Relation normalization maps varied, semantically similar extraction labels to a controlled relation vocabulary. In knowledge-graph extraction, generic prompts can produce unstable labels such as “works with,” “works at,” “specializes in,” or “is measuring”; specifying entity classes, relation types, aliases, and examples produces more consistent output. section 5.3.2
A practical normalization prompt combines:
- Alias instructions for initials, surnames, universities, and departments so references resolve to consistent entities. section 5.3.2
From varied labels to a controlled relation
An initial extraction assigns “thermodynamics” to “specializes in” and “specific heats” to “is measuring.” A revised prompt normalizes both as “works on,” while preserving distinct relations such as “works for,” “works with,” and “student of.” The result is more consistent, but variants such as “Raman spectra” versus “theory of Raman spectra” and residual relation differences still require validation. section 5.3.2
