This paper presents an enhancement of OWL2Vec4OA for ontology alignment, focusing on regression-based local matching. The enhancement integrates Sentence-BERT (SBERT) and Word2Vec embeddings, each combined with lexical and URI-based features. Diferent fusion strategies-concatenation, averaging, and merging-are explored to create richer embedding representations. The regression model leverages these hybrid embeddings to improve similarity prediction for local ontology matching. Regression models using these hybrid embeddings are Ontology alignment is essential for enabling interoperability across heterogeneous knowledge systems. Traditional approaches such as LogMap [1] and AML [2] have proven efective but struggle when the ontologies lack enough lexical information.
Our ontology alignment framework enhances OWL2Vec4OA embeddings for local ontology matching. It consists of two main steps: generating hybrid embeddings from SBERT and Word2Vec across lexical, URI, and combined channels, and predicting similarity scores between entity pairs using a regression model. This approach captures both semantic and structural information from the ontologies.
ments and evaluating local and global matching.
(ISWC 2025), November 2nd, 2025, Nara, Japan https://github.com/Sevinjt/OWL2Vec4OA/ (S. Teymurova) CEUR Workshop
ISSN1613-0073
OWL2Vec4OA (depicted in Figure 1) extends OWL2Vec* [12] by generating embeddings that capture both structural and textual semantics tailored to the ontology alignment task. Input ontologies are converted into RDF graphs, with simple axioms mapped directly and complex axioms transformed appropriately. Seed alignments from LogMap and AML identify high-confidence (“good”) and rejected (“bad”) mappings, which provide both positive and negative training data. The individual graphs are merged into a weighted graph, where edges reflect source confidence. Biased random walks on this graph produce sequences that train a Word2Vec model, resulting in embeddings that serve as features for machine learning models to predict semantic similarity between ontology entities.
2.2.1. Benchmark Datasets and Alignment Generation In this study, we utilise multiple benchmark datasets from the Ontology Alignment Evaluation Initiative (OAEI)1 The bio-ml benchmark consists of biomedical ontologies that are widely used in the biological and medical domains. These datasets provide diverse complexity levels and domain specificity, enabling comprehensive evaluation of our methodology. 2.2.2. Alignment Generation Methodology We first obtain alignments between the ontologies using two established matching systems: LogMap [ 1] and AML [2].
From these systems, we generate only four distinct sets of alignments which we use in OWL2Vec4Oa:
1. LogMap overestimation ( ): Comprises all correspondences identified by LogMap before
applying its mapping repair techniques and logical consistency checks.
2. Intersection of AML and LogMap mappings ( ): Contains only those correspondences that
are identified by both AML and LogMap, representing high-confidence alignments that are likely
to be correct:
=
∩
= {( 1, 2, ) ∣ ( 1, 2, 1) ∈
∧ ( 1, 2, 2) ∈ }
(
3. LogMap output ( ): Contains the final set of correspondences produced by LogMap after applying its repair techniques and logical consistency checks. 4. Union of AML and LogMap mappings ( ): Combines all correspondences identified by either
=
∪
= {( 1, 2, ) ∣ ( 1, 2, ) ∈
∨ ( 1, 2, ) ∈ }
(
2.2.3. Negative Sample Generation
• LogMap hard discards: Correspondences explicitly rejected by LogMap’s lexical and structural iflters. • LogMap discards: Correspondences initially considered but later rejected during the mapping process. • Conflicted mappings : Correspondences that cause logical inconsistencies in the aligned ontology.
imbalance is addressed via weighted loss functions, introducing non-determinism and reducing reliance on exhaustive negative sets. The regression loss is adjusted/lowered to stabilise training while balancing contributions from positive and negative samples. The training dataset includes approximately 1,000 to 44,000 negative samples, depending on the specific ontology pair. 2.2.4. Embedding Generation using enhanced OWL2Vec4OA In the second phase of our preprocessing pipeline, we utilise enhanced OWL2Vec4OA to generate vector representations of the ontology entities and their alignments. For each entity, embeddings are generated using SBERT and Word2Vec across three textual channels: lexical labels, URI tokens, and a combination of both. These embeddings are integrated using simple fusion strategies, such as concatenation, averaging or merging, to produce hybrid vectors that capture complementary semantic and structural information. Optionally, dimensionality reduction is applied to improve computational eficiency without sacrificing performance.
