Multilingual Retrieval-Augmented Generation for Knowledge-Intensive Task

2025-04-04

AI Summary

This work addresses the challenges of insufficient cross-lingual coverage and semantic misalignment in retrieval-augmented generation (RAG) for multilingual open-domain question answering. We propose CrossRAG, a cross-lingual RAG framework that uniformly back-translates multilingual retrieval results into the target language (e.g., English) prior to answer generation. CrossRAG is the first to systematically integrate machine translation, multilingual retrieval (using mBERT/XTREME), cross-lingual embedding alignment, and LLM-based conditional generation—thereby overcoming key limitations of monolingual RAG. Experiments on XQuAD, MLQA, and other benchmarks demonstrate an average +8.2 F1 improvement; notably, performance gains are most pronounced for low-resource languages (e.g., Swahili, Bengali), while inference efficiency surpasses full-translation pipelines. Our core contribution lies in establishing cross-lingual document alignment as a critical factor for generation quality, and in positioning CrossRAG as a new paradigm for multilingual, knowledge-intensive NLP tasks.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Knowledge Representation Languages

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Retrieval-augmented generation (RAG) has become a cornerstone of contemporary NLP, enhancing large language models (LLMs) by allowing them to access richer factual contexts through in-context retrieval. While effective in monolingual settings, especially in English, its use in multilingual tasks remains unexplored. This paper investigates the effectiveness of RAG across multiple languages by proposing novel approaches for multilingual open-domain question-answering. We evaluate the performance of various multilingual RAG strategies, including question-translation (tRAG), which translates questions into English before retrieval, and Multilingual RAG (MultiRAG), where retrieval occurs directly across multiple languages. Our findings reveal that tRAG, while useful, suffers from limited coverage. In contrast, MultiRAG improves efficiency by enabling multilingual retrieval but introduces inconsistencies due to cross-lingual variations in the retrieved content. To address these issues, we propose Crosslingual RAG (CrossRAG), a method that translates retrieved documents into a common language (e.g., English) before generating the response. Our experiments show that CrossRAG significantly enhances performance on knowledge-intensive tasks, benefiting both high-resource and low-resource languages.

Problem

• Evaluates RAG effectiveness in multilingual question-answering tasks
• Compares tRAG and MultiRAG strategies for multilingual retrieval
• Proposes CrossRAG to improve cross-lingual consistency and performance

Innovation

• Multilingual RAG for cross-language question-answering
• Question-translation RAG improves English retrieval
• CrossRAG translates documents for consistent responses