Abstract:
Ontology Matching is a critical task to establish semantic interoperability given the proliferation of ontologies and knowledge graphs with overlapping domains. While traditional Ontology Matching relied on heuristics and rule-based approaches to find corresponding entities between knowledge resources, recent advances in machine-learning have prompted the community to contemplate matching approaches that exploit machine-learning algorithms.
We present Matcha-DL, an extension of the matching system Matcha to tackle semi-supervised tasks using machine-learning algorithms. Matcha builds upon the algorithms of the established system AgreementMakerLight with a novel broader core architecture designed to tackle long-standing challenges such as complex and holistic ontology matching. Matcha-DL uses a linear neural network that learns to rank candidate mappings proposed by Matcha by using a partial reference alignment as a training set, and using the confidence scores produced by Matcha's matching algorithms as features.
Matcha-DL was evaluated in the 2022 and 2023 editions of the Bio-ML track of the Ontology Alignment Evaluation Initiative, achieving the highest F1 score in 4 of the 5 semi-supervised tasks. Furthermore, it was shown to benefit more than other competitors from the contextual information of ontologies.