When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval
A new study published on arXiv explores the effectiveness of mixing query embeddings in multilingual dense retrieval systems. Researchers found that interpolating embeddings from different languages can improve retrieval performance, outperforming monolingual queries in most tested cases. The study also identified an asymmetry where English dominance influences retrieval outcomes, with English queries being optimal for English document indices, while mixing benefits non-English indices. The research suggests that language mixing sensitivity is predictable and can be leveraged to enhance multilingual search capabilities. AI
IMPACT Optimizes multilingual search by demonstrating how to effectively blend query embeddings for improved retrieval accuracy.