ENTITY EXTRACTION USING A CODER-DECODER MACHINE LEARNING MODEL
Patent Information
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- INTUIT INC
- Filing Date
- 2023-05-31
- Publication Date
- 2026-07-15
AI Technical Summary
Existing data extraction systems face challenges in minimizing execution time while maintaining accuracy, as computers struggle to associate tokens with entity labels and require separate processing steps to group tokens into entities.
An encoder-decoder machine learning model processes document images directly to generate entity values and labels, bypassing the need for intermediate token labeling and grouping, using models like T5, BERT, ViT, or LayoutLMv2 to create an encoder hidden state vector and a decoder model that outputs raw text with entity labels and values.
This approach simplifies the data extraction process, reducing training and execution time while maintaining high accuracy by directly producing structured entity representations without additional processing steps.