Dr.E Bridges Graphs with Large Language Models Through Words

Significant efforts have been directed toward integrating powerful Large Language Models (LLMs) with diverse modalities, particularly focusing on the fusion of vision, language, and audio data. However, the graph-structured data, inherently rich in structural and domain-specific knowledge, have not yet been gracefully adapted to LLMs. Existing methods either describe the graph with raw text, suffering the loss of graph structural information, or feed Graph Neural Network (GNN) embeddings directly into LLM at the cost of losing semantic representation. To bridge this gap, we introduce an innovative, end-to-end modality-aligning framework, equipped with a pretrained Dual-Residual Vector Quantized-Variational AutoEncoder (Dr.E). This framework is specifically designed to facilitate token-level alignment with LLMs, enabling an effective translation of the intrinsic `language' of graphs into comprehensible natural language. Our experimental evaluations on standard GNN node classification tasks demonstrate competitive performance against other state-of-the-art approaches. Additionally, our framework ensures interpretability, efficiency, and robustness, with its effectiveness further validated under both fine-tuning and few-shot settings. This study marks the first successful endeavor to achieve token-level alignment between GNNs and LLMs.