← Go back

Autonomous Semantic Immune Systems for Knowledge Graph Validation

Knowledge Graphs (KGs) are the structural backbone of modern AI, representing facts as triples (Subject, Predicate, Object)—for instance, (MarieCurie, wonPrize, NobelPrize_Physics) . While Large Language Models (LLMs) enable the automatic extraction of facts from text at scale—a process known as Induced Semantics—they introduce a critical flaw: hallucinations. These are triples that are structurally valid (passing standard SHACL/OWL checks) yet factually false, e.g., (MarieCurie, wonPrize, NobelPrize_Literature).

Static validation breaks down here because it only checks for schema compliance, not world knowledge. To solve this, our project brings to life the Semantic Immune System (SIS)—a bio-inspired framework that treats a KG as a living ecosystem. Instead of relying on rigid human-written rules, the SIS autonomously learns to distinguish "Self" (correct facts) from "Non-Self" (semantic pathogens) through evolutionary logic. This project group focuses on the adaptive core of the SIS, moving beyond static validation toward a self-healing semantic infrastructure based on three foundational pillars:

1. Semantic Self & Innate Immunity: Formalizing the trusted "core" of a KG using embeddings and enforcing baseline structural constraints.
2. Adaptive Negative Selection: Automatically generating and filtering logical detectors (SPARQL-like rules) that react exclusively to hallucinated triples without touching the verified Self.
3. Clonal Selection & Semantic Memory: Evolving the best detectors via mutation (to catch novel error patterns) and caching pathogen signatures for ultra-fast secondary response.

Project Objectives

Our project group aims to design, implement, and evaluate a functional prototype of the adaptive SIS detection pipeline. Students will traverse the entire research lifecycle—from literature review and formal problem definition to implementation, experimental benchmarking, and critical analysis.

To achieve these objectives, we are investigating several advanced techniques, including:

• Utilizing embedding models (e.g., RDF2Vec, TransE) to represent the semantic space of a Knowledge Graph;
• Leveraging Large Language Models (LLMs) to generate and mutate candidate logical rules (detectors) through few-shot prompting;
• Implementing evolutionary algorithms (fitness evaluation, elitist selection, and mutation) to improve detector populations over time;
• Building a semantic memory cache (using FAISS) to enable rapid recognition of recurring hallucination families;
• Quantitatively benchmarking the adaptive system against static SHACL/OWL constraints to measure gains in precision, recall, and robustness under noisy oracle conditions.

Through participation in this project group, students will acquire hands-on experience across the entire pipeline of autonomous KG integrity management, covering conceptual formalization, dataset engineering, neuro-symbolic AI integration, and performance optimization.

FAQs

Q: What is the selection process for this project?
A: Candidates will need to submit a short assignment (implementing a simple embedding-based similarity check) and undergo an interview as part of the selection process.

Q: Is there a seminar connected to this PG?
A: No.

Q: What are the prerequisites for this PG?
A: The ideal candidate should possess foundational knowledge in Machine Learning and Natural Language Processing, along with programming skills in Python. Familiarity with Knowledge Graphs (RDF, SPARQL) is a strong plus but not strictly mandatory—willingness to learn quickly is essential. Experience with version control (Git) are also required. The domain is highly interdisciplinary (AI + symbolic logic + bio-inspired algorithms), so a steep but rewarding learning curve is expected.

Q: Do we need to implement the entire SIS framework?
A: No. The project strictly focuses on the Adaptive Detection & Evolution layer (Negative Selection, Clonal Selection, and Semantic Memory). The "Vaccination" and full-scale KG repair components are explicitly scoped out as future work, allowing us to dive deep into measurable experimental results.

Q: How do we evaluate if our system works?
A: You will inject controlled synthetic hallucinations into a KG (e.g., a subset of Wikidata) and measure how many are correctly quarantined (Precision/Recall/F1). Your primary baseline will be standard SHACL constraints. You will also analyze performance degradation when the "oracle" (ground truth) is noisy—a core discussion point in your thesis.

In case you have further questions, feel free to contact Manzoor Ali.

Course in PAUL

Project Group: Autonomous Semantic Immune Systems for Knowledge Graph Validation (in English)