CMOMgen: Complex Multi-Ontology Matching

• CMOMgen is a neurosymbolic framework that generates semantically sound OWL mappings for complex multi-ontology alignment.
• It combines retrieval-based class selection, pattern-guided in-context learning, and semantic consistency checks to ensure high-fidelity mappings.
• The approach leverages conceptual component extraction to reduce search space and improve scalability in large biomedical ontological ecosystems.

Complex Multi-Ontology Matching with CMOMgen encompasses a paradigm shift in ontology alignment, extending the classical 1:1 entity mapping to complex, compositionally expressive mappings between one source concept and composed logical constructions over arbitrarily many target ontologies. CMOMgen leverages neurosymbolic techniques—Retrieval-Augmented Generation (RAG), pattern-guided in-context learning, and semantic consistency verification—to automate the generation of semantically sound OWL mappings at scale, with robust empirical validation in large biomedical ontological ecosystems (Silva et al., 24 Oct 2025). The incorporation of conceptual component (CC) extraction further enables scalable, search-space-reducing constraints that are crucial for high-fidelity multi-ontology integration (Asprino et al., 2021).

1. Formal Definitions and Problem Scope

Complex Multi-Ontology Matching (CMOM) seeks mappings of the form

$M: S \to L(T_1,\ldots,T_n)$

where $S$ is a set of source concepts, $\mathcal{T} = \{T_1,\ldots,T_n\}$ is a collection of $n \geq 2$ target ontologies, $\bigcup_i T_i$ is the union of target ontology concepts, and $L(T_1,\ldots,T_n)$ is the set of logical OWL class expressions constructible via standard constructors ($\sqcap, \sqcup, \exists, \forall, \ldots$). The mapping $M(s) = E$, for $s \in S$, yields an expression $E$ that is semantically equivalent to $S$0 in OWL, i.e., $S$1.

Each CMOM mapping is a tuple $S$2, with $S$3 an optional confidence score. This formulation generalizes both pairwise and 1:n mapping, enabling representational richness needed for semantic integration of heterogeneous ontologies (Silva et al., 24 Oct 2025).

2. Core CMOMgen Architecture and Workflow

CMOMgen operationalizes complex multi-ontology alignment via an end-to-end pipeline with three principal modules:

• Class Selection (Retrieval): Implements a RAG methodology, combining recursive lexical coverage and embedding-based matching. Lexical candidate generation recursively covers the source label with non-overlapping target labels; the embedding-based component iteratively selects target concepts maximizing cosine similarity, updating the source embedding by subtraction until a threshold is reached. Each candidate set is scored using multiplicative weights (lexical) or embedding similarity (cosine; see Equations 4.2 and 4.5 in (Silva et al., 24 Oct 2025)).
• Pattern-Guided In-Context Learning (Composition): For each candidate target set, the namespace–cardinality pattern (e.g., UBERON:2, PATO:1) is extracted. A filtered pool of reference logical definitions is retrieved, containing only samples with compatible patterns. The prompt for a LLM such as GPT-4O-mini is constructed by injecting the selected classes and pattern-matched example mappings into a templated system and user message (see Section 4.2.3 in (Silva et al., 24 Oct 2025)).
• Semantic Filtering and Consistency Checking: Generated OWL expressions are parsed and checked for (i) OWL well-formedness, (ii) restriction to allowed properties (e.g. RO, BFO), and (iii) lightweight logical consistency. Only mappings passing these filters are retained.

The entire pipeline supports unrestricted-arity (n-ontology) matching and integrates retrieval, composition, and validation within a unified framework (Silva et al., 24 Oct 2025).

3. Mathematical Formulations, Algorithms, and Pattern Design

Key Formulas (from (Silva et al., 24 Oct 2025), Section 4.3):

• Lexical Candidates:

$S$4

• Embedding Candidates:

$S$5

• Candidate Set Scoring:

$S$6

Pattern-Guided Prompting

Prompt construction injects pattern-matched few-shot OWL examples alongside the class set resulting from the prior retrieval stage, enforcing compositional and syntactic pattern fidelity for the LLM.

Pseudocode Summary

$S$8 (Silva et al., 24 Oct 2025)

4. Integration of Conceptual Components and Scalability

Extraction of common conceptual components (CCs) from ontologies provides a scalable basis for constraining candidate mappings in the CMOMgen pipeline. Ontologies are transformed into "intensional graphs"—undirected and unlabeled—through a systematic process involving class, property, and axiom extraction (Asprino et al., 2021). Clauset-Newman-Moore community detection algorithm identifies dense communities (OODPs) whose feature vectors are then clustered into CCs via K-means on enriched semantic and lexical representations.

