Document Graph Neural Networks for Non-Sequential Key-Value Pairing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sequential-type machine learning models struggle to accurately identify question-answer pairs in FAQ-type documents due to their non-sequential arrangements, leading to incorrect pair identifications.
Innovation Solution
A graph neural network (GNN) is trained to recognize key-value pairs by generating document graphs with nodes representing sentences and edges indicating relationships, using attributes like font, location, and read-order to classify sentences as key or value types, and identify correct pair groupings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If sequential-type machine learning models are used to identify question-answer pairs, then the model implementation is simple, but the identification accuracy deteriorates when answers are arranged non-sequentially relative to questions
Solution Approach 1:
The patent transitions from sequential processing to graph-based structural representation by organizing document elements (questions, answers, headers) as nodes and their relationships as edges in a graph structure. This dimensional change allows the model to capture non-sequential relationships and hierarchical document structures, significantly improving identification accuracy while maintaining reasonable implementation complexity through established graph neural network frameworks
2Measurement precision
If graph neural network is used to identify key-value pairs, then the identification accuracy improves for non-sequential arrangements, but the model complexity increases
Solution Approach 1:
The patent segments the document processing task into distinct components: extracting document elements (questions, answers, headers), constructing graph structures with nodes and edges, training the graph neural network separately, and then applying the trained model for prediction. This segmentation allows each component to be optimized independently and simplifies the overall implementation despite the increased model capability
Solution Approach 2:
The patent introduces graph structures as an intermediary representation layer between the raw document text and the final key-value pair identification. This intermediary graph structure mediates the complex relationships in the document, making them more tractable for the neural network while maintaining clear interpretability of relationships between questions and answers
3Ease of operation
If applications extract and reproduce content based on read order, then the content reproduction is straightforward, but the content correspondence identification fails when answers are positioned away from their questions
Solution Approach 1:
The patent moves beyond linear read-order processing by constructing graph structures that represent document hierarchy and semantic relationships in multiple dimensions. Nodes represent document elements while edges encode various relationship types (parent-child, question-answer, header-content), allowing the system to reliably identify content correspondence regardless of spatial position or read order in the original document
Data Source
AI summary
Techniques for identifying content in key-value pairs of documents using a graph neural network (GNN) are disclosed. A system trains a GNN to identify key-value pair groupings in documents. The GNN classifies nodes in document graphs as key-type nodes and answer-type nodes. The GNN also classifies edges connecting nodes in the document graphs for keeping in the document graph or removing from the document graph. The resulting document graph includes key-value groupings in the document. Upon identifying content matching a query, a system returns as a query response content from among the key-value groupings.