Ontology Matching via Adaptive Blocking and Labeling Function Committee
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ontology matching algorithms are inefficient in identifying matches between source and target ontologies due to limited learning capabilities, requiring extensive human effort and struggling with extreme class imbalance, leading to suboptimal performance in finding true matches with affordable effort.
Innovation Solution
A computer-implemented method employing an adaptive blocking mechanism and two collaborative learning loops to filter non-matching pairs, select and weight labeling functions, and automatically create tuned labeling functions to enhance the efficiency and accuracy of ontology matching, utilizing adaptive blocking and tunable labeling functions to reduce candidates and improve prediction results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing ontology matching algorithms are used, then prediction performance can be achieved with low risk, but efficiency of identifying matches is limited and requires extensive human effort
Solution Approach 1:
The patent segments the ontology matching process into multiple independent labeling functions (LF1, LF2, LF3, LF4) that can operate in parallel. Each labeling function handles specific aspects of matching (e.g., string similarity, structural similarity, statistical properties), allowing the system to process large numbers of ontology pairs simultaneously while maintaining high prediction accuracy through ensemble voting.
Solution Approach 2:
The patent applies partial action by using multiple labeling functions that may not be individually perfect but collectively provide comprehensive coverage. The system evaluates more than enough labeling functions (4 functions with different strategies) to ensure robust matching, accepting that individual functions may have limitations but the ensemble provides excessive coverage for reliable matching.
2Measurement precision
If learning-based algorithms are used, then matching accuracy can improve, but they require lots of labeled data for training which is not available in ontology matching
Solution Approach 1:
The patent implements self-service by designing labeling functions that are self-sufficient and do not require external training data. Each labeling function uses inherent properties of the ontology data (string similarity, structural relationships, statistical characteristics) to autonomously generate predictions. The system serves itself by combining these self-generated predictions through voting, eliminating the need for labeled training data while maintaining high accuracy.
Solution Approach 2:
The patent uses disposable labeling functions that are simple, lightweight rules rather than complex trained models. These labeling functions are easy to implement, modify, and replace without requiring retraining. The system employs multiple simple labeling functions (string comparison, structural comparison, statistical properties) that can be independently applied to any ontology pair without needing substantial labeled data for training.
3Ease of manufacture
If heuristic-based algorithms are used, then matching can be performed without training data, but performance varies largely across different ontology datasets
Solution Approach 1:
The patent achieves universality by designing a multi-functional labeling function ensemble that can adapt to different ontology domains and datasets. The four labeling functions cover diverse matching strategies (string-based, structure-based, statistical-based), allowing the system to universally apply to various ontology types while maintaining consistent performance. The voting mechanism integrates results from all functions, providing robust performance across different datasets without requiring domain-specific tuning.
Solution Approach 2:
The patent applies parameter changes by allowing flexible adjustment of labeling function weights and thresholds based on specific ontology characteristics. The system can modify the voting weights of individual labeling functions or adjust their internal parameters to optimize performance for particular domains. This adaptability enables the system to maintain high reliability across different ontology datasets while keeping the overall architecture simple and easy to implement.
4Ease of operation
If active learning with random selection is used, then human annotation can be guided, but it fails to bootstrap learning capability due to extreme class imbalance
Solution Approach 1:
The patent applies preliminary action by having the labeling functions continuously operate in the background to generate predictions and update their weights before human annotation is needed. The system pre-processes ontology pairs using all four labeling functions, and the voting mechanism is already in place to guide annotation. This preliminary computation and weighting setup enables human annotators to focus on the most uncertain cases without requiring complex real-time interaction, thus guiding annotation while maintaining automation.
Data Source
AI summary
A method for ontology matching between a source and a target filters out non-matching pairs of the source and the target, to generate a dataset of possible matches. In a first loop, based on prediction results and uncertainty from a set of labeling functions of a labeling function (LF) committee, a data point is selected from the dataset and an annotation label is obtained for the data point. Additionally, labeling functions of the LF committee are selected and weighted based on prediction results against the dataset provided with annotation labels, and a weight of each of the selected LFs is adjusted to produce the prediction results and uncertainty of yet unlabeled data points of the dataset based on the data points of the dataset having already annotated a label. A second learning loop is executed that creates tuned labeling functions and augments the LF committee with the tuned labeling functions.