Bridging Graph-Structured Data and Large Language Models
Dr Bresson Xavier, Associate Professor, School of Computing
COM3 Level 2
MR21, COM3 02-61
Abstract:
This thesis proposes a unified framework that integrates graph-structured data with Large Language Models (LLMs), combining the structural learning strengths of Graph Neural Networks (GNNs) with the semantic and reasoning capabilities of LLMs. We introduce a three-stage pipeline for bidirectional graph–language integration. First, TAPE enriches graph nodes and edges with LLM-generated textual semantics to provide semantic priors for graph learning. Second, GraphViT decomposes graphs into informative substructures and represents them as sequences compatible with Transformer architectures. Third, G-Retriever retrieves task-relevant subgraphs based on natural language queries, encodes them with GNNs, and injects the resulting representations into LLMs as soft prompts for structure-aware generation. Together, this framework enables scalable, task-agnostic integration of graph and language modalities, improving both graph-based learning and LLM reasoning over structured data.