PhD Dissertation Defense: Ahmmad O. M. Saleh, MULTI-HOP QUESTION ANSWERING WITH LARGE LANGUAGE MODELS THROUGH KNOWLEDGE GRAPHS, Date & Time: 21 July, 2026 – 1:30 PM, Place: FENS 2019
MULTI-HOP QUESTION ANSWERING WITH LARGE LANGUAGE MODELS THROUGH KNOWLEDGE GRAPHS
Ahmmad O. M. Saleh
Computer Science and Engineering, PhD Dissertation, 2026
Thesis Jury
Prof. Yücel Saygin (Thesis Advisor)
Prof. Şule Öğüdücü
Assoc. Prof. Mehmet Emre Gürsoy
Asst. Prof. Onur Varol
Asst. Prof. Süha Orhun Mutluergil
Date & Time: 21st of July, 2026 – 1.30 PM
Place: FENS 2019
Zoom Link: https://sabanciuniv.zoom.us/j/
Keywords : Large Language Model, GraphRAG, Multi-hop Questions, Knowledge Graph, Text-to-Cypher
Abstract
Large Language Models (LLMs) have achieved remarkable success in natural language processing. However, they often struggle with knowledge-intensive and multi-hop questions that require reasoning over multiple pieces of information. Retrieval-Augmented Generation (RAG) addresses this limitation by incorporating external knowledge during inference, yet conventional text-based retrieval methods frequently fail to capture complex relationships among entities and facts. Knowledge graphs provide a structured representation of knowledge that can support more effective retrieval and reasoning. This thesis investigates graph-based retrieval techniques to improve multi-hop question answering through a unified Graph Retrieval-Augmented Generation (GraphRAG) framework. First, we propose Subgraph Retrieval-Augmented Generation (SG-RAG), a method that retrieves question-relevant subgraphs from a knowledge graph and transforms them into textual context for LLMs. Second, we introduce Merging and Ordering Triplets (MOT), a context optimization technique that reduces redundancy and organizes retrieved knowledge to improve contextual coherence. Third, we present a domain- agnostic Text-to-Cypher framework that translates natural language questions into executable Cypher queries, enabling direct retrieval of graph knowledge across diverse domains. To support this task, we develop an automated pipeline for generating and validating multi-domain Text-to-Cypher datasets. Experiments on the MetaQA benchmark and multiple graph schemas demonstrate that the proposed methods consistently outperform conventional prompting, text-based RAG, and existing graph-based baselines. Furthermore, integrating semantic parsing, graph retrieval, and context optimization into a single end-to-end pipeline yields robust performance on multi-hop question answering tasks. The results highlight the effectiveness of structured graph retrieval and graph-aware reasoning for building more reliable knowledge-intensive AI systems.