A brief history of semantics in data science

 

1. Before data science: meaning lived in people (pre-1990s)

 

Early information systems were built around bespoke databases and reports.  Meaning was implicit: developers and analysts knew what tables and fields represented, but this knowledge was rarely written down formally.  As long as systems were small and stable, this worked.

________________________________________

 

2. Relational databases: structure without meaning (1970s–1990s)

 

Relational databases introduced strong structural discipline (tables, keys, constraints), but they deliberately avoided encoding meaning beyond basic integrity rules.

 

Schemas described how data was stored, not what it meant in a business sense.

 

Semantics remained:

•in documentation

•in naming conventions

•in people’s heads

 

As systems multiplied, this gap widened.

________________________________________

 

3. Data warehousing and BI: semantics as translation (1990s–2000s)

 

With data warehouses and BI tools, organisations began explicitly mapping data into business concepts (facts, dimensions, measures).

 

This was an early form of semantic modelling, but it was:

•tool-specific

•fragile

•often duplicated across reports

•disconnected from operational systems

 

Semantics existed, but only inside reporting layers.

________________________________________

 

4. The Semantic Web: formal semantics (late 1990s–2010s)

 

The Semantic Web movement introduced formal, machine-readable semantics:

•ontologies

•RDF

•OWL

•explicit relationships

 

This was the first serious attempt to encode meaning so machines could reason about it.

 

However, adoption outside academia and specialist domains was limited due to:

•complexity

•tooling maturity

•lack of immediate business payoff

 

The ideas were sound; the timing was early.

________________________________________

 

5. Big data era: semantics sidelined (2010s)

 

Big data technologies prioritised:

•scale

•speed

•flexibility

 

Schema-on-read reduced upfront modelling effort, but also pushed semantics even further downstream.

 

Meaning was inferred late, inconsistently, and often differently by each team.  This worked for experimentation, but not for:

•governance

•consistency

•automation

 

Semantic debt accumulated.

________________________________________

 

6. Knowledge graphs: semantics return pragmatically (mid-2010s–present)

 

Knowledge graphs brought semantics back in a more practical, incremental form:

•explicit entities and relationships

•partial ontologies rather than full formal models

•integration across systems

 

They allowed organisations to:

•connect data without fully unifying it

•preserve context

•make relationships explicit

 

This marked a shift from “store everything” to “understand what connects”.

________________________________________

 

7. AI and LLMs: meaning becomes critical again (2020s)

 

Modern AI systems can generate answers, but they do not understand meaning in a business or organisational sense.

 

As AI systems are applied to:

•decision support

•automation

•regulated domains

 

the lack of explicit semantics becomes a risk, not just a nuisance.

 

This has renewed interest in:

•symbolic AI

•explicit definitions

•rules and constraints

•knowledge graphs as grounding layers

 

Semantics is no longer academic — it is operational.

________________________________________

 

 

Where we are now

 

Data science has learned that:

•scale without meaning creates fragility

•inference without structure creates risk

•automation without semantics creates surprises

 

The current direction is not a rejection of statistical AI, but a rebalancing:

•statistical models for pattern recognition

•symbolic structures for meaning, rules and control