Matching is seeded and pruned by CC clusters. Candidate entity pairs are restricted to those where classes/properties belong to identical or highly related CCs (hierarchical-strength threshold $S$7), reducing the candidate mapping space by over 90% in large domains. Hybrid strategies (lexical, structural, instance-based, logical validation) are subsequently applied within each CC as domain-appropriate (Asprino et al., 2021).

This CC framework enables efficient multi-ontology matching by aligning first at the component (CC) level and subsequently at the entity level, thus improving both computational feasibility and semantic plausibility of the resulting alignments.

5. Evaluation Methodologies and Empirical Results

Automatic Evaluation

Automatic assessment relies on two principal metrics:

• Relaxed Class-based Precision/Recall: Incorporates ancestor/descendant matches, rewarding correct but more general/specific mappings by partial credit (see 4.4.1, (Silva et al., 24 Oct 2025)).
• Graph-Edit-Distance (GED)-based Scores: Mappings are compared as OWL-derived graphs, assigning edit costs to node substitutions, insertions, deletions, and edge modifications, with the final mapping score normalized against maximal GED (see 4.4.1, (Silva et al., 24 Oct 2025)).

Quantitative Results

Task	Method	Prec	Rec	F₁
HP	CMOM baseline	0.211	0.211	0.211
	LM baseline	0.210	0.203	0.207
	w/o examples	0.275	0.255	0.265
	w/o classes	0.617	0.617	0.617
	CMOMgen (full)	0.634	0.632	0.633
MP	CMOM baseline	0.254	0.254	0.254
	LM baseline	0.203	0.193	0.198
	w/o examples	0.287	0.270	0.278
	w/o classes	0.645	0.642	0.643
	CMOMgen (full)	0.666	0.662	0.664
WBP	CMOM baseline	0.217	0.217	0.217
	LM baseline	0.206	0.197	0.201
	w/o examples	0.260	0.242	0.250
	w/o classes	0.731	0.730	0.731
	CMOMgen (full)	0.687	0.687	0.687

CMOMgen outperforms all baselines and ablations, achieving 0.63–0.69 in F1 across evaluation sets, demonstrating the benefit of both dedicated class retrieval and in-context example injection (Silva et al., 24 Oct 2025).

Manual Evaluation

Expert review of highest-confidence mappings assigned scores (1–5 for logical fidelity): 46% achieved exact matches (score 5), 24% near-exact (score 4), and only 30% scored ≤3. Average expert rating was 3.8/5, confirming generation of semantically strong OWL alignments (Silva et al., 24 Oct 2025).

6. Technical Considerations, Scalability, and Limitations

Key technical design choices include:

• LLM Infrastructure: GPT-4O-mini is used for OWL snippet synthesis, provided with explicit OWL system messages and pattern-matched in-context examples (Silva et al., 24 Oct 2025).
• Candidate Selection and Prompt Engineering: The interplay between class selection and example retrieval is critical. Ablations reveal degradation of class-based F1 from 0.63 (full system) to 0.21/0.20 (baselines).
• Search Space Pruning via CCs: Integrating conceptual component clustering at the pipeline front reduces the number of candidate pairs by over 90% in some benchmarks, scaling alignment to corpora containing hundreds of ontologies and tens of thousands of entities (Asprino et al., 2021).

Identified limitations include loss of class-expression information in initial graph abstraction, reliance on English or well-labeled entities, and dependency on reference examples for effective in-context prompting. Proposed improvements are to extend pattern extraction to union/intersection OWL constructs and to adopt nonparametric or hierarchical clustering for CC identification.

7. Implications and Future Directions

CMOMgen demonstrates that neurosymbolic approaches, unifying LLM synthesis with pattern-guided retrieval and ontology structural constraints, provide robust solutions for scalable, semantically rich multi-ontology alignment (Silva et al., 24 Oct 2025). Integrating conceptual component extraction further advances scalability and interpretability (Asprino et al., 2021). A plausible implication is the increasing automation of complex logical definition generation, reducing the burden on domain experts for large-scale ontology engineering.

Current limitations regarding LLM inference cost, ontology language diversity, and fine-grained logical expressiveness suggest fruitful areas for research: adaptive CC clustering, multilingual or cross-lingual support, and tighter integration of reasoning with neural model prompting. Integration with Ontology Design Pattern (ODP) catalogues and live updates also present promising avenues for future system extensions (Asprino et al., 2021).