1956
Dartmouth workshop
AI is named. The field begins with logic and search.
The shape of the next AI decade
Neuro-
Symbolic AI.
How we got to large language models — LLMs. How "knowledge graphs" and "ontologies" are learning to live inside them — and why that changes everything.
Speaker
Dr. Samir Sellami
LIRE Laboratory, Univ-Constantine 2
Dept. of Mathematics & Computer Science, ENSET-Skikda, Azzaba, Algeria
Dept. of Mathematics & Computer Science, ENSET-Skikda, Azzaba, Algeria
A 20-minute map
Four short acts.
- 01Where we came from. Two schools of AI — symbolic and neural.
- 02The limits of LLMs. Hallucination, provenance, and the absent ground truth.
- 03Knowledge graphs & ontologies. The quiet return of structured meaning.
- 04Knowledge-Infused Learning. The synthesis — and what it unlocks.
01
Two Big Schools
One goal,
two main directions.
Symbolists & connectionists (neural networks) have argued about the nature of intelligence for seventy years. Today, for the first time, they have to collaborate.
The discipline at a glance
AI has many tools
The Synthesis · This Talk
Neuro-Symbolic AI
①
Symbolic
⊕
②
Neural
= grounded, explainable AI
The goal — defined in 1956
“To develop machines that behave as though they were intelligent.”
— John McCarthy, Dartmouth Summer Research Project
①
Symbolic / Logical-AI
knowledge·driven · explicit
Knowledge Graphs
Ontologies
Semantic Web
Formalisms RDF OWL SPARQL SHACL
Formalisms RDF OWL SPARQL SHACL
Reasoning
Planning & Robotics
Expert Systems
NLP
Natural Language Processing · classical pipeline
Tokenization
›
Lemmatization
›
POS
›
Parsing
›
NER
→ KGs & Ontologies (with RDF/OWL/SPARQL/SHACL) are this talk's main concern.
②
Neural / Statistical-AI
data·driven · learned
Foundation
MLPs
Embeddings
Backprop
Machine Learning
Supervised
Unsupervised
Reinforcement
Deep Learning · Neural Networks
CNN
RNN
Transformers
Generative AI
← the talk's focus on the right
LLMs
Agentic AI
Diffusion · GANs
① grounds and constrains ② → Knowledge-Infused Learning (slides 14–15)
The symbolic era
Intelligence as the manipulation
of symbols.
1970s
Expert systems
MYCIN, DENDRAL. Hand-coded rules, transparent reasoning.
1990s
Ontologies & the Web
Description logics, OWL, the Semantic Web vision.
~2000
The winter
Brittleness. Scale failure. Funding collapses.
Beautifully explainable — you could read the rules. But every edge case had to be written by hand, and the world had too many edges.
ImageNet top-5 error rate (%)
The neural renaissance
Intelligence as pattern,
learned from data.
2012
ImageNet & AlexNet
Deep CNNs crush image classification. GPUs take over.
2017
Attention is all you need
The Transformer arrives. Sequences become scalable.
2020
GPT-3 & scaling laws
Bigger + more data = emergent capability.
Nov 2022
ChatGPT
AI leaves the lab. A civilisational moment.
What the symbolic tradition couldn’t do — see, hear, translate, converse — the neural tradition did, and decisively.
November 2022 — LLMs started a methodological revolution
A single prompt now does the work of an entire NLP pipeline.
Raw text goes in. Structured meaning comes out.
— What took five tools & a week now takes a sentence & a second.
Before
Raw text
›
Tokenize
›
POS
›
Parser
›
NER
›
Coref
›
Extractor
›
JSON
≈ 5 tools · ~1 week
Now
Raw text + 1 prompt
→
LLM
→
Structured JSON
≈ 1 prompt · ~1 second
Let’s be fair
Deep learning / Transformer did —
what symbolic AI never could.
Perception
- Image & scene recognition
- Speech recognition
- Handwriting transcription
Language manipulation
- Machine translation
- Text-to-speech / TTS
- Fluent conversational answers
Science & autonomy
- Protein structure prediction
- Autonomous driving
- Superhuman game-play
The turning point
“There ain’t no such thing
as a free lunch.”
as a free lunch.”
— Robert A. Heinlein, The Moon is a Harsh Mistress (1966)
↓ Every gain in convenience has a cost.
The honest critique
LLMs are Fluent. Confident.
but Often ungrounded.
- Hallucination. Confidently wrong answers that are syntactically perfect.
- No provenance. The model cannot tell you where a fact came from.
- No consistency. The same question can return contradictory answers.
- Hard to update. New facts require retraining — not editing.
- GPU and Tokens hungry. Require multi-GPU and consume vast amounts of tokens.
- Opaque. Knowledge is entangled in billions of weights.
The deeper problem
Statistics
is not
knowledge.
An LLM learns what words tend to follow what words. It does not, in any strong sense, know what a patient, a contract, or a regulation is.
