- The paper introduces a ratio-controlled embedding interpolation protocol that systematically improves retrieval effectiveness over 35 language pairs.
- It demonstrates that mixing yields significant gains (up to +2.9 nDCG) on non-English document indices, particularly when English is used as a mixing partner.
- The study reveals that increasing typological distance reduces gains and recommends using pure English queries when the document index contains English.
Embedding-Level Mixing for Multilingual Dense Retrieval: A Ratio-Controlled Analysis
Introduction
The paper "When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval" (2606.13537) systematically interrogates the retrieval dynamics of code-mixed queries within multilingual dense retrievers, focusing on how embedding-level query mixing influences retrieval effectiveness. The study leverages embedding interpolation between monolingual query embeddings and evaluates retrieval performance over different query-language and document-language configurations. A strong emphasis is placed on the operationalization and predictability of mixing gains, with extensive experimentation over 35 language pairs and varying mixing ratios using state-of-the-art multilingual retrievers.
Methodology
A central contribution is the introduction of a ratio-controlled embedding-level mixing protocol, which constructs mixed-language queries by convexly interpolating between monolingual query embeddings. Parallel translations from the mMARCO collection are used as the base queries, and the embedding mixing ratio is varied in controlled increments. Experiments are carried out using BGE-M3 and several other embedding models, including Qwen3, E5, and Jina, to ensure model-agnostic findings.
The experimental design considers three document index types per language pair: monolingual L1-only, monolingual L2-only, and a bilingual union, carefully controlling for confounds such as translation artifacts and language purity. Evaluation primarily uses nDCG@10, MRR@10, and Recall@10. Gains from mixing are reported as the improvement of the best mixed-ratio score over the better monolingual endpoint.
Key Empirical Findings
- Mixing Yields Consistent Gains When English Is Absent: In 88/105 (83.8%) of all (language pair, document-language) experimental settings, optimal embedding-level mixing outperforms the best monolingual endpoint. The observed mean gain is +0.70 nDCG, with some settings exceeding +2.9 nDCG improvement.
- Asymmetric Role of English: English occupies a uniquely dominant and asymmetric role. When English appears in the document index, mixing is neutral or, at best, provides no improvement. For non-English document indices, mixing with English consistently improves retrieval effectiveness.
- English Is Always the Strongest Mixing Partner: Across all non-English document settings, English as a mixing partner yields the highest gains, regardless of the paired language.
- Typological Distance Negatively Correlates with Gains: When controlling for the presence of English, mixing benefits decrease as the typological distance between the paired languages increases.
- Mixing Affects Recall Disproportionately: Gains are most pronounced in Recall@10, indicating that mixing is more effective at surfacing relevant documents into the candidate set than optimizing their top-rank ordering.
Theoretical and Practical Implications
The study demonstrates that language-mixing sensitivity in current multilingual dense retrievers is both structured and predictable. The findings clarify why simple embedding interpolation is a valid diagnostic and intervention tool: word-level code-mixed queries, when embedded, are found to lie nearly on the linear trajectory between their corresponding monolingual query embeddings, supporting the methodology.
Practically, the results prescribe a robust deployment rule: use pure English queries whenever the document index contains English, and apply embedding-level mixing otherwise to optimize non-English retrieval, especially in multilingual and bilingual document settings involving non-English languages. Additionally, these patterns generalize across major embedding model architectures and model scales.
From a theoretical perspective, the observed dominance of English can be traced to structural biases in pre-training datasets, where English constitutes the majority of the multilingual pre-training data. This "anchor" effect leads to denser and more robust semantic representations in the English subspace.
Discussion and Future Directions
The evidence highlights the systematic nature of mixing-ratio gains in controlled settings, but also the limitations of embedding-mixing as a surrogate for naturally occurring code-mixing in user queries. The research outlines several open directions:
- Extending to Naturally Code-Mixed Queries: Current findings are grounded in parallel-translation-based queries; future work should address the complexities of organic code-switching, including spelling variation, transliteration, and ad hoc mixing.
- Multi-way Mixing: The current focus on language pairs could be extended to simultaneous mixing among three or more languages, reflecting the realities of highly multilingual user bases and corpora.
- Fine-Grained Typology and Sociolinguistics: Investigation into more nuanced typological factors and sociolinguistically motivated mixing may yield deeper insights and model improvements.
Conclusion
The studied ratio-controlled embedding-level mixing protocol offers a systematic and computationally efficient mechanism for both analyzing and improving multilingual dense retrieval. The strong, reproducible asymmetry with respect to English confirms that code-mixing strategies must be adapted according to the document index composition. The insights provide concrete guidance for multilingual retrieval-augmented generation and IR system practitioners and suggest several research avenues into more realistic and diverse code-mixed retrieval scenarios.
