Source-linked AI summary
OntoAligner-Ensemble: Voting-Based Fusion across Heterogeneous Ontology Alignment Techniques
Hamed Babaei Giglou, Sören Auer, Peio Popov, Mahsa Sanaei, Jennifer D'Souza
TL;DR
Ontology alignment frameworks can deploy heterogeneous aligners, but systematic reconciliation of their complementary and conflicting predictions remains underexplored. OntoAligner-Ensemble addresses this gap with a modular two-stage framework combining voting-based fusion and post-fusion selection, and evaluates it across eight OAEI tasks. The results show improved precision–recall balance, with ensemble composition shaping whether precision or F1 is favored.
Problem
Unified OA ecosystems support heterogeneous aligners, but a general configurable mechanism for reconciling their complementary and conflicting predictions remains missing.
Method
OntoAligner-Ensemble integrates predictions from heterogeneous aligners through voting-based fusion followed by configurable post-fusion selection policies.
Results
Voting-based ensemble fusion consistently improves the precision–recall balance and frequently outperforms individual constituent baselines across eight benchmark tasks.
Takeaways & Limitations
Ensemble composition provides a practical way to trade precision against recall and F1 across diverse ontology-alignment scenarios.
Takeaways & Limitations
The framework reduces relation decisions to binary match versus no-match classification and does not distinguish equivalence, subclass, and superclass relations.
Abstract
from arXiv · showhide
Ontology alignment (OA) has evolved through several methodological paradigms, ranging from lexical and structural aligners to knowledge graph embedding (KGE) models and, more recently, Large Language Model (LLM)-based approaches. Although modern OA frameworks provide unified ecosystems for deploying these heterogeneous aligners, mechanisms for systematically reconciling their complementary and sometimes conflicting predictions remain relatively underexplored. We present OntoAligner-Ensemble, a modular and aligner-agnostic framework that combines candidate correspondences through a configurable two-stage process comprising voting-based fusion strategies followed by post-fusion selection policies. The framework supports any aligner implemented within OntoAligner that produces candidate correspondences, enabling diverse alignment paradigms to be integrated through a unified decision process. To demonstrate its effectiveness, we instantiate the framework using representative lightweight string-aligner, KGE-based, and Retrieval-Augmented Generation aligners powered by both open-weight and API-based LLMs. We evaluate individual aligners and ensemble configurations across eight benchmark tasks from five OAEI tracks spanning biomedical to beyond-equivalence. The results show that ensemble fusion consistently improves the balance between precision and recall and frequently outperforms standalone aligners across diverse domains. Furthermore, our analysis reveals that ensemble composition directly affects the precision-recall trade-off: heterogeneous cross-paradigm ensembles generally improve precision, whereas homogeneous LLM ensembles more often achieve higher overall F1-scores. These findings demonstrate that systematic ensemble learning offers a robust and reproducible strategy for OA while providing practical guidance for selecting ensemble compositions under different alignment scenarios.
1. Introduction
Ontology alignment has progressed across multiple methodological paradigms, but unified frameworks still lack configurable mechanisms for reconciling heterogeneous and conflicting predictions. OntoAligner-Ensemble addresses this gap with voting-based fusion and post-fusion selection, evaluated across diverse OAEI tasks.
- Ontology alignment connects semantically corresponding elements across different ontologies to support interoperability between heterogeneous data and knowledge sources.
- Modern frameworks unify lexical, structural, embedding-based, and generative aligners, but do not by themselves specify how their complementary or conflicting outputs should be combined.
- LLM-based aligners have largely been studied individually, while prior OntoAligner work primarily considered open-weight models.
- OntoAligner-Ensemble combines heterogeneous aligner predictions through a unified, configurable mechanism using voting-based fusion and post-fusion selection.
- The evaluation compares five aligners and two ensemble configurations across eight tasks from five OAEI tracks spanning multiple ontology domains.
2. Problem Formalization
The paper formalizes ontology alignment as producing semantic correspondences between concepts from source and target ontologies, with confidence scores attached to predicted relations.
