When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval

The study presents a systematic evaluation of query embedding interpolation in multilingual dense retrieval, focusing on the mMARCO dataset. It finds that an optimal mixing ratio of parallel query translations significantly enhances retrieval performance, outperforming monolingual queries in 88 out of 105 cases, particularly when retrieving from non-English document indices. The results indicate that English serves as the most effective mixing partner, and the sensitivity to language mixing is predictable, providing insights for practitioners on optimizing multilingual retrieval systems.