Discovering Subsumption Relationships in Ontology Alignment

27 concepts19 papers

When we try to merge two different databases, knowledge graphs, or ontologies, our first instinct is to look for identical concepts. If Ontology A contains Laptop and Ontology B contains Notebook, an alignment system should ideally recognize them as equivalent ($Laptop \equiv Notebook$).

For years, automated ontology alignment focused almost exclusively on this type of horizontal, flat equivalence matching. However, real-world knowledge systems are rarely designed at the exact same level of granularity. As Xi Chen 0003 observes in Variations on aligning linked open data ontologies (om 2013), exact concept matches are remarkably rare across web-scale datasets like Linked Open Data. If Ontology A defines Laptop but Ontology B only contains the broader category Computer, a traditional equivalence matcher will fail entirely.

To truly integrate heterogeneous data, systems must move beyond simple synonyms and discover subsumption relationships—directional, vertical "is-a" links where one concept is a specialized subtype of another ($Laptop \sqsubseteq Computer$). Early alignment frameworks, such as Falcon-AO developed by Wei Hu et al. in The Results of Falcon-AO in the OAEI 2006 Campaign (Ontology Matching 2006), demonstrated that ignoring these hierarchical inclusions severely limits the ability of downstream applications to perform deep semantic reasoning. Similarly, in decentralized architectures, Yong Li et al. in A Peer-to-Peer Approach to Semantic Web Services Discovery (International Conference on Computational Science (4) 2006) showed that capturing partial matches via subsumption is vital for finding functionally compatible services described at different levels of specificity.

Why Subsumption is a Harder Problem

Discovering subsumption is fundamentally more challenging than finding equivalences. Equivalence is symmetric and often highly lexical; if two concepts share a name or a close embedding vector, they are likely equivalent. Subsumption, however, is asymmetric and frequently lexically disjoint.

Consider the concepts Viral Pneumonia and Lung Disease. A machine learning model relying purely on string similarity or surface-level text processing will find little overlap. Furthermore, as noted in the Results of the Ontology Alignment Evaluation Initiative 2022 (OM@ISWC 2022) by Mina Abd Nikooie Pour et al., even state-of-the-art machine learning models that excel at matching identical entities experience a sharp decline in accuracy when tasked with discovering vertical semantic links.

The breakdown occurs because traditional logic programming and unification paradigms struggle to reason about hierarchical type constraints without explicit structural context, a limitation early highlighted by Chris Mellish et al. in Using Classification as a Programming Language (ijcai 1993). To resolve an asymmetric link, a system cannot just look at the entities themselves; it must evaluate their positions within their respective taxonomic hierarchies.

Algorithmic Strategies for Hierarchical Alignment

To bridge this gap, researchers have developed mechanisms that combine machine learning, formal logic, and language models to infer directional links.

Classification-Based Learning

One prominent strategy is to frame subsumption discovery as a binary classification problem. Instead of relying on hand-crafted rules, the CSR (Classification-Based Learning of Subsumption Relations) framework, introduced by Vassilis Spiliopoulos et al. in Learning Subsumption Relations with CSR: a Classification based Method for the Alignment of Ontologies (OM 2007) and expanded in CSR: Discovering Subsumption Relations for the Alignment of Ontologies (ESWC 2008), trains machine learning models on the internal taxonomic structures of the ontologies. By learning how concepts are clustered and structured internally, CSR can predict whether an unmapped pair across two ontologies exhibits an "is-a" relationship, uncovering complex links that formal logic reasoners miss when initial equivalence mappings are absent.

Description Logic and Hybrid Reasoning

Rather than bypassing formal logic, another approach uses it as a scaffolding. Quentin Reul developed KOSIMap (KOSIMap: Use of Description Logic Reasoning to Align Heterogeneous Ontologies (Description Logics 2010); KOSIMap: Ontology Alignments Results for OAEI 2009 (OM 2009)), demonstrating that utilizing inferred logical hierarchies through description logic reasoning significantly enhances alignment accuracy. By extracting implicit background knowledge via reasoners, KOSIMap uncovers hidden relationships that are not explicitly defined in the asserted structures, while simultaneously pruning logically inconsistent matches.

Modern frameworks often create a feedback loop between statistical embeddings and logical constraints. For example, the PRASE framework, proposed by Zhiyuan Qi et al. in Unsupervised Knowledge Graph Alignment by Probabilistic Reasoning and Semantic Embedding (ijcai 2021), pairs probabilistic reasoning with semantic embeddings. This allows the system to use the contextual power of embeddings to suggest initial alignments, which are then refined and corrected by integrating higher-level class hierarchies and ontologies during the reasoning process.

+-----------------------+     Suggested Alignments     +----------------------------+
|  Semantic Embeddings  | ---------------------------> |   Probabilistic Reasoner   |
| (Contextual Clues)    | <--------------------------- | (Enforces Class Hierarchy) |
+-----------------------+      Refined Constraints     +----------------------------+

To make these hybrid deep-learning and logic workflows accessible, Yuan He et al. introduced DeepOnto: A Python Package for Ontology Engineering with Deep Learning (arxiv 2023), providing tools for language model probing to identify non-obvious subsumption relations that traditional string matching misses. When automated logic or embeddings remain highly uncertain, systems like LogMapLLM, introduced by Sviatoslav Lushnei et al. in Large Language Models as Oracles for Ontology Alignment (arxiv 2025), deploy Large Language Models as targeted "oracles" to resolve ambiguous, complex hierarchical mappings without incurring the massive computational cost of an LLM-only alignment pipeline.

Overcoming the Evaluation Gap

For a long time, progress in subsumption discovery was bottlenecked by how alignments were evaluated. Traditional metrics like standard precision and recall operate on simple set-to-set comparisons, treating alignments as flat pairs and ignoring deductive consequences. As Daniel Fleischhacker et al. argued in Implementing Semantic Precision and Recall (OM 2009), evaluating systems based on their true deductive impact requires a fundamental shift toward subsumption correspondences.

Furthermore, early benchmarking campaigns lacked the data to track this. The Results of the Ontology Alignment Evaluation Initiative 2009 (OM 2009) by Jérôme Euzenat et al. noted that the evaluation of hierarchical relationships was severely immature due to a lack of refined reference alignments for non-equivalence mappings.

To solve this evaluation bottleneck, Yuan He et al. constructed Bio-ML (Machine Learning-Friendly Biomedical Datasets for Equivalence and Subsumption Ontology Matching (arxiv 2022)). Bio-ML provides curated train/test splits and ranking-based metrics specifically tailored for both equivalence and subsumption matching in complex biomedical domains. By establishing reference datasets where models are explicitly penalized if they fail the logical inference required for parent-child classification, this framework allows modern machine learning and logic-based systems to systematically improve how they discover the vertical structures that bind disparate ontologies together.

Key Foundational Work

The Results of Falcon-AO in the OAEI 2006 Campaign (2006) — contributed 2 concept elements to this direction.
A Peer-to-Peer Approach to Semantic Web Services Discovery (2006) — contributed 1 concept elements to this direction.
Using Classification as a Programming Language (1993) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 17 papers
  2. Version 1 · 2026-10-04 · 27 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_subsumption_relationship_discovery_2f388fbc,
  title = {subsumption relationship discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subsumption-relationship-discovery/rdr_2f388fbce7b344d48d4bf79ccdf431e7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_subsumption_relationship_discovery_2f388fbc_v2,
  title = {subsumption relationship discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subsumption-relationship-discovery/rdr_2f388fbce7b344d48d4bf79ccdf431e7/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake