FIBO and In-Place Semantic Mapping for Banking

Banks drown in report sprawl because every team invents its own definition of customer, exposure, product, or counterparty. The fix is a shared, machine-readable vocabulary — and in finance, that vocabulary already exists. FIBO (the Financial Industry Business Ontology) is finance-specific by design, with classes for instruments, parties, obligations, and contracts. You do not have to invent an ontology; you have to map to one.

The knowledge backbone — a layered standards stack

FIBO is not used alone. It sits inside a layered modeling approach built entirely on open standards:

Layer	Standard	What it gives you
Glossary / taxonomy	SKOS	Human-readable business concepts and synonyms
Ontology	OWL (bootstrapped from FIBO)	Banking classes, properties, relationships
Physical mapping	R2RML / OBDA	Relational sources → RDF, virtual or materialized
Validation	SHACL	"Every exposure resolves to one counterparty"
Provenance	PROV-O	Source, pipeline, and steward decisions per fact
Catalog publication	DCAT	Federated discovery across domains

Because every layer is an open standard, the model stays portable across clouds and vendors — no lock-in at the semantic layer.

What the mapping actually looks like

Two datasets in different systems both describe a FIBO concept. The mapping makes that explicit with R2RML and constrains it with SHACL:

@prefix fibo: <https://spec.edmcouncil.org/fibo/> .
@prefix bank: <https://ABCbank.com/kg/> .

bank:tbl_oracle_acct  a dcat:Dataset ;
  bank:storesConcept  fibo-fbc:DepositAccount ;
  bank:sourceSystem   "oracle.crm.acct" ;
  r2rml:logicalTable  [ r2rml:tableName "acct" ] .

bank:tbl_sqlsrv_loan  a dcat:Dataset ;
  bank:storesConcept  fibo-loan:Loan ;
  bank:cdeField       "outstanding_principal" ;
  sh:property [ sh:path bank:lei ;
                sh:minCount 1 ; sh:datatype xsd:string ] .

Once oracle.crm.acct and sqlsrv.loan both declare their FIBO concept and share an lei, an agent can join them by meaning — it knows both belong to the same LegalEntity — without anyone hand-coding that relationship per query.

Mapping at scale — the agentic pipeline

The obstacle has always been effort: mapping thousands of columns to ontology classes by hand is infeasible. Schema-Miner Pro is an agentic pipeline that drafts the mappings and routes only the uncertain ones to humans:

1. Ingestion & drafting — an LLM parses table headers, schemas, and descriptions from Databricks foreign catalogs and drafts a lightweight initial mapping.
2. Hybrid grounding — lexical heuristics first (cust_id → hasIdentifier), then a FAISS semantic nearest-neighbor search over FIBO OWL embeddings for ambiguous fields.
3. Template generation — approved mappings are exported as RDF/Turtle, linking columns to canonical FIBO identifiers.

A critical design choice: when no alignment exists, the pipeline returns an empty result rather than a forced mapping. A wrong mapping is worse than a missing one, because it silently corrupts every downstream inference.

# every new column → draft → ground → confidence score
draft = llm.map_column(col, schema_ctx)         # "cust_segment" → CustomerSegment?
if lexical_match(col):                            # direct name match
    mapping, c = lexical_match(col), 0.95
else:
    hits = faiss.search(embed(col), fibo_owl, k=3)  # semantic fallback
    mapping, c = hits.best_or_none()              # None if no alignment
route = "commit" if c >= 0.85 else "sme_review"   # confidence gate

The confidence gate keeps it trustworthy

Every proposed mapping carries a confidence score c ∈ [0,1]. Mappings at c ≥ 0.85 auto-commit; anything below locks the column and routes to an SME dashboard. Stewards validate or adjust — they curate proposals, they do not rebuild from scratch — and every correction re-trains the FAISS memory, lowering future error rates. This human-in-the-loop spine is covered in its own article.

Ready-made banking patterns

FIBO-aligned mapping unlocks reusable, governed data products the whole enterprise consumes: Customer 360, Loan, Deposit, Credit, plus AML and regulatory-reporting products. Publish once, discover through the Data Marketplace, reuse everywhere.

The path to implement

1. Pick one high-value domain (e.g. counterparty exposure) and bootstrap an OWL ontology from FIBO.
2. Run Schema-Miner Pro over its foreign catalogs to draft column → FIBO mappings.
3. Constrain the model with SHACL ("every exposure resolves to one counterparty") and capture PROV-O provenance.
4. Route low-confidence mappings to stewards; commit high-confidence ones to the RDF store.
5. Publish the governed product to the Data Marketplace and attach agents.

The result: one vocabulary, many systems — report sprawl becomes governed self-service, and agents reason over customer, exposure, and counterparty instead of table names.

Map your estate to FIBO with Schema-Miner Pro →