Core Concepts
Incorporating multi-perspective domain insights into retrieval-augmented generation (RAG) systems significantly improves their performance and reliability in knowledge-intensive fields.
Abstract
The paper introduces a novel multi-view retrieval framework, MVRAG, designed to enhance the effectiveness of retrieval-augmented generation (RAG) systems in knowledge-dense domains like law and medicine.
The key highlights are:
Intention Recognition: The framework utilizes a large language model to identify the underlying intent and assign relevance scores to different professional perspectives, forming a Perspective Vector.
Query Rewriting: The Perspective Vector guides the query rewriting process, where the original query is tailored to each identified perspective to retrieve contextually relevant documents.
Retrieval Augmentation: The retrieved documents are re-ranked based on their alignment with the multi-perspective query and integrated into a structured prompt for final inference.
The experiments conducted on legal and medical case retrieval datasets demonstrate significant improvements in recall, precision, and F1 scores compared to baseline models. The multi-view approach proves effective in capturing the complex relationships and local nuances inherent to specialized domains, enhancing the reliability and interpretability of RAG systems.
Stats
"The recall@100 for the bge-m3 model improved from 3.125% to 16.53% with the multi-view framework."
"The recall@100 for the bge-large-en model in the medical dataset increased from 8.791% to 15.14%."
Quotes
"Our multi-view query rewriting technique significantly improves the relevance and performance of information retrieval, representing a paradigm shift from traditional methods."
"The framework's integration into RAG systems substantially increases document retrieval scope and accuracy, ensuring high-relevance information retrieval for domain-specific tasks."