11. The intelligence layer | Semantic Layer course

Section 1

The rung that was inserted

Next section

Section 2

Ontology induction, four ways in

Next section

Section 3

The governance gate

Next section

Section 4

Knowing how good your ontology is

Next section

Section 5

When a true fact stops being true

Next section

Section 6

Short-term memory

Next section

Section 7

Where the intelligence layer runs

Hands on

Induce and review

35 min

Auto-generate a candidate ontology plus R2RML mappings from Meridian's Postgres schema into a draft named graph, run ELK for consistency, diff it against your hand-written ontology, then promote the good parts and reject the bad.

make lab-up            # once, starts Postgres + Fuseki + Ontop
make lab lab11-induction

Full instructions, starter files and the solution are in labs/modules/lab11-induction/README.md. The lab checks its own work, so you will know if it passed.

Check yourself

Quiz

7 questions, pass mark 70%. Answers are explained either way, so a wrong guess is not wasted.

1 The AWS guidance inserts an Intelligence tier between Knowledge and Wisdom in the classic DIKW ladder. What does that tier hold?

Faster copies of the knowledge layer, for query performance Meta-knowledge: knowledge about the knowledge, including its coverage, freshness and consistency The machine learning models that generate embeddings The business rules, moved out of the knowledge layer

2 Why can virtualized knowledge accessed through the VKG not go stale, while materialized knowledge can?

Virtualized queries are cached with a short time-to-live, so staleness is bounded to a few minutes There is no copy: the SPARQL query is rewritten to SQL and runs against the live source at query time, so it returns whatever the source holds now The VKG runs a background refresh job that re-syncs the graph every hour Ontop validates every result against SHACL shapes before returning it

3 Induction proposes a class ex:RestructuredLoan defined as any loan with restr_flag = 'Y', confidence 0.88. Ayesha Kapoor from Risk rejects it. Why?

The confidence score was too low for an automatic promotion restr_flag is never cleared, so the flag records a past event while the proposed class asserts a current state, and every loan ever restructured would be permanently in it OWL cannot define a class from a datatype property value FIBO already has a class for it, so Meridian should align instead of defining its own

4 Which two things must be true of induced ontology output before any service acts on it? (pick two)

Select all that apply.

It lands in a draft named graph flagged as DRAFT, which the VKG and reasoning services are configured not to read It is converted to SHACL shapes before it can be stored It passes a promotion gate that includes human approval and a reasoner consistency check It reaches a confidence score above 0.95 on every proposal

5 Daniel Okafor's annual income of 54,000 was extracted from a payslip in March 2026. He changed jobs in June. What does truth maintenance do with the March assertion?

Deletes it, since it is now incorrect Marks it SUPERSEDED, records the new assertion with a link between them, and re-runs reasoning over everything derived from it Lowers its confidence score and leaves it in place Nothing, because the payslip it came from is still a valid document

6 A session context graph types its nodes against the ontology (mrb:Query, mrb:Intent, mrb:Document) instead of storing a plain chat transcript. What does that buy you?

Smaller storage, because IRIs compress better than text The session content is ontology-grounded, so the reasoner can infer across the session graph joined with the main graph, and retrieved entities are IRIs rather than ambiguous names The session survives a restart of the graph store It removes the need for the intent classifier at query time

7 What do OWL2Vec* and RDF2Vec produce, and what is it used for in the Intelligence layer?

SHACL shapes generated from observed data patterns, used for validation Vector embeddings of ontology and graph structure, used to compute similarity between a discovered class and a reference ontology concept when the names do not match Compressed binary serializations of large graphs, used to speed up reasoning Natural-language definitions for induced classes, used to fill in rdfs:comment