arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2504.04771cs.CL
收藏

通过辩证推理论证改进多语言检索增强语言模型

Improving Multilingual Retrieval-Augmented Language Models through Dialectic Reasoning Argumentations

School of Informatics, University of Edinburgh(爱丁堡大学信息学院);University of Rome Tor Vergata(罗马第二大学)

Leonardo Ranaldi, Federico Ranaldi, Fabio Massimo Zanzotto, Barry Haddow, Alexandra Birch

AI总结:

本文提出Dialectic-RAG(DRAG)框架,通过论证性解释和辩证推理解决多语言检索增强生成中的知识冲突问题,实验证明其能以低算力开销显著提升RAG性能并增强对知识扰动的鲁棒性。

摘要:

检索增强生成(RAG)是增强大型语言模型(LLMs)以系统性获取更丰富事实知识的关键。然而,使用RAG带来了内在挑战,因为LLMs必须处理潜在冲突的知识,尤其是在多语言检索中,检索到的知识异质性可能提供不同的观点。为了使RAG更具分析性、批判性和依据性,我们引入了Dialectic-RAG(DRAG),这是一种由论证性解释引导的模块化方法,即通过比较、对比和解决冲突视角来系统性评估检索信息的结构化推理过程。给定一个查询和一组多语言相关文档,DRAG选择并例示相关知识以提供辩证解释,通过批判性地权衡对立论点和过滤无关内容,明确确定最终响应。通过一系列深入实验,我们展示了该框架作为上下文学习策略以及构建演示以指导较小模型的影响。最终结果表明,DRAG显著改进了RAG方法,仅需低影响的算力开销,并对知识扰动具有鲁棒性。

英文摘要:

Retrieval-augmented generation (RAG) is key to enhancing large language models (LLMs) to systematically access richer factual knowledge. Yet, using RAG brings intrinsic challenges, as LLMs must deal with potentially conflicting knowledge, especially in multilingual retrieval, where the heterogeneity of knowledge retrieved may deliver different outlooks. To make RAG more analytical, critical and grounded, we introduce Dialectic-RAG (DRAG), a modular approach guided by Argumentative Explanations, i.e., structured reasoning process that systematically evaluates retrieved information by comparing, contrasting, and resolving conflicting perspectives. Given a query and a set of multilingual related documents, DRAG selects and exemplifies relevant knowledge for delivering dialectic explanations that, by critically weighing opposing arguments and filtering extraneous content, clearly determine the final response. Through a series of in-depth experiments, we show the impact of our framework both as an in-context learning strategy and for constructing demonstrations to instruct smaller models. The final results demonstrate that DRAG significantly improves RAG approaches, requiring low-impact computational effort and providing robustness to knowledge perturbations.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