A Method Translates Text Embeddings Across Vector Spaces Without Paired Data

Summary

The paper introduces an unsupervised method for translating text embeddings between different vector spaces without paired examples, encoders, or a predefined set of matching items. The method maps embeddings to and from a universal latent representation based on the conjecture that models share a common semantic structure. The authors report high cosine similarity across model pairs that differ in architecture, parameter count, and training data. The approach is intended to preserve the geometry of an embedding space while moving an unknown embedding into another space. The paper also identifies a security risk for vector databases: an attacker who has access only to embedding vectors may recover sensitive information about the documents behind them. According to the abstract, that information can be sufficient for document classification and attribute inference. The work therefore connects cross-model representation transfer with the privacy and security of systems that store embeddings.