Analogy:
the LLM (left) is the engine of the car — powerful, but driving like a teenager by instinct;
the knowledge graph (right) is the steering wheel — a city-map expert giving direction.
Structure returns
The quiet
return of
meaning.
Knowledge graphs and ontologies — the unfashionable cousins of machine learning — are suddenly the hottest technology in enterprise AI.
Graph vs. ontology
The graph is
the memory.
The ontology
is the grammar.
Memory
Facts of the world — written down, recallable.
Grammar
Rules of form — what may combine with what.
-
RDF / KG. Entities & relationships.
«Ibuprofen treats inflammation.» «Tehran is the capital of Iran.»
-
OWL. Classes, properties, constraints.
«A Patient has exactly one DOB.» «A Contract has ≥2 Parties.»
-
SHACL. Rules & validation.
Enforce what must be true. Catch what cannot.
An industry signal
Ontology is
having its moment.
having its moment.
— Tony Seale, Knowledge Graph Engineer · LinkedIn, 2025
A quiet corner of CS for 20 years.
Now every major enterprise is building one.
Now every major enterprise is building one.
Knowledge-Infused Learning — KiL
Neural intuition
× symbolic truth.
Structured human knowledge — domain rules, knowledge graphs, physical laws — fed directly into statistical ML.
+ Accurate
Grounded in ground truth
+ Explainable
Provenance for every claim
+ Efficient
Less data, cleaner signal
Analogy:
the LLM is the smooth, charismatic salesperson — promising a deal to the CEO;
the knowledge graph & ontology are the strict back-room accountant — auditing every fact before the deal is signed.
The numbers behind the claim
The benchmarks now back this up.
+71%
Biomedical multiple-choice QA
Llama-2 with KG-RAG (SPOKE) vs. base model on biomedical MCQ.
Soman et al., Bioinformatics, 2023.
Soman et al., Bioinformatics, 2023.
+77.6%
Retrieval quality (MRR)
Customer-service KG-RAG deployed at LinkedIn — −28.6% median resolution time.
Xu et al., ACM SIGIR, 2024.
Xu et al., ACM SIGIR, 2024.
> −40%
Hallucination rate
Multi-evidence RAG with biomedical KGs in public-health QA.
Xu et al., Frontiers in Public Health, 2025.
Xu et al., Frontiers in Public Health, 2025.
Substantial
Comprehensiveness & diversity
GraphRAG vs. conventional RAG on global sense-making over 1M-token corpora.
Edge et al., Microsoft Research, 2024.
Edge et al., Microsoft Research, 2024.
Different domains, different metrics — same direction. Structure doesn't replace the model; it makes the model accountable.
Three points of entry
Three points where you can inject knowledge
Before, during, or after the model.
Real systems combine all three. The architecture is a pipeline, not a swap.
01
Healthcare & medical QA
A clinician asks about a drug interaction.
Without KG
The LLM invents imaginary facts. No source, no audit trail, no verification.
With KG
Grounded in UMLS / SNOMED ontologies. Every claim traceable. Ontology catches contradictions before they reach the patient.
Result
A toy becomes a tool. Regulators are now asking for this explicitly.
02
Legal & compliance
Review a thousand contracts without missing a clause.
Without KG
LLM alone misses obligations, mis-attributes parties, contradicts itself across documents.
With KG
Legal ontology: parties, obligations, jurisdictions, effective dates. Cross-document reasoning becomes possible.
Result
Consistency at scale. EU regulators are pushing this direction hard.
03
Enterprise knowledge
Ten years of documents. One assistant that actually knows your business.
Without KG
A general model guesses. Answers drift. Private data must be pasted every time.
With KG
Enterprise KG unifies wikis, databases, emails, product specs. The LLM becomes a natural-language interface over your own ground truth.
Result
A decisive competitive advantage — and a moat no competitor can just download.
04
Education & research
How do we train the next generation when LLMs seem to make the basics optional?
The risk
Students skip fundamentals — programming, querying, data design, ontology engineering — because the chatbot does it for them.
The answer
Teach them to build the grounding layer — three concrete tracks every CS programme can add now:
① Ontology engineering — OWL, SHACL, domain modelling.
② Knowledge-graph systems — SPARQL / Cypher, Graph-RAG, evaluation.
③ A grounded-AI capstone — build & SHACL-validate a KG-backed RAG system end-to-end.
Why it matters here
The neural layer will commoditise. The knowledge layer is where expertise lives — and where Algerian research can lead.
A call to action
Build with
both hands.
Generative AI is powerful and accessible — and it needs control. Knowledge graphs and symbolic reasoning provide that grounding. The future of AI is not neural, it is not symbolic. It is both.
Thank you
Dr. Samir Sellami
LIRE Laboratory, Univ-Constantine 2 · ENSET-Skikda, Azzaba · NCSDS-26