💬 NLP

Multilingual Retrieval-Augmented Generation for Knowledge-Intensive Task

This paper investigates multilingual retrieval-augmented generation strategies and proposes CrossRAG, a method that translates retrieved documents into a common language before response generation, to significantly enhance performance on knowledge-intensive tasks across both high- and low-resource languages.

Original authors: Leonardo Ranaldi, Barry Haddow, Alexandra Birch

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Leonardo Ranaldi, Barry Haddow, Alexandra Birch

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a brilliant chef (the AI) who wants to cook a perfect meal (the Answer) for a guest who speaks a specific language, like Italian. The problem is, your chef's memory is a bit fuzzy, and they sometimes make things up (hallucinations) or forget specific facts.

To fix this, you give the chef a library of cookbooks (Retrieval-Augmented Generation, or RAG) to look at while they cook. This paper explores what happens when the guest speaks many different languages, and the library contains cookbooks in every language of the world.

Here is the story of their journey, explained simply:

1. The Problem: The Language Barrier

In the past, researchers only tested this system with English. It worked great! But what if the guest asks a question in Spanish, and the library only has English books?

The Old Way (Monolingual RAG): You tell the chef, "Only look at the Spanish cookbooks."
- Result: If the fact isn't in a Spanish book, the chef guesses or says "I don't know," even if the answer is right there in a French or German book.
The "Translate First" Way (tRAG): You translate the guest's Spanish question into English, find the English answer, and then translate it back.
- Result: This often fails. Translating a tricky question can change its meaning, leading the chef to the wrong cookbook entirely. It's like asking for "a hot dog" in a language where that phrase means "a sausage with a bun," but the translator turns it into "a sausage in a bun," and the chef gets confused.

2. The First Big Step: The "Global Library" (MultiRAG)

The researchers tried a new approach: MultiRAG.
Instead of limiting the chef to one language, they said, "Look at ALL the cookbooks in the library, regardless of language!"

The Good News: This is a huge win! The chef now has access to way more information. If the answer is missing in Spanish, it might be in Japanese or English. This helped the chef answer questions much better, especially for rare languages (Low-Resource languages) where information is scarce.
The Bad News: The kitchen got chaotic. The chef was handed a stack of cookbooks: one in German, one in Swahili, one in English, and one in Korean.
- The Confusion: The chef got overwhelmed. They struggled to mix these different languages together to form a single, coherent answer. Sometimes they answered in the wrong language, or they got confused by conflicting information. It was like trying to bake a cake while reading a recipe written in five different languages at the same time.

3. The Solution: The "Universal Translator" (CrossRAG)

To fix the chaos, the researchers invented CrossRAG.
Here is the new workflow:

Search Everywhere: The chef still looks at all the cookbooks in every language (just like MultiRAG).
Translate the Evidence: Before the chef starts cooking, a Translator Robot takes all those foreign cookbooks and translates the relevant pages into English (a common language the chef is an expert in).
Cook and Serve: The chef reads the English summaries of the foreign books, cooks the answer, and then serves the final dish back to the guest in their original language (e.g., Italian).

Why this works:

The chef gets the best of both worlds: the massive knowledge of the global library + the clarity of reading in a language they understand perfectly.
It stops the chef from getting confused by mixed languages.
It works incredibly well for both common languages (like Spanish) and rare ones (like Finnish or Bengali).

The Analogy of the "Confused Tourist"

Imagine you are a tourist in a foreign country asking a local for directions to a specific museum.

Monolingual: You only ask people who speak your language. If no one speaks it, you are lost.
MultiRAG: You ask everyone in the square, regardless of language. You get directions from a German, a Chinese, and a Brazilian. But they are all shouting different things in different languages, and you can't understand the mix. You get lost.
CrossRAG: You ask everyone in the square. Then, a friendly guide (the Translator) listens to all of them, translates their directions into your language, and gives you a clear, single set of instructions. You find the museum easily.

The Bottom Line

The paper proves that to make AI smart in a multilingual world, we shouldn't just translate the question. We should:

Search for answers in every language.
Translate the answers into a language the AI understands best.
Let the AI generate the final response in the user's language.

This method makes AI more accurate, less likely to lie, and much better at helping people in languages that usually get ignored by technology.

1. Problem Statement

Retrieval-Augmented Generation (RAG) has significantly improved Large Language Models' (LLMs) factual accuracy and reduced hallucinations in monolingual (English) settings. However, its application in multilingual contexts remains under-explored. Existing approaches face specific challenges:

Limited Coverage: Naïve translation of queries (Question-Translation) often leads to retrieval failures due to translation errors or limited domain coverage in the target language.
Information Asymmetry: Monolingual retrieval (retrieving only in the query language) often fails because high-quality knowledge may exist only in other languages (e.g., English).
Cross-Lingual Inconsistency: When retrieving documents in multiple languages (Multilingual RAG), LLMs struggle to synthesize information from heterogeneous sources, leading to incorrect answers, wrong output languages, or hallucinations, particularly in low-resource languages.