2. The four alignment sets described above ( , , , and ).
The OWL2Vec4OA framework generates embeddings by: 1. Converting the OWL ontologies into RDF graphs. 2. Applying random walks on the RDF graphs to generate sequences of entities and relationships.
a SBERT models on these sequences.
2.3.1. Input Ontologies and Preprocessing Input ontologies 1 and 2 are projected into RDF graphs using OWL2Vec4OA’s axiom transformation rules. Seed mappings from LogMap and AML provide high-confidence positive alignments. Each entity’s URI, label, and annotation are extracted to form textual corpora for embedding generation.
2.3.2. Multi-Channel Embedding Generation For each entity, we generate three textual representations: • Lexical (lexc): tokenized rdfs:label and annotation strings. • URI: tokenized local names extracted from entity IRIs. • Mixed (lexc + URI): combined lexical and URI tokens.
• Word2Vec - trained on the ontology corpus.
• SBERT - a transformer-based contextual embedding model.
2.3.3. Fusion and Dimensionality Reduction To integrate multiple channels, embeddings are fused using: v v = 13 (v = [v ; v
; v ]
+ v
+ v )
(
2.3.4. Regression-Based Similarity Score Prediction Using the embeddings generated by enhanced OWL2Vec4OA, we train supervised machine learning models to improve LogMap’s output alignments. Our objective is to refine LogMap’s results by identifying false positives and recovering false negatives. Hybrid embeddings serve as input to a Siamese regression model that predicts continuous similarity scores between candidate entities. The architecture is designed to model semantic relatedness eficiently while addressing class imbalance via sample weighting. Training uses early stopping to prevent overfitting and ensures robust similarity estimation. The resulting similarity scores are then used for both local ranking of candidates and global alignment selection.
• Positive samples: Alignments from LogMap’s output ( ), which are assumed to be correct but potentially incomplete. • Negative samples: A combination of LogMap hard discards, LogMap discards, and conflicted mappings, which represent entity pairs that should not be aligned.
Class imbalance is addressed through weighting:
=
∑
=1
⋅
(
ifcation model ) and predict the scores between alignments (regression model ) We employ two Siamese
Architecture Details:
To model semantic similarity between biomedical concepts, we leveraged a pre-trained model (OWL2Vec4OA*) to convert ontology terms into dense vector embeddings. Positive pairs were scaled into the similarity interval [0.6, 1.0], while negative pairs were scaled into [0.0, 0.6] using min-max normalization. Entities lacking pre-trained embeddings were discarded to maintain consistency in feature representation.
• Input: 256-dimensional embeddings (from fused SBERT/Word2Vec channels) • Encoder: three Conv1D layers with tanh activation, batch normalisation, and max-pooling • Output: a single regression head producing similarity in [0, 1] • Loss: mean absolute error (MAE) between predicted and true similarity • Optimisation: Lion optimiser, learning rate 0.001, batch size 1024, early stopping (patience = 5) • Sample weights derived from class frequency to address imbalance • PCA-compressed embeddings (2–96 dimensions) are used for robustness testing
Entity similarity is measured using standard vector-based metrics, which are combined to rank candidate target entities for each source entity. Alignments are determined by applying a similarity threshold to regression predictions, retaining only high-confidence pairs. This approach allows the system to prioritize precise matches while maintaining overall coverage.
Similarity between ontology entities is computed using multiple metrics implemented in our code:
(
cos(v , v ) sim sim (, ) = (, ) = 1 1 1 + ‖v − v ‖2 1 + ‖v − v ‖1
Evaluation Metrics: • Local Matching: evaluating the ranking of candidate target entities for each source entity. • Global Matching: selecting high-confidence alignments across all entity pairs. • MRR: measures the rank of the correct entity for each source.
• Hits@K: percentage of queries where the correct entity is within the top K candidates.
