vaibhavalakshmiravideshik/mesh-snomed-entity-alignment-15k

MeSH-SNOMED Entity Alignment 15K

MeSH-SNOMED Entity Alignment 15K is a biomedical heterogeneous knowledge graph alignment benchmark for cross-ontology matching between MeSH and SNOMED CT. It is designed to evaluate entity alignment systems under realistic large-graph conditions, where gold-aligned concepts are embedded in much larger biomedical graphs containing many structurally relevant but non-aligned background entities. This release is intended for the accompanying EMNLP 2026 submission.

The benchmark starts from an existing set of MeSH-SNOMED matches and packages those alignments together with graph context extracted from the two source ontologies. The purpose of the dataset is not to newly adjudicate the semantic validity of every original match, but to support research on whether alignment systems can recover semantically corresponding biomedical entities across large, heterogeneous graph spaces.

Overview

MeSH and SNOMED CT are both major biomedical resources, but they were created for different primary purposes and exhibit different ontology engineering choices. MeSH, maintained by the U.S. National Library of Medicine, is a controlled and hierarchically organized vocabulary used extensively for literature indexing, cataloging, and retrieval, especially in MEDLINE/PubMed and related NLM systems. SNOMED CT, maintained by SNOMED International, is a very large clinical terminology designed for representing clinical meaning in health records and related health information systems. Aligning these resources is therefore valuable because it connects literature-oriented indexing structure with clinically oriented semantic structure.

This is precisely why the benchmark is heterogeneous. The two sides differ in relation inventories, attribute styles, graph density, concept granularity, and local topology. A model cannot succeed here by assuming the same schema or the same structural semantics on both sides. Instead, it must reason across ontology mismatch, lexical mismatch, attribute mismatch, and graph asymmetry. That makes this benchmark relevant not only for biomedical entity alignment, but also for broader research on heterogeneous KG alignment, ontology matching, representation learning, terminology interoperability, and semantic harmonization across biomedical systems.

Resource Links

The official MeSH home page is available at NLM MeSH, and the linked-data representation is available through MeSH RDF, with additional documentation at MeSH RDF documentation and downloadable RDF resources at NLM MeSH RDF downloads. The official SNOMED CT overview is available at What is SNOMED CT?, with broader organizational and documentation access through SNOMED International, NLM’s SNOMED CT overview, and the SNOMED documentation portal. The dataset repository itself is hosted at vaibhavalakshmiravideshik/mesh-snomed-entity-alignment-15k, while viewer-oriented helper files are kept separately under viewer/links/, viewer/attributes/, and viewer/relations/. Image assets used in the card are stored in assets/.

Benchmark at a Glance

Property	Value
Task	Biomedical entity alignment
Setting	Heterogeneous cross-ontology / cross-KG matching
Source ontologies	MeSH, SNOMED CT
Gold alignment file	ent_links
Gold alignment size	15,000 pairs
URI-expanded alignment file	ent_links_uri
MeSH attribute triples	11,385,523
SNOMED CT attribute triples	1,391,104
MeSH relation triples	6,948,511
SNOMED CT relation triples	1,331,550
Primary language	English
Intended venue	EMNLP 2026

Repository Structure

The canonical benchmark release at repository root consists of exactly six files: ent_links, ent_links_uri, attr_triples_1, attr_triples_2, rel_triples_1, and rel_triples_2. These are the authoritative files for experimentation and reproducible use. In addition, the repository contains a viewer/ subdirectory holding small TSV-based preview files used only to support the Hugging Face Dataset Viewer. Those preview files are not a replacement for the canonical benchmark release; they exist so that the dataset page remains browsable and interpretable through the Hub interface.

The alignments viewer configuration corresponds to the tabular preview of the alignment files, the attributes configuration corresponds to small previews of attribute triples, and the relations configuration corresponds to small previews of relation triples. This separation is necessary because the three data families have different schemas and should not be cast into a single tabular representation.

