques for training an item extraction machine learning model. Embodiments include extracting text and bounding box coordinates from a structured document and creating structured text by adjusting formatting of the extracted text based on the extracted bounding box coordinates and adding table delimiter tags to the extracted text based on detecting one or more tables in the structured document. Embodiments include providing the structured text to a language processing machine learning model along with a prompt instructing the language processing machine learning model to generate a label indicating variables present in the structured text and values for the variables. Embodiments include receiving the label from the language processing machine learning model in response to the structured text and the prompt and training the item extraction machine learning model through a supervised learning process based on training data comprising the structured text and the label.