Pioneering Temporal Knowledge Graph Completion with Large Language Models: A Text Generation Approach

ABSTRACT Traditional temporal knowledge graph completion (TKGC) methods often rely on random encoding or pre‐trained small‐scale language models to initialize entity and relation embeddings. Despite the remarkable reasoning and understanding capabilities of large language models (LLMs), their potential remains untapped in the context of TKGC. To bridge this gap, we propose a novel framework, large‐language model‐based temporal knowledge graph completion, which reformulates the TKGC task as a text generation problem. In this approach, entities, relations, and temporal information within quadruples are converted into text sequences, enabling the use of task‐specific prompts to guide the LLM in generating predictions. This innovative approach fundamentally differs from existing methods by leveraging the generative capabilities of LLMs. Experimental evaluations on four public benchmark datasets demonstrate the effectiveness of our model, showcasing its superior performance and the ability of LLMs to significantly advance the state‐of‐the‐art in temporal knowledge graph completion.