Characterize subtle framing changes undetected by semantic similarity

Identify and characterize subtle linguistic framing changes between Grokipedia and English Wikipedia articles that maintain high embedding-based textual similarity, including single-word shifts and active–passive voice changes, to assess their impact on content interpretation and neutrality.

Background

The authors caution that embedding-based similarity can miss subtle shifts in framing that preserve high textual similarity, such as word choice changes or alterations in grammatical voice.

They explicitly state that identifying and characterizing these subtle changes remains open, motivating methods that go beyond cosine similarity to capture interpretive differences.