Q348 : Design of a Question Answering System baxsed on Knowledge Graphs and Large Language Models
Thesis > Central Library of Shahrood University > Computer Engineering > MSc > 2026
Authors:
[Author], Fatemeh Jafarinejad[Supervisor]
Abstarct: Abstract With the emergence and rapid advancement of Large Language Models (LLMs)، Retrieval-Augmented Generation (RAG) systems have attracted significant attention as an effective approach for accessing up-to-date external knowledge in knowledge-intensive tasks، particularly in addressing the issue of outdated knowledge after model training. However، the inability of these models to fully understand structured relationships and perform multi-hop reasoning often leads to reduced accuracy when answering complex questions. Knowledge Graphs (KGs)، by providing a structured representation of relationships among entities، offer strong potential for overcoming these limitations. Nevertheless، the effective integration of knowledge graphs with large language models remains an open research challenge. In this research، a hybrid frxamework baxsed on Knowledge Graphs and Retrieval-Augmented Generation is proposed، in which the relationships and mextadata extracted from the knowledge graph are transformed into rich and interpretable textual representations using large language models. This textual representation acts as an intermediate laxyer that enables more effective utilization of structured knowledge during both the retrieval and response generation processes. To implement this frxamework، the E5-large-v2 model was used for vector representation، while Qwen 2.5 and GPT-4o-mini were employed as the core response generation models. The proposed system was evaluated on standard benchmark datasets including SimpleQA، GrailQA، and Zero-Shot RE. Experimental results demonstrate that، in the relation extraction task (Zero-Shot RE)، the proposed approach achieved an accuracy of 88.34% using an offline and relatively lightweight model. Compared to the baxseline method، which achieved an accuracy of 84%، this represents an absolute improvement of 4 percentage points. Furthermore، the experimental findings indicate that leveraging graph-enriched textual representations significantly improves answer accuracy، information coverage، and model robustness in complex reasoning scenarios. Overall، this study demonstrates that transforming knowledge graphs into textual representations can effectively reduce the gap between graph-structured knowledge and large language models، while also enabling the practical use of lighter and more accessible models in complex knowledge-intensive applications.
Keywords:
#_Large Language Models، Knowledge Graphs، Retrieval-Augmented Generation، Computational Chemistry_ Keeping place: Central Library of Shahrood University
Visitor: