Semantic Concept Matching via Weak Supervision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current ontology matching techniques face challenges in efficiently homogenizing data from heterogeneous sources, requiring numerous labeling functions, low coverage, and low accuracy due to sparsity of labels, which increases computational complexity and reduces data availability.
Innovation Solution
The method enhances ontology matching by computing pairwise similarity scores between semantic concepts using multiple similarity metrics, aggregating these scores to increase the density of the labeling matrix, and applying a label aggregator to apply a single label per concept pair, thereby reducing the need for manual labeling and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If numerous labeling functions are used to improve ontology matching accuracy, then matching accuracy is improved, but device complexity and computational resources required increase
Solution Approach 1:
The patent uses similarity metrics to create virtual copies of labeled concept pairs to infer labels for unlabeled pairs. Instead of requiring numerous labeling functions to cover all concept pairs, the system copies information from well-labeled pairs to unlabeled pairs through similarity-based inference, reducing the number of actual labeling functions needed while maintaining accuracy.
Solution Approach 2:
The patent introduces similarity metrics as an intermediary mechanism between labeled and unlabeled concept pairs. This intermediary allows the system to transfer labeling information indirectly through semantic similarity, avoiding the need for direct labeling of all pairs and reducing overall system complexity.
2Measurement precision
If manual labeling is performed for all concept pairs to ensure accuracy, then labeling accuracy is improved, but loss of time and computational resources increase
Solution Approach 1:
The patent performs preliminary labeling of a subset of concept pairs using automated labeling functions, then uses similarity metrics to infer labels for remaining pairs. This preliminary action on a representative sample allows the system to avoid manual labeling of all pairs while maintaining accuracy through similarity-based generalization.
Solution Approach 2:
The system uses its own similarity metrics and labeled data to automatically infer labels for unlabeled concept pairs, making the labeling process self-service. Instead of requiring continuous manual intervention, the system serves itself by using inferred labels to improve its matching capabilities over time.
3Productivity
If the labeling matrix is sparse with few labels to reduce computational complexity, then processing efficiency is improved, but reliability and accuracy of ontology matching decrease
Solution Approach 1:
The patent copies labels from well-labeled concept pairs to unlabeled pairs that are semantically similar. This copying mechanism increases the effective number of labels in the matrix without requiring actual manual annotation of all pairs, thereby improving reliability while maintaining processing efficiency through automated inference.
Solution Approach 2:
The patent merges information from multiple similar concept pairs into a single inferred label. By combining similarity-based inferences across multiple pairs, the system increases label density and reliability without proportionally increasing computational complexity, as the same similarity metrics are applied across all pairs.
Data Source
AI summary
A method augments data labels of a machine learning task. The method includes applying at least one labeling function for each of a plurality of pairs of semantic concepts and producing a labeling matrix, computing pairwise similarity scores using similarity metrics for each of the pairs of semantic concepts, augmenting the labeling matrix using the pairwise similarity scores to increase a density of the labeling matrix, and inputting the labeling matrix to a label aggregator to apply a single label for each of the pairs of semantic concepts labeled by the at least one labeling function.