Semantic heterogeneity is when database schema or datasets for the same domain are developed by independent parties, resulting in differences in meaning and interpretation of data values. Beyond structured data, the problem of semantic heterogeneity is compounded due to the flexibility of semi-structured data and various tagging methods applied to documents or unstructured data. Semantic heterogeneity is one of the more important sources of differences in heterogeneous datasets.
Yet, for multiple data sources to interoperate with one another, it is essential to reconcile these semantic differences. Decomposing the various sources of semantic heterogeneities provides a basis for understanding how to map and transform data to overcome these differences.
Conflicts between independently designed database schemas were studied extensively in the 1980s as part of research in schema matching. Batini, Lenzerini, and Navathe compared existing integration methodologies in 1986 and calloused the naming and structural conflicts each had to resolve. William Kent's 1989 paper "The many forms of a single fact" used examples to show how one fact can be represented in many different forms.
Sheth and Larson's 1990 survey of federated database systems defined semantic heterogeneity as disagreement about the meaning, interpretation, or intended use of the same or related data. They separated these conflicts of meaning from differences among database systems and data models.
Later work concentrated on automating the detection and resolution of such conflicts. Rahm and Bernstein reviewed techniques for automatic schema matching in 2001, and Euzenat and Shvaiko's Ontology Matching covers the corresponding problem for ontologies. Halevy argued in 2005 that semantic heterogeneity is inevitable because schemas will always be built by independent parties. Machine learning and, more recently, large language models have been applied to schema matching, ontology alignment and entity matching.
One of the first known classification schemes applied to data semantics is from William Kent in the late 80s. Kent's approach dealt more with structural mapping issues than differences in meaning, which he pointed to data dictionaries as potentially solving.
One of the most comprehensive classifications is from Pluempitiwiriyawej and Hammer, "Classification Scheme for Semantic and Schematic Heterogeneities in XML Data Sources". They classify heterogeneities into three broad classes:
Structural conflicts arise when the schema of the sources representing related or overlapping data exhibit discrepancies. Structural conflicts can be detected when comparing the underlying schema. The class of structural conflicts includes generalization conflicts, aggregation conflicts, internal path discrepancy, missing items, element ordering, constraint and type mismatch, and naming conflicts between the element types and attribute names.
Domain conflicts arise when the semantics of the data sources that will be integrated exhibit discrepancies. Domain conflicts can be detected by looking at the information contained in the schema and using knowledge about the underlying data domains. The class of domain conflicts includes schematic discrepancy, scale or unit, precision, and data representation conflicts.
Data conflicts refer to discrepancies among similar or related data values across multiple sources. Data conflicts can only be detected by comparing the underlying sources. The class of data conflicts includes ID-value, missing data, incorrect spelling, and naming conflicts between the element contents and the attribute values.
Moreover, mismatches or conflicts can occur between set elements (a "population" mismatch) or attributes (a "description" mismatch).
Michael Bergman expanded upon this schema by adding a fourth major explicit category of language, and also added some examples of each kind of semantic heterogeneity, resulting in about 40 distinct potential categories . This table shows the combined 40 possible sources of semantic heterogeneities across sources:
A different approach toward classifying semantics and integration approaches is taken by Sheth et al. Under their concept, they split semantics into three forms: implicit, formal and powerful. Implicit semantics are what is either largely present or can easily be extracted; formal languages, though relatively scarce, occur in the form of ontologies or other description logics; and powerful (soft) semantics are fuzzy and not limited to rigid set-based assignments. Sheth et al.'s main point is that first-order logic (FOL) or description logic is inadequate alone to properly capture the needed semantics.
The traditional approach to resolution is mapping where a practitioner studies the sources, writes rules to translate between them, and revises the rules when a schema changes. Wiederhold proposed "mediators" for this: software modules that sit between applications and data sources and hold the knowledge needed to translate. Later data integration systems accept queries against a single mediated schema and answer them through stored mappings to each source.
The core problem to this approach is that writing mappings by hand is slow, error-prone, and can have biases based on the practitioner. Schema matchers compare element names, structures and data values to propose likely correspondences. The same idea extends to ontologies. Entity matching (also called record linkage) works on the data itself, deciding whether two records refer to the same real-world entity.
Neural networks changed the matchers. Trained on labeled pairs, they learn the similarity judgments that older systems encoded in rules. Pre-trained language models represent records as embeddings, so a matcher can tell that "HTN" and "high blood pressure" name the same condition. Large language models judge matches from a prompt, with a few labeled examples or none. Knowledge graphs supply what the records leave out, such as synonyms and type hierarchies.
Standards address the problem before integration begins. The FAIR principles, published in 2016, ask data publishers to describe their data with shared, formal vocabularies. Datalake systems store data in its original form and add semantics afterward with a metadata layer that links stored files to knowledge graphs, and ontology-based data access rewriting queries over an ontology into queries over the sources.
Besides data interoperability, relevant areas in information technology that depend on reconciling semantic heterogeneities include data mapping, semantic integration, and enterprise information integration, among many others. From the conceptual to actual data, there are differences in perspective, vocabularies, measures and conventions once any two data sources are brought together. Explicit attention to these semantic heterogeneities is one means to get the information to integrate or interoperate.
In practice, such conflicts arise when independent groups record the same facts. The same product can appear under its full retail name in one catalog and under an internal stock code in another. Geographic information systems built by different agencies describe the same terrain in different terms. NASA lost the Mars Climate Orbiter in 1999 because ground software reported thruster impulse in pound-force seconds while the navigation software expected newton-seconds. In medicine, where one condition can go by many names, the Unified Medical Language System links more than two million names from over 60 families of biomedical vocabularies to some 900,000 common concepts.
Data integration
Data mapping
Enterprise information integration
Heterogeneous database system
Interoperability
Ontology-based data integration
Schema matching
Semantic integration
Semantic matching
Semantics
