Multi-Grained Knowledge for Retrieval-Augmented Question Answering on Hyper-long Contexts
Keywords: Knowledge-based Question Answering, Retrieval-Augmented, Large Language Model Generation, Information Extraction, Hyper-long Contexts
TL;DR: This paper presents a Multi-Grained Retrieval-Augmented Generation (MGRAG) method for hyper-long context question answering, integrating multi-grained entity graph with iterative retrieval and reasoning.
Abstract: In the task of hyper-long context question answering (QA), a key challenge is extracting accurate answers from vast and dispersed information, much like finding a needle in a haystack. Existing approaches face major limitations, particularly the input-length constraints of Large Language Models (LLMs), which hinder their ability to understand hyper-long contexts. Furthermore, Retrieval-Augmented Generation (RAG) methods, which heavily rely on semantic representations, often experience semantic loss and retrieval errors when answers are spread across different parts of the text.
Therefore, there is a pressing need to develop more effective strategies to optimize information extraction and reasoning.
In this paper, we propose a multi-grained entity graph-based QA method that constructs an entity graph and dynamically combines both local and global contexts. Our approach captures information across three granularity levels (i.e., micro-level, feature-level, and macro-level), and incorporates iterative retrieval and reasoning mechanisms to generate accurate answers for hyper-long contexts.
Specifically, we first utilize EntiGraph to extract entities, attributes, relationships, and events from hyper-long contexts, and aggregate them to generate multi-granularity QA pairs. Then, we retrieve the most relevant QA pairs according to the query. Additionally, we introduce LoopAgent, an iterative retrieval mechanism that dynamically refines queries across multiple retrieval rounds, combining reasoning mechanisms to enhance the accuracy and effectiveness of answering complex questions.
We evaluated our method on various datasets from LongBench and InfiniteBench, and the experimental results demonstrate the effectiveness of our approach, significantly outperforming existing methods in both the accuracy and granularity of the extracted answers. Furthermore, it has been successfully deployed in online novel-based applications, showing significant improvements in handling long-tail queries and answering detail-oriented questions.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5772
Loading