File Matching Rationale Alignment Neural Networks Optimal Transport
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current file matching technologies lack interpretability and sparsity in alignment, making it difficult to understand how predictions are made, especially in tasks like similar document recommendation and fake news detection, due to the complexity of deep neural networks and the non-sparse nature of attention-based models.
Innovation Solution
A self-explaining neural network system that uses optimal transport algorithms to compute alignment between vector representations of file objects, with techniques such as adding dummy points and duplicating points to achieve sparse and interpretable alignments, allowing for the selection and alignment of matching objects to be jointly identified and optimized.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If attention-based models are used for aligning rationales in file matching, then alignment computation is performed at multiple intermediate layers, but the alignments become non-sparse and difficult to interpret
Solution Approach 1:
The patent extracts the alignment computation from multiple intermediate layers and concentrates it at the final layer only. This extraction principle resolves the contradiction by removing unnecessary intermediate alignment computations that contribute to complexity while preserving the essential alignment capability at the output layer, ensuring sparsity and interpretability.
Solution Approach 2:
Instead of applying attention-based alignment at multiple intermediate layers as conventional approaches do, the patent inverts the approach by applying alignment computation only at the final layer. This inversion eliminates the sparsity problem caused by multiple intermediate alignments while maintaining the adaptability of the model.
2Reliability
If neural attention is applied for aligning rationales, then alignment between document objects is achieved, but most pairs have non-zero alignment values making interpretation difficult
Solution Approach 1:
The patent changes the parameter of alignment sparsity by modifying when and how attention is applied. By concentrating alignment computation at the final layer and using a specific attention mechanism design, the patent achieves sparse alignments where only relevant object pairs have non-zero values, thereby improving interpretability while maintaining alignment accuracy.
3Measurement precision
If multiple intermediate layers with attention computation are used, then comprehensive alignment is achieved, but the prediction cannot be attributed to specific alignment choices
Solution Approach 1:
The patent extracts alignment computation from intermediate layers and places it only at the final layer. This extraction enables clear attribution of predictions to specific alignment choices between document objects, as the alignment at the final layer directly determines the matching decision without being obscured by multiple intermediate alignment operations.
4Productivity
If standard optimal transport is applied for alignment, then alignment computation is performed, but the alignments lack sufficient sparsity for interpretability
Solution Approach 1:
The patent modifies the optimal transport computation by applying it only at the final layer rather than at multiple intermediate layers. This parameter change in the computation timing and location produces sparse alignments that are both computationally efficient and interpretable, resolving the contradiction between productivity and alignment sparsity.
Data Source
AI summary
The present disclosure describes a system, method, and computer program for matching an input file to one of a plurality of datastore files and displaying the rationale for the match. Neural networks are trained to create vector representations of objects in the input file and the datastore files. The cost of each possible pairing of vector representations between the input file and a datastore file is computed, and an optimal transport algorithm is used to identify the vector pairings that result in the lowest total cost of alignment. The datastore file with the lowest total cost of alignment to the input file is identified as the best matching file. The alignment results are used to display the rationale for the match. To constrain the alignment results of the optimal transport algorithm, one or more dummy points and, in certain embodiments, duplicate points are added to one or both of the vector sets to achieve alignments with the desired sparsity patterns.