- Ontology alignment considers concepts drawn from a source ontology and a target ontology as the entities to be matched.
- Semantic relation r∈{≡, ⊑, ⊒} represents equivalence, subclass, or superclass correspondence between source and target concepts.
- The confidence score S_r∈[0, 1] indicates the likelihood assigned to a predicted correspondence.
3. Related Work
Prior OA research combines lexical, structural, semantic, learned, and ensemble-based evidence, but ensemble approaches remain relatively underexplored. OntoAligner-Ensemble builds on this literature by integrating multiple paradigms through configurable voting.
- Traditional OA aggregates lexical, structural, and semantic similarity measures, with effectiveness depending on the characteristics of the ontologies being aligned.
- Machine-learning approaches learn combinations of matching evidence, including meta-level frameworks that combine outputs from multiple ontology aligners.
- ROME uses Bagging and Boosting, while other work applies genetic programming, active meta-learning, Random Forest classification, or interactive ensemble strategies.
- Biomedical alignment has combined word-embedding similarities with existing aligners, showing that learned semantic features can complement traditional matching methods.
- Ensemble approaches remain relatively underexplored because combining multiple aligners can increase execution time and resource requirements.
- OntoAligner-Ensemble investigates a unified framework that integrates multiple aligners and leverages complementary strengths across diverse OA scenarios.
4. OntoAligner-Ensemble Framework
OntoAligner-Ensemble is an aligner-agnostic framework that combines heterogeneous ontology aligners through configurable voting-based fusion and post-fusion selection. Its modular design supports unified integration of diverse prediction formats, voting strategies, and selection policies.
- Framework Design: Multiple independent pipelines generate candidate correspondences that are converted into a unified representation before fusion.Duplicate source–target pairs are reduced to the highest-ranked occurrence according to confidence score.
- Framework Design: The framework integrates any OntoAligner-compatible aligner that produces source–target correspondence predictions.Aligners operate through the shared AlignerPipeline abstraction, including parsing, encoding, alignment, and optional post-processing.
- Voting-Based Fusion: Fusion aggregates candidates proposed by at least one aligner and supports weighted voting, RRF, Condorcet voting, Borda count, and score averaging.Candidates are ordered by decreasing fused score after applying the selected voting strategy and parameters.
- Post-Fusion Selection: Post-fusion selection converts the ranked candidate pool into a final alignment using configurable decision policies.The framework provides top-1 per source, top-k per source, threshold-based, and greedy bijective selection.
- Framework Design: This formulation adapts voting and selection policies to different datasets and application requirements while integrating heterogeneous OA systems.The framework is designed to remain independent of the underlying alignment algorithms.
5. Experimental Setup
The experiments evaluate OntoAligner-Ensemble across eight tasks from five OAEI tracks using five aligners spanning lightweight, KGE-based, and RAG-based approaches. Two ensemble configurations use equal-weight majority voting followed by top-1-per-source selection, with manually chosen hyperparameters.
- Datasets: Eight ontology-alignment tasks from five OAEI tracks cover domains including beyond-equivalence, circular economy, anatomy, material science, and biodiversity.The selected tasks include G1–Web, G2–Diseases, G3–Text, CEON–BiOnto, CEON–MatOnto, Mouse–Human, and MI–MatOnto, with one biodiversity task also selected.
- Aligners: Five aligners represent three categories: lightweight string matching, ConvE-based KGE, and three RAG configurations.All aligners use OntoAligner’s AlignerPipeline interface for unified execution and prediction formatting.
- Aligners: The RAG configurations pair Qwen, GPT, and Gemini generators with retrievers from their respective model ecosystems.The configurations include both open-weight and API-based LLM ecosystems.
- Experimental Protocol: Hyperparameters were selected manually from preliminary observations and prior experience without systematic optimization.The authors cite the large configuration space and computational cost as reasons for avoiding exhaustive or automated tuning.
