LLM Hallucination in Knowledge Engineering

The specific challenge of LLM confabulation when used for knowledge engineering tasks, creating, extending, or validating knowledge structures.

Manifestations

Relationship Invention: Creating plausible but false relationships between entities Property Confabulation: Assigning properties entities don’t have Hierarchy Errors: Placing concepts in wrong taxonomic positions Citation Fabrication: Generating nonexistent sources Definition Drift: Subtly redefining terms from their intended meaning

Why Knowledge Engineering is Vulnerable

• Requires factual accuracy (hallucination is directly harmful)
• Involves specialized domains (may be underrepresented in training)
• Errors propagate through connected knowledge structures
• Human reviewers may lack expertise to catch subtle errors

Mitigation Strategies

Verification Pipeline:

• Generate with LLM
• Verify against authoritative sources
• Human expert review for novel claims

Prompt Engineering:

• Request only high-confidence assertions
• Ask for source citations (then verify them)
• Use conservative extraction prompts

Hybrid Approaches:

• LLM for candidate generation
• Rule-based validation
• Human final approval

Related: 05-atom—uniform-confidence-problem, 06-molecule—knowledge-graph-construction