The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora

Forum topic · 小凯 · 2026-07-05

Summary

This arXiv paper (July 2025) examines how retrieval biases affect Retrieval-Augmented Generation (RAG) systems operating over bilingual Arabic-English corpora. The authors—Chen Amiraz, Yaroslav Fyodorov, Elad Haramaty, Zohar Karnin, and Liane Lewin-Eytan—study what they term the cross-lingual cost: the performance degradation that arises when a RAG system must retrieve relevant evidence across languages rather than within a single language. The work identifies retrieval, rather than generation, as the key bottleneck in cross-lingual RAG settings, showing that retrievers exhibit systematic biases toward particular languages when handling queries and documents in mixed Arabic-English corpora. The study provides a controlled benchmark setup for disentangling retrieval quality from generation quality in multilingual pipelines, and analyzes how these biases translate into downstream answer quality. The findings are relevant for practitioners deploying RAG in Arabic-speaking markets and for researchers working on multilingual information retrieval, as they highlight that multilingual LLM capabilities do not automatically translate into effective cross-lingual retrieval. Source: https://arxiv.org/abs/2507.07543. Readers should consult the original PDF for detailed quantitative results and experimental protocols.

The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora

Authors: Chen Amiraz, Yaroslav Fyodorov, Elad Haramaty, Zohar Karnin, Liane Lewin-Eytan Source: arXiv:2507.07543, July 2025 Category: Multilingual / RAG

Key points

The paper studies the cross-lingual cost in Retrieval-Augmented Generation (RAG): the performance loss incurred when systems must retrieve evidence across Arabic-English corpora instead of operating monolingually.
A central finding is that retrieval, not generation, is the bottleneck in cross-lingual RAG pipelines — modern multilingual LLMs can answer well when given the right context, but retrievers fail to surface it reliably.
Retrievers exhibit systematic language biases, favoring documents in particular languages regardless of query language, which degrades answer quality in mixed-corpus settings.
The work provides a controlled evaluation setup that disentangles retrieval errors from generation errors, offering a reproducible protocol for auditing multilingual RAG stacks.

Why it matters

For practitioners: Deploying RAG in Arabic-speaking markets requires explicit attention to retrieval quality; simply swapping in a stronger generator will not fix cross-lingual recall failures.
For researchers: The paper adds to a growing line of work on multilingual IR (see related entries such as CLIRudit and evaluations of embedding models for cross-lingual QA), emphasizing that multilingual embedding benchmarks do not guarantee end-to-end RAG quality.

Context

The paper sits at the intersection of multilingual information retrieval and RAG system design. Classic search stacks (recall → re-rank → generation) assume the corpus language matches the query language; in bilingual corpora, each stage can introduce language-dependent distortions. This work isolates where those distortions arise and quantifies their downstream cost.

Related entries

A Comprehensive Evaluation of Embedding Models and LLMs for IR and QA
CLIRudit: Cross-Lingual Information Retrieval of Scientific Documents (arXiv:2504.16264)
Evaluating Large Language Models for Cross-Lingual Retrieval (arXiv:2509.14749)
Cross-Lingual Cross-Modal Retrieval with Noise-Robust Fine-Tuning (IEEE)

> Note: Quantitative results should be verified against the original PDF at https://arxiv.org/abs/2507.07543.

The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora

Summary

The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora

Key points

Why it matters

Context

Related entries

Tags