English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

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

Forum topic · 小凯 · 2026-07-05

Summary

This paper, "The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora" (arXiv:2507.07543, July 2025), studies how retrieval-augmented generation (RAG) systems behave when the knowledge corpus is bilingual, mixing Arabic and English documents. Authored by Chen Amiraz, Yaroslav Fyodorov, Elad Haramaty, Zohar Karnin, and Liane Lewin-Eytan, the work examines retrieval biases that arise when queries are posed in one language while relevant evidence exists in the other. The central finding highlighted by the paper is a cross-lingual cost: retrievers systematically favor content whose language matches the query, which can degrade answer quality when the best evidence is stored in the other language, even when the generator model itself could handle both languages. The study analyzes this behavior across Arabic-English corpora and discusses implications for building fair, effective multilingual RAG pipelines. For practitioners, the takeaway is that retrieval—rather than generation—is often the bottleneck in cross-lingual settings, so multilingual evaluation of the retriever is essential. Full details, experimental setup, and quantitative results are available in the original paper at https://arxiv.org/abs/2507.07543. Readers should consult the PDF for exact numbers and benchmarks.

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

Source: arXiv:2507.07543 (July 2025)

Authors: Chen Amiraz, Yaroslav Fyodorov, Elad Haramaty, Zohar Karnin, Liane Lewin-Eytan

Overview

This paper investigates retrieval biases in retrieval-augmented generation (RAG) systems operating over bilingual Arabic-English corpora. As RAG deployments increasingly serve multilingual user bases, the language composition of the knowledge base and the language of the query can interact in unexpected ways, affecting which documents get retrieved and, ultimately, answer quality.

Key Points

  • The paper studies a cross-lingual cost: when a RAG system operates over a mixed Arabic-English corpus, retrieval quality can suffer depending on the language relationship between the query and the underlying evidence.
  • The core issue is language-matching bias in retrieval: retrievers tend to favor documents in the same language as the query, so when the best supporting evidence is written in the other language, it may be under-retrieved — even if the underlying generator (LLM) is fully capable of using it.
  • This isolates the retriever, not the generator, as the bottleneck in cross-lingual RAG: the generation component can often produce good answers in either language, provided the right context is retrieved.
  • The study uses Arabic-English corpora as the testbed, a pairing that is practically relevant for Middle East deployments and representative of asymmetric-resource language settings.
  • The work has practical implications for multilingual RAG system design and evaluation: teams should evaluate retrieval in cross-lingual conditions (query language ≠ evidence language) rather than only in matched-language settings, and consider language-aware retrieval strategies when indexing bilingual knowledge bases.

    • Why It Matters

    Most RAG benchmarks are English-centric or evaluate retrieval with queries and documents in the same language. This paper highlights a failure mode that such evaluations miss: systems can look healthy in monolingual tests yet systematically under-retrieve the most relevant evidence in real bilingual deployments. For organizations serving Arabic-speaking users with mixed-language knowledge bases, this is a direct production concern.

    Reference

  • Paper: The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora, arXiv, July 2025.
*Note: For exact experimental setups, datasets, and quantitative results, consult the original PDF.*

Tags

#rag#retrieval-augmented-generation#cross-lingual#arabic-nlp#information-retrieval#retrieval-bias#multilingual-llm#arxiv

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208759