Included Files and Format

The file ent_links contains the gold MeSH-SNOMED alignment pairs, one per line, in a two-column tab-separated format. The file ent_links_uri contains a URI-based expanded alignment resource over the same ontology pair. The files attr_triples_1 and attr_triples_2 store attribute triples, where each line has the form head, predicate, attribute_value. The files rel_triples_1 and rel_triples_2 store relation triples, where each line has the form head, relation, tail.

A representative alignment example is:

<http://id.nlm.nih.gov/mesh/C537369>	<http://snomed.info/id/440350001>

A representative attribute triple example is:

<http://id.nlm.nih.gov/mesh/A01.111>	<http://www.w3.org/2000/01/rdf-schema#label>	A01.111@en

A representative relation triple example is:

<http://id.nlm.nih.gov/mesh/A01.111>	<http://id.nlm.nih.gov/mesh/vocab#parentTreeNumber>	<http://id.nlm.nih.gov/mesh/A01>

Dataset Statistics

File	Lines
ent_links	15,000
ent_links_uri	46,823
attr_triples_1	11,385,523
attr_triples_2	1,391,104
rel_triples_1	6,948,511
rel_triples_2	1,331,550

The graph validation analysis indicates that the benchmark is both large and structurally realistic. The ground-truth side contains 13,219 MeSH entities and 14,866 SNOMED CT entities. On the graph side, rel_triples_1 contains 2,456,953 entities and rel_triples_2 contains 379,282 entities. Of the gold entities, 13,191 of the 13,219 MeSH entities were found in KG1 relations, while 14,452 of the 14,866 SNOMED entities were found in KG2 relations. Only 28 MeSH gold entities were missing from KG1 relations, and 414 SNOMED gold entities were missing from KG2 relations.

Validation of KG Construction

The central dataset-construction question for this release is whether KG1 (MeSH) and KG2 (SNOMED CT) were created correctly as entity-alignment graphs given the existing ground-truth matches. Based on the supplied validation analysis, the answer is yes. The checks focused on whether gold entities appear in graph context, whether neighborhoods look realistically open rather than artificially closed, and whether missing entities are better explained by source sparsity than by benchmark assembly errors.

One of the most important results is that the neighborhoods of gold entities are dominated by non-gold background entities rather than by other gold entities. In a sample of 500 MeSH gold entities, only 533 of 9,745 observed neighbors were themselves gold entities, while 9,212 were non-gold background entities. This is not a flaw; it is one of the properties that makes the benchmark realistic. Real ontology alignment does not happen in a closed world consisting only of aligned nodes. Instead, candidate entities are embedded in much larger environments full of semantically related but non-equivalent distractors.

A small number of gold entities are isolated in the released graph representation. For MeSH, 28 gold entities were missing from rel_triples_1, and those same 28 were also absent from attr_triples_1. This suggests true source-data sparsity rather than a graph-construction mistake. Such cases are important to acknowledge, particularly for graph-based or embedding-based evaluation.

How to Use the Dataset

This repository should be treated primarily as a raw benchmark file release rather than as a standard row-based Hugging Face dataset with a single tabular schema. The canonical files are intended to be downloaded and parsed directly. The viewer/ files exist only to make browsing possible through the Dataset Viewer and should not be treated as the main release artifacts.

A typical usage pattern is to fetch the canonical files directly from the Hub and then parse them as tab-separated benchmark files:

from huggingface_hub import hf_hub_download

repo_id = "vaibhavalakshmiravideshik/mesh-snomed-entity-alignment-15k"
repo_type = "dataset"

ent_links_path = hf_hub_download(repo_id=repo_id, repo_type=repo_type, filename="ent_links")
rel1_path = hf_hub_download(repo_id=repo_id, repo_type=repo_type, filename="rel_triples_1")
attr1_path = hf_hub_download(repo_id=repo_id, repo_type=repo_type, filename="attr_triples_1")

The dataset is not currently packaged as a single load_dataset(..., split=...)-style tabular corpus because the canonical benchmark is naturally composed of distinct file families with different semantics and schemas. The manual viewer configurations included in the README are therefore a presentation aid rather than a replacement for direct benchmark use.

