Harnessing Large Language Models for Temporal Knowledge Graph Forecasting with Multi-view Prompting

In recent years, Temporal Knowledge Graphs (TKGs) have attracted considerable attention from researchers. An important research direction is forecasting future events in TKGs based on historical information, i.e. TKG forecasting. Existing methods primarily rely on modeling temporal relationships between entity IDs to make predictions, which brings limitations in understanding the entity content and the relationships between entities. Fortunately, the emergence of Large Language Models (LLMs) has brought new insights into solving this problem. In this paper, we propose a novel LLM-enhAnced TKG forEcasting (LATE) model. Specifically, we first design multi-view prompts that enable LLMs to generate comprehensive descriptions for entities as external knowledge. Then, we get the enhanced representations of entities based on the descriptions and we further design a contrastive learning strategy to reinforce this process. Finally, we construct a TKG forecasting model based on self-attention mechanisms and dual-pattern LSTM-based temporal modeling to accomplish the forecasting task. Experiments on three datasets demonstrate the effectiveness and superiority of our model compared with several state-of-the-art baseline methods.