2. Methodology

The authors propose and evaluate four distinct RAG pipelines to address these issues, tested across three knowledge-intensive benchmarks: MLQA, MKQA, and XOR-TyDi QA.

A. Baseline Approaches

Monolingual RAG (monoRAG): The query and the retrieval corpus are in the same language ( $L_{SL}$ ). This limits the knowledge base to documents available only in that specific language.
Translation RAG (tRAG): The query is translated into English, retrieval is performed on an English-only corpus, and the answer is generated. This suffers from translation errors and limited retrieval scope.
Multilingual RAG (MultiRAG): The query is kept in the original language, but retrieval is performed across a union of corpora in all available languages ( $\bigcup D_i$ ). This maximizes information coverage but introduces the challenge of processing retrieved documents in mixed languages.

B. Proposed Solution: Crosslingual RAG (CrossRAG)

To mitigate the inconsistencies of MultiRAG, the authors propose CrossRAG.

Mechanism: Documents are retrieved using the MultiRAG strategy (across all languages). However, before the LLM performs inference, all retrieved documents are translated into a common language (English) using an external translation tool (e.g., Google Translate or GPT-4o).
Goal: The LLM receives the query in the original language but processes the "Reference Evidence" entirely in English, ensuring consistent semantic understanding while retaining the breadth of multilingual knowledge.

C. Experimental Setup

Retrieval System: Cohere's multilingual embedding model (Cohere_Embed_V3) and re-ranker on a Wikimedia dump.
Models Evaluated: GPT-4o, Llama-3-8b-instruct, and Command-R.
Evaluation Metrics: Flexible exact-match accuracy and character-level 3-gram recall.

3. Key Contributions

Systematic Evaluation of Multilingual RAG: The paper provides a comprehensive analysis of how extending retrieval beyond English affects accuracy and consistency, identifying that while multilingual retrieval increases information access, it degrades performance if the LLM cannot handle mixed-language contexts.
Identification of Low-Resource Limitations: The study reveals that LLMs are proficient at understanding multilingual queries but fail to extract information effectively from retrieved documents in low-resource languages when those documents are not translated.
Proposal of CrossRAG: The introduction of a document-level translation pipeline that decouples retrieval scope from inference language. This allows models to leverage global knowledge bases while operating in a single, high-performance language (English) during generation.
Robustness Analysis: The authors demonstrate that CrossRAG is more robust to variations in document ordering compared to MultiRAG, which is sensitive to the sequence of retrieved languages.

4. Key Results

Performance Gains:
- MultiRAG outperforms monoRAG and tRAG across all models, confirming that accessing heterogeneous knowledge bases is beneficial.
- CrossRAG achieves the highest performance, significantly outperforming MultiRAG.
  - GPT-4o: +3.5% average improvement over MultiRAG.
  - Command-R: +2.2% average improvement.
  - Llama-3-8b: +2.9% average improvement.
Impact on Low-Resource Languages (LR): The benefits of CrossRAG are most pronounced in low-resource languages.
- For Command-R, CrossRAG improved LR performance by +5.0% over MultiRAG.
- For Llama-3-8b, the improvement was +4.1%.
- This suggests that translating retrieved content to English helps models overcome the scarcity of high-quality reasoning capabilities in low-resource languages.
Translation Quality: Using GPT-4o as a translator yielded slightly better results than Google Translate for CrossRAG, but both significantly outperformed non-translated baselines.
Language Consistency: CrossRAG significantly improved the rate of answers generated in the correct language, whereas MultiRAG often failed to follow language instructions when presented with mixed-language contexts.

5. Significance and Conclusion

This paper establishes that multilingual retrieval is essential for knowledge-intensive tasks but requires language alignment during the generation phase to be effective.

Theoretical Insight: It highlights a critical bottleneck in current RAG systems: the inability of LLMs to synthesize information from documents in multiple languages simultaneously without degradation in performance.
Practical Application: The CrossRAG approach offers a computationally efficient strategy (using existing translation tools) to unlock the full potential of multilingual knowledge bases without requiring massive retraining of the LLM.
Future Direction: The work argues for a strategic combination of retrieval expansion and document translation to build reliable, culturally aware, and linguistically diverse AI systems, moving beyond the English-centric paradigm of current RAG pipelines.