= {( 1, 2) ∣ ( 1, 2) ≥ }
The study retained only entity pairs surpassing a defined similarity threshold in the final ontology alignment. Using two Siamese neural network architectures with OWL2Vec4OA embeddings—which integrate textual and structural ontology information—the model efectively captured rich semantic relationships between concepts. Training ran for up to 250 epochs with early stopping and balanced sampling, preserving the best-performing checkpoints. The framework used dual Conv1D towers and a similarity regression branch to estimate concept relatedness, keeping pairs with similarity scores above 0.60. All models, configurations, and results were fully documented and exported for reproducibility.
The regression model utilises an identical Siamese architecture with an additive attention mechanism for enhancing semantic similarity representation: ( q, k, v) = ∑
exp( ( ∑ exp( ( q, ki)) q, kj)) i v
= {( 1, 2) ∣ ( 1, 2) = 1 ∧ ( 1, 2) ≥ } where is a confidence threshold.
This section presents the consolidated evaluation results of the enhanced OWL2Vec4OA and retrieval experiments conducted across four datasets: NCIT–DOID, OMIM–ORDO, SNOMED–Neoplas(NCIT), and SNOMED–Neoplas (Pharm). Each experiment compared two embedding strategies—SentenceBERT (SBERT) and Word2Vec—while employing multiple similarity metrics (Cosine, Euclidean,
similarity metrics and embedding dimensions were part of the evaluation pipeline, they are omitted from the table below for clarity and are discussed only within this descriptive text.
terms of Mean Reciprocal Rank (MRR) and Hits@ (for = 1, 5, 10 ). The highest scores were achieved by SBERT in combination with cosine similarity, highlighting that contextualised sentence-level embeddings capture richer semantics than static word vectors. 0.879 84.15 92.53 95.27 0.707 63.45 78.62 83.80
Overall, The results show that embedding type and dimensionality greatly influence ontology alignment accuracy, with particularly strong gains in the SNOMED–Neoplas (NCIT). The largest improvements are observed in Hits@1 and MRR, indicating that SBERT is substantially more efective at ranking correct matches at the top of the retrieval list. While Word2Vec occasionally achieves slightly higher scores at broader cutofs (Hits@5 or Hits@10), SBERT remains the dominant model, demonstrating superior semantic representation quality for ontology alignment tasks.
The regression-based similarity prediction model was re-evaluated using the expanded BIO-ML datasets, incorporating full query coverage across all four ontology mapping tasks: NCIT–DOID, OMIM–ORDO, SNOMED–Neoplas (NCIT), and SNOMED–Neoplas (Pharm). The enlarged datasets improved overall robustness, yielding consistent gains in ranking quality.
Table 2 summarizes the best-performing configurations. The NCIT–DOID task produced the strongest results, achieving an MRR of 0.8795 and Hits@1 of 84.2%, slightly improving over the previous maximum. OMIM–ORDO also benefited from the extended dataset, reaching MRR 0.7079 and Hits@1 63.6%, with notable improvements at higher retrieval depths.
For SNOMED–Neoplas (NCIT), the model attained an MRR of 0.8286 and Hits@1 of 77.2%, reflecting stronger neoplasm alignment performance. The SNOMED–Neoplas (Pharm) configuration achieved the highest pharmacological alignment accuracy, with MRR 0.8647 and Hits@1 79.7%, and strong precision across broader cutofs.
NCIT–DOID 0.8795 OMIM–ORDO 0.7079 SNOMED–Neoplas (NCIT) 0.8286 SNOMED–Neoplas (Pharm) 0.8647 84.2 63.6 77.2 79.7
Combining SBERT and Word2Vec embeddings with regression-based similarity scoring continues to yield strong results for ontology alignment. The updated experiments demonstrate consistent improvements across all BIO-ML tasks, with notable gains in MRR and Hits@1 for NCIT–DOID,
During the preparation of this work, the authors used Grammarly in order to grammar and spell check, and improve the text readability. After using the tool, the authors reviewed and edited the content as needed to take full responsibility for the publication’s content.