Deeper and Deeper: A Lightweight Semi-Supervised Deep Reinforcement Adaptive Learning-Based Ontology Alignment

Ontology alignment, also known as ontology matching, is pivotal for addressing semantic heterogeneity on the Semantic Web. Essentially, it entails linking entities across different ontologies or knowledge graphs in order to resolve ambiguity and enhance the interoperability of data. While various techniques exist, many still rely on rule-based or logic-based approaches, often requiring human intervention and domain specificity. Despite these challenges, ontology alignment remains crucial for seamlessly integrating disparate knowledge sources and facilitating effective data integration. In this paper, we tackle the limitations of current ontology alignment models by introducing a novel, lightweight, semi-supervised deep reinforcement learning model called Deep Reinforcement Adaptive Learning for Ontology Alignment (DRAL-OA). The DRAL-OA method incorporates both syntactic and structural information into the training phase. In addition, this approach is semi-supervised, utilizing a portion of the training data and automatically generating the rest, which reduces the need for human intervention. Moreover, DRAL-OA uses non-domain-specific language models to ensure broad applicability and reduce the need for extensive domain expertise. We evaluate our proposed approach using two datasets from the Ontology Alignment Evaluation Initiative (OAEI). In our experiments, we have shown that the proposed model can achieve high-quality alignments with F-measures on par with other state-of-the-art systems, all while maintaining a very short runtime and a compact model size.