Reconciling Semantic Heterogeneity across Disparate Data Sources

590 concepts540 papers9 proposals

When independent teams, companies, or communities build databases, they inevitably design different schemas for the same real-world concepts. One database might store an employee's earnings under the column salary, another under income, and a third under a cryptic shorthand like pay_empr. This mismatch is known as semantic heterogeneity.

To a human, the relationship between "salary" and "income" is obvious. To a computer, however, they are entirely distinct strings. Traditional data integration techniques—such as exact string matching or relying on rigid primary-key/foreign-key constraints—fail when confronted with these variations. As Ravi Gummadi demonstrates in Using Mined Attribute Dependencies to Improve Retrieval Accuracy over Web Databases in the Absence of Primary-Key Foreign-Key Dependencies, reconstructing fragmented entities across web databases is severely bottlenecked by these naming variations.

To build systems that can automatically merge catalogs, clean data lakes, or query distributed sensors, we must find ways to align these disparate schemas into a shared semantic space.

From Lexical Strings to Semantic Meaning

The earliest and most intuitive approach to resolving semantic heterogeneity is to move past literal string matching and look at the actual meaning of terms. As argued by Fausto Giunchiglia et al. in Towards Explaining Semantic Matching, true semantic interoperability requires determining relationships based on conceptual definitions rather than simple lexical similarity.

To achieve this, ontology alignment systems typically combine two distinct signals:

  1. Linguistic Vibe: Analyzing the labels themselves using external dictionaries (like WordNet) to find synonyms, hypernyms, or abbreviations.
  2. Structural Layout: Analyzing how concepts are organized hierarchically. If two terms have similar parent and child concepts, they are likely equivalent.

For example, the RiMOM framework developed by Juanzi Li et al. in Result of Ontology Alignment with RiMOM at OAEI'06 dynamically assesses both the linguistic and structural characteristics of the ontologies before deciding which mapping strategy to prioritize. If the labels are highly descriptive, linguistic matching dominates; if the labels are obscured or missing, the system relies heavily on structural similarity propagation.

Capturing Context via Subgraphs

What happens when labels are completely uninformative? Consider a database where columns are named col_A, col_B, and col_C. Linguistic analysis is useless here.

To solve this, Peng Wang et al. introduced a powerful intuition in LILY: the Results for the Ontology Alignment Contest OAEI 2007: semantic subgraphs. Instead of looking at a concept in isolation, the system extracts its immediate structural neighborhood.

[Employee] ──(has_attribute)──> [col_A] ──(measured_in)──> [USD]

Even if col_A is an anonymous label, the fact that it connects an Employee to a currency (USD) provides a structural fingerprint. By comparing these subgraphs across different schemas, alignment engines can infer that col_A in one system serves the same structural role as income in another.

Structural and Dimensional Mismatches

Semantic heterogeneity is not merely a naming problem; it is also a structural one. Two schemas might agree on what a concept means but disagree entirely on how to model it.

In Towards a Benchmark for Instance Matching, Alfio Ferrara et al. point out that alignment algorithms frequently stumble when data is aggregated differently—such as when one system represents an address as a single flat string, while another flattens it into a complex object with separate entities for street, city, and postal code.

Furthermore, deep semantic conflicts often occur at the value level. As Jovan Varga notes in Dimensional enrichment of statistical linked open data, even when automated systems successfully identify that two columns represent financial metrics, they often fail to resolve conflicting units of measurement or differing currencies. True alignment requires enriching these schemas with explicit dimensional metadata to ensure data compatibility.

The Modern Era: Hybrid and Pretrained Models

Historically, schema matching relied on hand-crafted rules, graph propagation frameworks like X-SOM (Carlo Curino et al.), or multi-agent architectures like the Intelligent Mapping Toolkit (Enabling the Interoperability of Large-Scale Legacy Systems). While effective, these systems were often "single-table" specialists, requiring manual tuning for every new dataset.

The modern paradigm shifts this burden onto machine learning models that can learn generalized representations of tabular data. The authors of Towards Cross-Table Masked Pretraining for Web Data Mining (Chao Ye et al.) address this by using masked pretraining across thousands of diverse tables. By learning how column headers co-occur across the web, the model maps diverse headers into a shared semantic space, automatically recognizing that salary and income behave identically across different contexts.

Traditional:  [Schema A] ───(Hand-crafted Rules)───> [Schema B]
Modern:       [Schema A] ───> [Shared Semantic Space] <─── [Schema B]

To make this practical at scale, frameworks like Magneto (Yurong Liu et al.) employ a hybrid approach. Because large language models (LLMs) possess deep semantic reasoning but are computationally expensive, the authors combine them with small language models (SLMs) in a retrieve-and-rerank pipeline. The SLM quickly prunes the search space, while the LLM performs the heavy lifting of reasoning over cryptic column names and heterogeneous data values.

Real-World Impact

Resolving semantic heterogeneity unlocks massive operational and analytical capabilities across various domains:

  • Data Lakes: In massive enterprise data repositories, redundant data often hides behind different names. By identifying semantic containment without relying on identical column labels, the R2D2 pipeline (Raunak Shah et al.) can automatically flag duplicate datasets, saving millions in storage and compliance costs.
  • Industrial IoT: In large-scale engineering environments, sensors from different manufacturers use inconsistent naming conventions for identical physical properties. Systems like Optique (Evgeny Kharlamov et al.) bridge this gap by providing a unified semantic layer, allowing engineers to query thousands of streaming and static data sources simultaneously without manually rewriting queries for each schema variation.
  • Sustainability and Science: Tracking global environmental impacts requires integrating highly fragmented datasets that use conflicting classification systems. Projects like the open-source ontology for product footprinting (Agneta Ghose et al.) rely on these semantic web technologies to bridge different industrial domains, making life cycle assessments transparent and reproducible.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 494 papers
  2. Version 1 · 2026-10-04 · 489 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_semantic_heterogeneity_ontology_misalign_fa80d19c,
  title = {semantic heterogeneity (ontology misalignment)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-heterogeneity-ontology-misalignment/rdr_fa80d19c7b0f468f95e8d647241d94b1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_semantic_heterogeneity_ontology_misalign_fa80d19c_v2,
  title = {semantic heterogeneity (ontology misalignment)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-heterogeneity-ontology-misalignment/rdr_fa80d19c7b0f468f95e8d647241d94b1/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake