Kelvin Koa and Baishi Li introduced the Construct Validity Protocol, a pipeline designed to validate NLP embeddings as measures of social concepts. The protocol addresses the "Proxy Presumption," where geometric properties are assumed to directly measure constructs like novelty and bias without accounting for confounding attributes such as topic or authorship. The authors also proposed Counterfactual Neutralization, a method using LLMs to reduce confounding in embedding space, alongside a standardized Validity Suite for testing.
No score is assigned. Sources and their independence are shown in the citation chain below.