InformationContent: Semantic-level comparison using WordNet synsets and IC scores

This metric aims to evaluate semantic similarity between a generated label and a gold-standard label using WordNet synsets and Information-Content (IC) scores such as Resnik, Lin, or Jiang-Conrath. Each label is mapped to a synset, and the IC of their lowest common ancestor in the WordNet taxonomy is normalised to a similarity in [0, 1]. The method was considered but excluded from the final pipeline because IC requires a shared taxonomy, which is generally not the case for LLM-generated outputs.

Example:

Generate_Label_1: dog

Generate_Label_2: cat

Generate_Label_3: software

Gold_Standard_Label_1: hound

Gold_Standard_Label_2: feline

Gold_Standard_Label_3: Wine

Result:

1. InformationContent(Generate_Label_1, Gold_Standard_Label_1) ≈ 0.90
2. InformationContent(Generate_Label_2, Gold_Standard_Label_2) ≈ 0.85
3. InformationContent(Generate_Label_3, Gold_Standard_Label_3) ≈ 0.05

What is being measured?

This metric captures taxonomy-aware semantic similarity. It is strong when both sides share a stable hierarchy and weak otherwise, which is why it remains in the catalogue only as a documented baseline.