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

2026-06-11 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A study on multilingual dense retrieval explores the impact of mixed-language queries using embedding-level interpolation. Researchers conducted a ratio-controlled analysis on the mMARCO dataset, employing the BGE-M3 model to vary the mixing proportion of parallel query translations. The findings reveal that an optimal mixing ratio surpasses the best monolingual performance in 88 out of 105 evaluated cases. A significant asymmetry driven by English dominance was observed: mixing consistently benefits retrieval from non-English document indices, while indices containing English documents perform optimally with pure English queries. English also emerged as the most effective mixing partner for all non-English document languages. Furthermore, after accounting for English dominance, mixing gains showed a negative correlation with typological distance. These patterns demonstrate that language-mix sensitivity is structured, predictable, and robust across different model families and scales.

Key takeaway

For NLP Engineers or Research Scientists developing multilingual dense retrieval systems, you should consider implementing query embedding interpolation. This technique can significantly boost performance for non-English document indices, outperforming monolingual approaches in many cases. However, prioritize pure English queries when your document collection includes English, as mixing can be detrimental there. Leverage English as a strong mixing partner for improving retrieval in other languages, while also noting that gains diminish with increasing typological distance.

Key insights

Query embedding interpolation significantly improves multilingual dense retrieval, but English dominance creates an asymmetric benefit.

Principles

• Optimal query mixing often outperforms monolingual retrieval.
• English queries are best for English-containing document indices.
• English is the strongest mixing partner for non-English languages.

Method

Systematically evaluate retrieval performance by varying mixing proportion of parallel query translations via embedding-level interpolation.

In practice

• Interpolate query embeddings for non-English document retrieval.
• Prioritize pure English queries for English-rich document collections.
• Use English as a mixing partner for other languages.

Topics

• Multilingual Dense Retrieval
• Query Embedding Interpolation
• BGE-M3 Model
• mMARCO Dataset
• English Language Dominance
• Typological Distance

Best for: AI Engineer, Machine Learning Engineer, AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality — New IBM Granite multilingual embedding models offer superior retrieval quality and extended context for diverse language and code applications.
• Rerankers Aren’t Magic Either: When the Cross-Encoder Layer Is Worth the Cost — Cross-encoder rerankers offer inconsistent performance gains over strong embeddings, often failing on complex query types.
• IndicDB -- Benchmarking Multilingual Text-to-SQL Capabilities in Indian Languages — IndicDB reveals a significant "Indic Gap" in LLM Text-to-SQL performance for Indian languages compared to English.
• IdioLink: Retrieving Meaning Beyond Words Across Idiomatic and Literal Expressions — Idiom understanding in retrieval requires semantic abstraction beyond lexical overlap.
• Global vs. Local Sentence Embeddings for Brazilian Portuguese: Revisiting Monolingual Models in the Age of Foundation Models — Optimal model choice for PT-BR sentence embeddings balances vocabulary coverage and adaptation flexibility.
• mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection — Finetuning mid-sized LLMs with QLoRA and augmented data improves multilingual online polarization detection.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.