GraphXtract: Adaptive Graph-Based Sentence Selection for Resource-Efficient Lay Summarization with Large Language Models
Year: 2026
Type: conference-paper
Abstract: Lay summarization of biomedical articles is essential for making scientific research accessible to non-expert readers. While Large Language Models (LLMs) are effective for this task, improving their performance typically requires fine-tuning, which demands high computational resources. To address this issue, we propose a resource-efficient framework that focuses on adaptive sentence selection rather than model training. In our approach, we represent scientific articles as semantic graphs to capture the relationships between sentences. We train a Graph Neural Network (GNN) to predict the importance of each sentence based on ROUGE similarity labels. These selected "Key Facts" are then used in a "Draft-and-Refine" prompting strategy to guide the LLM. We evaluated our framework on the PLOS dataset using various general and biomedical LLMs. The results show that our proposed method, GraphXtract, outperforms baseline methods like similarity-based selection and PageRank. Specifically, the combination of GraphXtract and the MedGemma model achieves a high ROUGE-1 score of 0.4255. Additionally, using Gemma 3 yields the highest factual consistency with a SummaC score of 0.5133. These findings demonstrate that our graph-based approach generates accurate and informative lay summaries efficiently without the need for expensive fine-tuning.
Institutions: Özyeğin University, Galatasaray University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7171554242