Graph-Enhanced Retrieval Augmented Generation Through Cross-Attention and Contrastive Learning

Read full text Summarise with VIRAT
Authors: Gianluca Bonifazi, Christopher Buratti, Michele Marchetti, Federica Parlapiano, Giulia Quaglieri, Davide Traini, Domenico Ursino, Luca Virgili
Cited by: 0
References: 17

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Large Language Models (LLMs) have achieved remarkable success in text generation, but they frequently struggle with factual inaccuracies and limited structural reasoning. While Retrieval-Augmented Generation (RAG) aims to mitigate these issues by sourcing external knowledge, traditional approaches rely heavily on flat vector-based similarity searches. This paradigm often fails to capture the relationships among pieces of information, leading to incomplete or fragmented context. To address this issue, we propose Graph-Enhanced RAG through cross-attention (GERAG), a graph-aware RAG framework that formulates retrieval as a projection of textual queries onto a graph topology. The framework introduces a cross-attention module trained in a self-supervised manner to map textual representations to Knowledge Graph (KG) nodes, by jointly optimizing a contrastive objective and a topological regularization term. This training scheme enforces structural consistency between text embeddings and the KG, enabling structure-aware retrieval and reasoning over entities and relationships, while an adaptive selection mechanism limits redundant context. Experimental results across three fact-checking datasets (MOCHEG, HoVer, and CONFACT) and two LLMs demonstrate that GERAG consistently outperforms state-of-the-art graph-based RAG methods.

Fields of Study & Subjects

Topic Modeling Information Retrieval and Search Behavior Multimodal Machine Learning Applications

Publication details

Journal / source
International Joint Conference on Neural Networks (IJCNN)
Pages
1–6
Publication date
1 June 2026
Language
EN