Encoder-Decoder Model for Merging Extraction Outputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document extraction models have limitations in accuracy and efficiency when dealing with different types of documents, as they are optimized for specific tasks and may not perform well across varying document types.
Innovation Solution
The method involves training an encoder-decoder model using a combination of extraction model outputs, character recognition data, and ground truth key-value pairs from training documents, framing the problem as a multimodal text generation task to improve the accuracy of key information extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple extraction models are used to handle different document types, then the adaptability and accuracy for specific document types improve, but the device complexity and difficulty of merging outputs increase
Solution Approach 1:
The patent merges outputs from multiple extraction models by concatenating their key-value pairs and training an encoder-decoder model to process the combined data. The encoder processes inputs from multiple models while the decoder generates unified outputs, effectively combining the capabilities of different specialized models into a single integrated system that maintains adaptability across document types.
Solution Approach 2:
The encoder-decoder model serves as a universal framework that can process outputs from various extraction models trained on different document types. This multi-functional model learns to handle diverse input formats and generates consistent outputs across different domains, eliminating the need for separate processing pipelines for each document type.
2Measurement precision
If extraction models are optimized for specific tasks and document types, then the measurement precision for those specific types improves, but the reliability across varying document types deteriorates
Solution Approach 1:
The patent employs feedback mechanisms through the training process where the encoder-decoder model learns from concatenated outputs of specialized models and ground truth data. The model continuously refines its understanding by comparing its predictions against ground truth, enabling it to maintain high precision for specific document types while improving overall reliability across varying types through iterative learning.
Solution Approach 2:
The system changes the operational parameters by training the encoder-decoder model on aggregated data from multiple specialized models. This parameter transformation allows the model to internalize the strengths of different specialized models and adapt its extraction behavior to match the optimal performance characteristics of each document type while maintaining consistency across all types.
3Device complexity
If a single extraction model is used for all document types, then the device complexity is reduced, but the measurement precision and adaptability deteriorate
Solution Approach 1:
The patent segments the extraction task into multiple specialized extraction models, each optimized for specific document types, and then combines their outputs through the encoder-decoder framework. This segmentation allows each component to maintain high precision for its specialized domain while the integration layer provides overall system simplicity by unifying the outputs into a single processing pipeline.
Data Source
AI summary
Systems and methods for training an encoder-decoder model are disclosed. An example method includes receiving, over a communications network, a plurality of extraction model outputs from a corresponding plurality of extraction models, each extraction model output received from a corresponding extraction model and each extraction model output including a respective plurality of key-value pairs corresponding to extracted text from one or more training documents, receiving, over the communications network, character recognition data corresponding to the one or more training documents, receiving, over the communications network, ground truth key-value data corresponding to the one or more training documents, and training the encoder-decoder model based at least in part on the plurality of extraction model outputs, the character recognition data, and the ground truth key-value data, wherein the trained encoder-decoder model is configured to generate key-value pairs for subsequent outputs of the plurality of extraction models.