NeSyMatch: A Neuro-Symbolic Approach for Knowledge Alignment

Knowledge alignment is both a critical and complex challenge in the broader task of knowledge integration. With the remarkable success of transformer-based models across diverse domains, their potential has been increasingly explored for knowledge alignment. In this work, we present NeSyMatch, a novel neuro-symbolic knowledge alignment system that integrates transformer-based learning with symbolic reasoning techniques. NeSyMatch begins with a lexical matcher to identify and align lexically similar concepts. For the remaining concept pairs, it employs transformer-based similarity methods, selecting the most effective model through an ablation study comparing two fine-tuned variants. Finally, it refines the alignments using a large language model (LLM) and structural cues such as sibling relations. While NeSyMatch is designed to be domain-agnostic, its performance was evaluated on the Bio-ML track dataset from the Ontology Alignment Evaluation Initiative (OAEI). Experimental results show that NeSyMatch consistently outperforms prominent ontology matching (OM) systems, including BERTMap, LogMap, and Matcha in many scenarios.