- Ensemble Configurations: Ens. (All) combines all available aligners, whereas Ens. (LLMs) combines only the three RAG-based LLM aligners.These configurations compare heterogeneous cross-category integration with a homogeneous LLM ensemble.
- Ensemble Configurations: All ensemble experiments use equal aligner weights, majority voting, and top-1-source selection after fusion.The minimum valid vote is 3 for Ens. (All) and 2 for Ens. (LLMs).
6. Results and Discussion
Across eight OAEI tasks, individual aligners showed task-dependent precision–recall profiles, while voting ensembles often improved balance and performance. Ensemble composition shaped the trade-off: heterogeneous ensembles favored precision, whereas homogeneous LLM ensembles more often favored recall and higher F1.
- RQ1: Individual aligners: No individual aligner consistently achieved the best performance across all tasks, confirming that relative effectiveness was task-dependent.
- RQ1: Individual aligners: For six of eight tasks, all three RAG/LLM-based aligners exceeded the corresponding established baseline in recall.On Fish–Zooplankton, GPT-5.4-Nano achieved 93.3% recall versus 53.3% for LogMapLt.
- RQ1: Individual aligners: Lightweight and RAG/LLM-based aligners generally favored recall, while the KGE aligner generally favored precision.The fuzzy string aligner reached 93.1% recall on CEON–BiOnto and 100.0% on CEON–MatOnto, whereas ConvE reached 100.0% precision on Fish–Zooplankton and 93.3% on MI–MatOnto.
- RQ2: Voting-based ensemble: At least one ensemble achieved or tied for the highest F1-score on five of eight tasks, showing that voting can exploit complementary alignment evidence.
- RQ2: Voting-based ensemble: 96.5% F1 on Fish–Zooplankton was achieved by Ens. (LLMs), exceeding GPT-5.4-Nano at 93.3% and the LogMapLt baseline at 64.0%.
- RQ2: Voting-based ensemble: Ensemble gains were not uniform: established baselines remained strongest on CEON–MatOnto and Mouse–Human, while Qwen led MI–MatOnto with 42.2% F1.Ens. (LLMs) reached 41.6% on MI–MatOnto, while Ens. (All) reached 36.3%; the LogMap baseline reached 52.5% on CEON–MatOnto and Matcha reached 94.1% on Mouse–Human.
- RQ3: Ensemble composition: Ens. (All) generally favored precision, whereas Ens. (LLMs) more often preserved recall and achieved higher F1 on six of eight tasks.Ens. (All) combines lightweight, KGE-based, and RAG/LLM aligners; Ens. (LLMs) combines the three RAG/LLM aligners.
- RQ3: Ensemble composition: Ensemble composition varied by domain: Ens. (All) reached 84.3% precision but 23.1% recall on MI–MatOnto, while Ens. (LLMs) achieved 41.6% F1.
7. Discussion and Conclusion
The paper positions ensemble alignment as useful for reliable semantic interoperability and concludes that voting fusion can improve precision–recall balance across diverse OA settings. It also identifies relation-type classification as an unresolved framework limitation.
- Reliable semantic interoperability supports regulatory reporting, reference data integration, and enterprise analytics involving heterogeneous ontologies.
- The framework reduces equivalence, subclass, superclass, and disjointness decisions to binary match versus no-match outputs.
- Distinguishing specific semantic relations remains a key future direction for OntoAligner.
- Across eight benchmark tasks from five OAEI tracks, voting fusion consistently improved precision–recall balance and frequently outperformed individual baselines.
Declaration on Generative AI
The manuscript reports limited use of generative AI tools for presentation and experiment code development. The authors state that scientific content and research decisions were developed and verified by the authors.
- ChatGPT was used for grammar, spelling, readability checks, and presentation improvements, with all suggestions reviewed and adapted by the authors.
- The authors state that scientific content, research design, analysis, and conclusions were developed and verified without AI involvement.
- Generative AI tools were also used solely to assist with code development for the reported experiments.