Benchmark Interpretation and Reporting

This benchmark should be interpreted as a needle-in-a-haystack heterogeneous biomedical KG alignment problem. The gold pairs are the evaluation targets, while the surrounding ontology graphs provide lexical and structural evidence. The two graph sides differ in schema, attribute style, and topology, and most nearby entities are not themselves gold-aligned. A successful system must therefore recover the correct counterpart despite many plausible but incorrect alternatives.

The repository does not impose a single official train/validation/test split, because different studies may evaluate under transductive retrieval, inductive generalization, candidate generation plus reranking, or zero-shot LLM settings. For this reason, papers using the benchmark should report their split construction clearly, specify whether matching requires exact URI equality, state the candidate space used at evaluation time, distinguish lexical, structural, and attribute signals when possible, and document how many gold pairs were actually evaluable by the method.

This last point matters especially for graph-embedding methods. If a method depends on graph structure or attribute evidence, isolated entities may not receive usable representations. Authors should therefore report the number of evaluated gold pairs, the number of excluded pairs, and whether exclusions were caused by missing relations, missing attributes, or both.

Intended Scope and Limitations

This dataset is intended for heterogeneous knowledge graph alignment, biomedical entity alignment, ontology matching, knowledge graph representation learning, candidate generation for terminology mapping, alignment reranking, and graph-based or LLM-based biomedical matching evaluation. It is not intended to serve as a clinical decision support resource, a substitute for official MeSH or SNOMED CT releases, or a definitive source of fully adjudicated one-to-one mappings for every downstream clinical purpose.

The main limitations are worth stating clearly. The ground-truth links come from existing matching resources rather than newly adjudicated annotation created specifically for this release. Not all biomedical concepts are one-to-one alignable across MeSH and SNOMED CT. Some alignments may depend on granularity, synonymy, or ontology modeling choices. A small number of gold entities are isolated in the released graph files. Finally, both MeSH and SNOMED CT continue to evolve over time, so future source releases may differ from the versions reflected here.

Citation

If you use this benchmark, please cite the dataset and, once available, the accompanying paper.

@misc{mesh_snomed_ea_15k_2026,
  title        = {MeSH-SNOMED Entity Alignment 15K},
  author       = {Ravideshik, Vaibhavalakshmi and collaborators},
  year         = {2026},
  howpublished = {Hugging Face dataset repository},
  note         = {Dataset for accompanying EMNLP 2026 submission}
}

The original source resources should also be cited:

@misc{mesh_nlm,
  title        = {Medical Subject Headings (MeSH)},
  author       = {{U.S. National Library of Medicine}},
  howpublished = {https://www.nlm.nih.gov/mesh/}
}

@misc{snomed_ct,
  title        = {SNOMED CT},
  author       = {{SNOMED International}},
  howpublished = {https://www.snomed.org/what-is-snomed-ct}
}

License and Acknowledgements

This repository redistributes benchmark files derived from external biomedical resources, and users are responsible for ensuring that their use complies with the original providers’ terms and access conditions. MeSH is provided by the U.S. National Library of Medicine. SNOMED CT has jurisdiction- and license-specific usage conditions, and users should consult SNOMED International and relevant NLM guidance before redistribution or downstream commercial use.

We thank the maintainers of MeSH, the U.S. National Library of Medicine, and SNOMED International for making large-scale biomedical terminology resources available to the research community.

References

The benchmark is grounded in official source resources and documentation from the underlying biomedical terminologies. Particularly relevant references include the MeSH home page, the MeSH RDF portal, the MeSH RDF documentation, the official MeSH RDF download site, the SNOMED CT overview page, the SNOMED International homepage, the NLM SNOMED CT overview, and the SNOMED documentation portal.